Journal of Advanced Sciences Applied Engineering - … · · 2016-07-03Journal of Advanced...

University of Tebessa, Tebessa (Algeria)

Journal of Advanced Sciences &

Applied Engineering

Volume: 01 Issue: 01 Year: 2014

ISSN: 0000-000x

0

10

20

30

40

50

60

70

80

90

100

-30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30

Déformation radiale (‰) Déformation axiale (‰)

Co

ntr

ain

te a

xial

e (M

Pa)

Non confiné

1couche

3 couches

Journal Policy Journal of Advanced Sciences & Applied

Engineering is a multidisciplinary international

journal which contains a whole of academic

publications having for main aims, the presentation

the university scientific research work and the

exchange of ideas and experiments with the other

national and international institutions and the

contribution through scientific production in the

evolution, technological development and scientific.

The scientific papers publish original writings (in

English, and French) raising as well of the basic

research as of the applied research in several

disciplines: Analysis, Algebra, Applied

Mathematics, Computational Mathematics,

Mathematical Chemistry and Biology, Theoretical

physics, Applied Physics, Nuclear Physics,

Organic materials, Theoretical chemistry,

Chemistry of the environment, Chemical

Engineering, Electronics, Electronic devices

Electrical engineering, Automation and control,

Material science, Nano-materials, Nano-systems,

Nanostructures, Spintronics & spin transport,

Nano-magnetism, Surface and interface,

Superconductivity, Semiconductors, Energy

storage and generation, New energies, Networks

and communication, Computer science,

Mechanical engineering, Hydropneumatic order

and industrial automatisms, Geophysical,

Geotechnics, Electromechanics Hydraulic, Ground

sciences, Civil engineering, Town and country

planning, Biological Sciences, Biotechnology,

Agroalimentary industries, Agronomic Sciences,

Architecture, Mining environment, Mining

geotechnics, Mining exploitation, Industrial

maintenance.

Journal of Advanced Sciences & Applied Engineering (JASAE)

University of Tebessa, Tebessa, Algeria Tel:+213 (0) 37584637 Fax: +213 (0) 37584637 Email: [email protected] Printing: SARL DAR EL-HOUDA, Ain Mlila, Oum El-Bouaghi, 04000 Algeria Website: http://www.darelhouda.com Email: [email protected] Tel: +213 (0) 21966220 Fax: +213 (0) 21966111

Journal of Advanced Sciences & Applied Engineering

A multidisciplinary international journal of Sciences and Engineering – Algeria (Univ. Tebessa)

Honorary Editor Said Fekra

(Rector of the University of Tebessa)

Editor-in-Chief Abderrachid Bechiri Department of Physics, University of Tebessa, Algeria Fax: +213 (0) 37584637. E-mail: [email protected]; [email protected]

Editors

Salah Chenikher Department of Electronics, University of Tebessa, Algeria E-mail: [email protected] BenmakhloufDepartment of Physics, University of Tebessa, AlgeriaE-mail: [email protected] NinouhDepartment of Civil engineering, University of Tebessa, Algeria

E-mail: [email protected] RouiliDepartment of Civil engineering, University of Tebessa, AlgeriaE-mail: [email protected] Messabhia Department of Civil engineering, University of Tebessa, Algeria

E-mail: [email protected] Louafi Department of Electromechanics, University of Tebessa, Algeria E-mail: [email protected] Ridda LaouarDepartment of Computer science, University of Tebessa, Algeria

E-mail: [email protected] Djabri Department of Biology, University of Tebessa, Algeria

E-mail: [email protected] Baali Department of Geology, University of Tebessa, Algeria E-mail: [email protected]

Advisory Board Abdelaziz Khadraoui, Univ. Geneva, Switzerland

Abdelhamid Bouchair, Univ. Clermont Ferrand, France

Abdelkader Ali Benamara, Univ. Chlef, Algeria

Abdelkrim Amirat, Univ. Souk Ahras, Algeria

Abdelkrim Gouasmia, Univ. Tebessa, Algeria

Abderrachid Bechiri, Univ. Tebessa, Algeria

Aderrahmane Gahmous, Univ. Tebessa, Algeria

Alexander Granovsky Univ. Moscow, Russia

Ali Rahmouni, Univ. Saida, Algeria

Ammar Dibi, Univ. Batna, Algeria

Ammar Zellaghi, Univ. Oum El Bouaghi, Algeria

Amor Djemel, Univ. Constantine, Algeria

Anis Moumen, Univ. Kenitra, Morocco

Antonio Bianconi, RICMASS, Roma, Italy

Antonio Pulido-Bosch,Univ. Almeria, Spain

Badis Bennecer, Univ. Guelma, Algeria

Belgacem Djabri, Univ. Tebessa, Algeria

Bernard Barbara, Institut Néel, CNRS, Grenoble, France

Djamel Boukredimi, Univ. Oran, Algeria

Farid Mokhat, Univ. Oum El Bouaghi, Algeria

Fatima Hamdache, USTO, Algeria

Fauziah Ahmad, Univ. Sains Malaysia Penan,Malaysia

Fella Benmakhlouf, Univ. Tebessa, Algeria

Hamid Khachab, Univ. Bechar, Algeria

Hamid Mecheik, Univ. UQAM, Montreal, Canada

ILhami Colak, Univ. Gazi, Turkey

Jacqueline Signorini, Univ. Paris 8, France

Jaky Mania, Polytech Lille, France

Jian-Xin Shen, University Hangzhou, China

Kais Bouallegue, High Institute of Applied Sciences and Technology

of Sousse, Tunisia

Kamel Haouam, Univ. Tebessa, Algeria

Layachi Gouaidia, Univ. Tebessa, Algeria

Liès Dekar, Univ. Medea, Algeria

Mabrouk Touahmia, Univ. Hail, Saudi Arabia

Martin Achmus, Leibniz Universität Hannover, Germany

Mohamed El Bachir Menai, Univ. King Saud, Saudi Arabia

Mohamed Faouzi Harkat, Univ. Annaba, Algeria

Mohamed Ridda Laouar, Univ. Tebessa, Algeria

Mohamed Sahnoun, Univ. Mascara, Algeria

Mostafa Ezziyyani, Univ. Tanger, Morocco

Mouhamed Khetaoui, Univ. Tizi ouzou, Algeria

Mounir Terki-Hassein, Univ. Mostaganem, Algeria

Musa Resheidat, Jordan University, Jordan

Nadir Bouarissa, Univ. M’sila, Algeria

Okba Kazar, Univ. Biskra, Algeria

Philippe Audra, Polytech Nice, France

Rachid Bensalem, Univ. Annaba, Algeria

Said Mesloub, Univ. King Saud, Saudi Arabia

Salah Chenikher, Univ. Tebessa, Algeria

Tarek Bouktir, Univ. Setif, Algeria

Tarek Ninouh, Univ. Tebessa, Algeria

Tayeb Benouaz, Univ. Tlemcen, Algeria

Wolfgang Kleemann, Univ. Duisburg-Essen, Germany

Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014)

A new class of attractor generated by fractal processes Pages 1-7 Kais Bouallegue, Youcef Soufi, Abdessattar Chaari

Application of Differential Evolution Algorithm to Optimal Power Flow incorporating FACTS: a case study Pages 8-15 Linda Slimani, Tarek Bouktir

Calculation of transient electromagnetic field in an electrical Substation EHV / HV Pages 16-20 A. Bendakir, D. Dib, T. Ruibah

Design of a fractional order PID controller for a class of fractional order MIMO systems Pages 21-27 Abdelhamid Djari, Toufik Bouden, Abdesselem Boulkroune

Effects of Different Parameters on Power System Transient Stability Studies Pages 28-33 M. Amroune, T. Bouktir

Energy Conversion System Performance and Analysis Pages 34-38 Y. Soufi, T. Bahi, S. Ghoudelbourk, H. Merabet and S. Lekhchine

Face localization using neural networks trained with geometric and skin characteristics Pages 39-44 M. Saaidia, A. Gattal, M. Maamri, M. Ramdani

Fuzzy Regression Analysis using Quadratic and Support Vector Machines Approaches Pages 45-49 Riadh Djabri, Fayçal Megri, Noureddine Guerfi

Nonlinear model-based coagulant dosing control at water treatment plants Pages 50-54 Toufik Benzaraa, Messaoud Ramdani, Khaled Mendaci, Abdenabi Abidi

PSO Optimization with Autoregressive Modeling and Support Vector Machines for Bearing Fault Diagnosis Pages 55-59 Thelaidjia Tawfik, Salah Chenikher

Evaluation of antioxidative and antibacterial potentials of Crataegus monogyna Jacq. from Mahouna mountain (Algeria) Pages 60-63 Samah Djeddi and Hanifa Boutaleb

Screening of antagonistic activity of indigenous bacteria against two fusarium species Pages 64-66 S. Mezaache-Aichour , N. Sayah, N. Haichour, A. Guechi, M. M. Zerroug

A Survey of the Possible Role of German Cockroaches as a Source for Bacterial Pathogens Pages 67-70 Taha Menasria, Samir Tine, Souad EL-Hamza, Djaouida Mahcene, Fatima Moussa, Leyla Benammar, Mohamed Nacer Mekahlia

Mechanical behaviour of square concrete columns wrapped with CFRP composite Pages 71-75 Riad Benzaid, Nasr eddine Chikh, Habib Mesbah

Modeling Of Chloride Penetration In Concrete Structures In Cold Regions Pages 76-82 H’mida Hamidane, Ayman Ababneh

Modeling and Measuring the Fundamental Period of Vibration for Low to Medium Rise Residential Buildings in Jordan Pages 83-87 Musa Resheidat, Hanan Al Nimry, Marwa Al Jamal

Behavior Geotechnical and Geologic of the Grounds Northeast Algerian Affected by the Landslides Pages 88-91 A.Saihia, M. Meksaouine

Leader Election in mobile ad hoc networks using Omega failures detector Pages 92-95 Leila Melit, Nadjib Badache

Forecasting Crude Oil Price Based on Artificial Intelligent Model: A Smoothed Feedforward Neural Network Pages 96-99 Manel Hamdi, Chaker Aloui

A new Class of Attractor Generated

elSenaof

unsiFradnobe

plsean

mth

shadhaseclM

plbitiInfrmof

usofTth

by Fractal Processes

1

Kais Bouallegue1,+, Youcef Soufi2 and Abdessattar Chaari3Department of Electrical Engineering. High Institute of Applied Sciences and Technology of Sousse, Tunisia

2Department of Electrical Engineering,University of Tebessa-Algeria.3Department of Electrical Engineering,National Engineering School of Sfax.

+Email: kais [email protected]

Abstract—In this paper, we first present the mod-ling of nonlinear numeric processes of fractal shape.cond, we study the convergence of nonlinear dy-mic numeric process resulting from a finite numberidentical processes in a cascade.

Keywords-: fractal; Attractor; Julia process.

I. Introduction

A fractal is an object that displays self-similarityder magnification and can be constructed using a

mple motif (an image repeated on ever-reduced scales).actals have generated a great deal of interest since thevent of the computer. Talking about fractals while ig-ring the dynamic processes which created them wouldinadequate.

Fractals and chaos offer a rich environment for ex-oring and modelling the complexity of nature. In somense, fractal geometry is used to describe, model andalyze the complex forms found in nature.Fractals and chaos are also linked by the fact thatany of the contemporary pacesetting discoveries ineir fields were only possible using computers.The nonlinear numeric dynamic processes of fractalape are described by iterative procedures that are wellapted to an algorithmic treatment. These processesve been studied since the end of the 19th century byveral mathematicians such as: Gaston M. Julia, Wa-aw Sierpinski, Peano, Helge von Koch (see [9] and [10],andelbrot [5], Bransley [1].The dynamic numeric processes are extensively ap-ied in several domains of science such as chemistry,ology, high energy physics, mathematics, communica-ons, energy networks and automatic control systems.this paper, we are interested in the modelling of some

actal processes. The convergence study, the transfor-ation as well as the stability analysis and the variationthese processes are proposed.In SectionII, we present definitions and terminologiesed throughout the paper. We also recall the notionsrecurrence relations, system of recurrence relations.

hese relations will be used in the third section undere name of transformations. SectionIII is devoted to

the modelling of some dynamic numeric processes. Arecursive algorithm for periodic mathematical functionsand for hyperbolic functions is studied.

In SectionIV, The convergence and the process attrac-tors are, also, treated.

We close the paper by a conclusion about the mainideas developed in this work.

II. Definitions and Terminologies

A dynamic numeric process is a mapping from Rn into

itself, that we iterate. It is described by functions bindingthe inputs to the outputs. In addition, it is bound closelyto the initial conditions of the process and to the numberof iterations. Indeed, the algorithm that characterizesthe process function provides variables which are rein-serted to be the new inputs after the execution of thefirst iteration in the output. This operation is iterative(or recursive) and can be defined as a simple relationshipwhich affects the output toward the input. It could alsobe accompanied by any transformation.

A. Recurrence relations

Let E be a nonempty set, d a nonzero natural number,f : Ed −→ E be a mapping and (Xn)

n∈Nbe a sequence

of elements of E satisfying the following relation

Xn+d = f(Xn, Xn+1, . . . , Xn+d−1). (1)

The above relation is called recurrence relation oforder d.

It is worth noting that recurrence relations are ofinterest in several applied sciences. For instance, let usconsider the signal x(t) = sin(ωt). A discretization ofthis signal may be represented by the following recur-rence relation:

Xn+1 = 2 cos(ω)Xn −Xn−1,

with initial conditions: X0 = 0 and X1 = sin(ω).

It is known that any dynamical system ruled by aquadratic polynomial behaves as if it were ruled bythe logistic equation, which is given by the followingrecurrence relation:

1 © 2014 UTA

Xn+1 = µXn(1−Xn), where µ ∈ [0, 4] and X0 ∈ [0, 1].

B. Systems of recurrence relations

1) Systems of dependent variables: Let us consider thefollowing systems S1 and S2:

(S1)

x(t) = cos(ωt)y(t) = sin(ωt)

(S2)

x(t) = cosh(ωt)y(t) = sinh(ωt)

(2)

Discretizations of systems S1 and S2 can be given asfollows:

xn+1 = α1xn − β1ynyn+1 = β1xn + α1yn

xn+1 = α2xn + β2ynyn+1 = β2xn + α2yn

(3)

with α21 + β2

1 = 1 and α22 − β2

2 = 1.

When we want to change the orbit or trajectorydirection for the states, we use the reverse system asfollows:

(S−1

1 )

xn+1 = α1xn + β1ynyn+1 = −β1xn + α1yn

(S−1

2 )

xn+1 = α2xn − β2ynyn+1 = −β2xn + α2yn

When we need to switch between systems S1 and S2

we have to insert an integer k of value k = ±1. Ifk = −1 then the system S1 is selected, else system S2 isconsidered. Let us consider the integer m permitting tochose the direction of the states(xn+1, yn+1). If m = 1the direct direction of the states else opposite directionis adopted.

In such conditions the system S becomes autonomous:

S

xn+1 = αxn + kmβynyn+1 = −mβxn + αyn

with α2 + km2β2 = 1

The parameters k and m allow to describe the systemS by a variable structure. The system S is a systemof nonlinear dynamic equations to two non separatevariables. For each iteration number i between 0 andn, if the value of xn+1 is attributed to xn and the valueof yn+1 to yn , figure 1 a, then the model is consideredas a model of a dynamic system without inverting thestates. In contrary if the value of xn+1 is attributed toyn and the value of yn+1 to xn , 1 b, then the modelis considered as a model of a dynamic system whileinverting the states.

Yi+1 Xi+1

XiYi

S(Xi,Yi)

(a) Model of a dynamic system without inverting thestates

Yi+1Xi+1

XiYi

S(Xi,Yi)

(b) Model of a dynamic system while inverting thestates

Figure 1. System of recurrence equation

A system of recurrence relation with two non separatevariables has the following structure:

xn+1 = f(xn, yn, . . . , x1, x0, y0)yn+1 = g(xn, yn, . . . , x1, x0, y0)

2) Systems of independent variables: The followingsystem 4, which is described by the recursive equationsT1 and T2, has the same structure as mentioned previ-ously (see Figure 1). It is a nonlinear dynamic systemwith two separate variables:

xn+1 = T1(xn, . . . , x1, x0)yn+1 = T2(yn . . . , y1, y0)

(4)

C. Fractals

The fractals are complex processes which have a cer-tain form of order or structure. This structure gives thema particular role from the set of the irregular shapes.

K. Bouallegue et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 1-7

2 © 2014 UTA

There are many techniques which allow us to makethe construction of classical fractals as the concept ofiterated function systems, L-systems, and the complexpolynomials or Complex Square Roots and Quadraticequations.

1) Iterated function systems: A theory of self-similarand iterated function system was developed by Hutchin-son [8] in 1981 and popularized by Barnsley [2].

An Iterated Function Systems (IFS, for short) is a sys-tem of contracting transformations defined on a metricspace.

An important class of IFS on the Euclidean space R2

is given using affine transformations:

wi(x) = Aix+Bi i = 1, 2, . . . , N, x,Bi ∈ R2,

with N is the number of transformations of the system,Bi are translation vectors and Ai are 2×2 real matriceswith eigenvalue si such that |si| < 1. For example, theSierpinski gasket is constructed from the triangle withvertices (0, 0), (1, 0) and (1

2, 1

2) by successive deletion

of the central triangles. It is a self-similar set withthe iterated function system consisting of three affinetransformations:

w1(x, y) = (1

2x, 1

2y)

w2(x, y) = (1

2x+ 1

2, 1

2y)

w3(x, y) = (1

2x+ 1

4, 1

2y + 1

2)

(5)

The Fractals as the Sierpinski Gasket and Barnsley’sFern can be generated by the concept of iterated functionsystems.

(a) Sierpinski Gasket

(b) Bransley ’s fern

Figure 2. Fractals generated by IFS

2) Complex square roots and quadratic equations:In [6], Heinz-Otto Peitgen et al. have proposed an al-gorithm which computes the model of Julia set usingcomplex square root:

Figure 3. Julia set

III. Modelling of Dynamic Numeric Processes

This section deals with the modelling of dynamicnumeric process.

A. Model of a dynamic process

In Complex Square Roots and Quadratic equationsalgorithm described in [6], the only treatment functionis the function called “square root”. Our approach con-sists in applying the same algorithm while using threefunctions f , g and h. Besides, these functions vary andrepresent the hindrances of the process; which makes thealgorithm very dynamic. Let P1 be a numeric processwith two variables. The process inputs areXi and Yi andthe initial constants Xc and Yc (initial conditions). Thecomputation is done with a = Xi −Xc and b = Yi − Ycafter the algorithm treatment. The generated outputsXi+1 and Yi+1 are reinserted into Xi and Yi for thenext iteration. Figure 4 illustrates our proposed modelof the process. The graph is constituted with a circlewhich represents the internal treatments modelling. Theinputs ( Xi, Yi, Xc and Yc) are associated to this circle.In the graph, the circle arcs represent an affectation ofXi+1 to Xi and Yi+1 toYi.

The algorithm has the following structure: The itera-tion value i varies between 0 and n, where n is equal to106. The initial conditions are x0 = −1, y0 = 1, xc = −1and yc = 0.

Yg Xg

XiYi

P1

+−

+−

Xc

Yc

Figure 4. Model of the process

The process treats a random function, which allowsthe switching between the following three cases:


3 © 2014 UTA

Algorithm 1 The algorithmic of Processes

Ensure: Calculate (yi+1, xi+1) = P (xi, yi, f, g, h)1: if xi < xc then

2: xi+1 = g[f [(xi − xc)2 + (yi − yc)2] + xi2

]3: yi+1 = yi−yc

2xi+1

4: end if

5: if xi = xc then

6: xi+1 =√

|yi−yc|2

7: if xi > 0 then

8: xi+1 = yi−yc2yi+1

9: end if

10: if xi < 0 then

11: yi+1 = 012: end if

13: end if

14: if xi > xc then

15: yi+1 = h[f [(xi − xc)2 + (yi − yc)2]− xi2

]16: xi+1 = yi−yc

2xi+1

17: if yi < yc then

18: yi+1 = −yi+1

19: end if

20: end if

The following table presents 10 studied cases whilepreserving the same starting values and the same num-ber of iterations. In this computation, the initial condi-tions are x0 = −1, y0 = 1, xc = −1 and yc = 0.

Process f(x) g(x) h(x) Figure:P1

√x

√x

√x Fig: 7(a)

P2 arctanx√x

√x Fig: 5(b)

P3 tanh x√x

√x Fig: 5(c)

P4

√x arctanx

√x Fig: 5(d)

P5 arctanx arctanx expx Fig: 5(e)P6

√x

√x arctan Fig: 5(f)

P7

√x tanhx arctanx Fig: 5(g)

P8

√x tanhx

√x Fig: 5(h)

P9

√x tanhx tanhx Fig: 5(i)

P10

√x arctanx coshx Fig: 5(j)

(a) P1(b) P2

(c) P3 (d) P4

(e) P5

(f) P6

(g) P7(h) P8

(i) P9(j) P10

Figure 5. Multi Processes

It seems interesting to ask the following questions:

– Can this kind of process be applied to control, tofilter, to observe, to treat nonlinear systems as well asto analyze nonlinear natural phenomena?

–What are their limits in these cases?

To answer these questions, it is necessary to study andapply some new methods such as the convergence, theparameter variation and the transformation in nonlinearspace. This is the goal of the section that follows.


4 © 2014 UTA

IV. A new class of attractors

To analyze the algorithm performances, we have ap-plied it to the convergence and the transformation stud-ies.

A. Convergence

A process is said to be in cascade if the outputs ofprevious iterations are inputs for the next ones.

Definition IV-A.1: Let P be a dynamic process. Sup-pose that there exists n ∈ N \ 0, 1 such that theassociation result of P in n cascades is identical to thatof P in (n + 1) cascades. Then a such n is called theconvergence order of P .

The following graph represents the interconnection incascade of k identical processes, P1

XiYi

P1

+−

+−

Xc

Yc

P1

...

P1

Figure 6. Processes in “n cascades”

In what follows, we give three examples of convergenceorder of some processes (P2, P5 and P10):

Example IV-A.2:

Process n: number of cascades Figure:P2 n = 1 Fig: 7(a)P2 n = 2 Fig: 7(b)P2 n = 3 Fig: 7(c)P2 n = 4 Fig: 7(d)P2 n = 5 Fig: 7(e)P2 n = 6 Fig: 7(f)P2 n = 7 Fig: 7(g)P2 n = 8 Fig: 7(h)P2 n = 9 Fig: 7(i)P2 n = 10 Fig: 7(j)P2 n = 11 Fig: 7(k)

(a) n=1 (b) n=2 (c) n=3

(d) n=4 (e) n=5 (f) n=6

(g) n=7 (h) n=8 (i) n=9

(j) n=10 (k) n=11

Figure 7. Convergence of Processes P2

Example IV-A.3:


5 © 2014 UTA

Process n cascades Figure:P5 n = 2 Fig: 8(b)P5 n = 3 Fig: 8(c)P5 n = 4 Fig: 8(d)P5 n = 5 Fig: 8(e)P5 n = 6 Fig: 8(f)P5 n = 7 Fig: 8(g)P5 n = 8 Fig: 8(h)P5 n = 9 Fig: 8(i)P5 n = 10 Fig: 8(j)

(a) n=1 (b) n=2 (c) n=3 (d) n=4

(e) n=5 (f) n=6 (g) n=7 (h) n=8

(i) n=9 (j) n=10


Example IV-A.4:

Process n cascades Figure:P10 n = 2 Fig: 9(a)P10 n = 3 Fig: 9(b)P10 n = 4 Fig: 9(c)P10 n = 5 Fig: 9(d)P10 n = 6 Fig: 9(e)P10 n = 7 Fig: 9(f)P10 n = 8 Fig: 9(g)P10 n = 9 Fig: 9(h)P10 n = 10 Fig: 9(i)P10 n = 11 Fig: 9(j)P10 n = 12 Fig: 9(k)P10 n = 13 Fig: 9(l)P10 n = 14 Fig: 9(m)

(a) n=1 (b) n=2 (c) n=3 (d) n=4

(e) n=5 (f) n=6 (g) n=7 (h) n=8

(i) n=9 (j) n=10 (k) n=11 (l) n=12

(m) n=13 (n) n=14


We present a table for order of convergence of someprocesses:


6 © 2014 UTA

Process Oder of convergenceP1 n = 7P2 n = 11P3 n = 10P4 n = 14P5 n = 10P6 n = 10P7 n = 11P8 n = 10P9 n = 12P10 n = 14

V. Conclusion

We have shown, in this paper, that the dynamicprocess is analyzed as the reinsertion of the outputsagain to the inputs. The processes are based on periodicor variant mathematical functions. The initial conditionsand the constants act or influence the process behaviour.The convergence is done while setting into cascade thesame process. To each process’s attractor, we have asso-ciated a number which represents the convergence orderof the process.

References

[1] M. F. Barnsley, S. Demko: Iterated function systems and the global construction of fractals. Proc. Roy. Soc.

Lon-don Ser. A. 399, 243-275, 1985.

[2] M. F. Barnsley: Fractals everywhere. Academic Press. Inc. Boston. MA., 1988.

[3] Huang Dong-Wei, Gao Qin, Wang Hong-Li, Jian-FengFeng, Zhi-Wen Z hu: On chaotic motion of some stochas-tic nonlinear dynamic system. Chaos, Solitons and Frac-tals 31, 242-246, 2007.

[4] Z. Jing ,Z. Yang ,T. Jiang: Complex dynamics in Duffi- ng Van der Pol equation. Chaos, Solitons and Fractals 27,

722-747, 2006.

[5] B. Mandelbrot: The fractal geometry of nature. Freemanand Co. San Francisco. Calif., 1982.

[6] H. O. Peitgen, H. Ju¨rgens, D. Saupe: Chaos andfractals. New frontiers of science. New York. Springer-Verlag, 1992.

[7] S. L. Singh, P Bhagwati, K Ashish: Fractals via iteratedfunctions and multifunctions. Chaos Solitons and Frac-tals., 2007.

[8] J. E. Hutchinson: Fractals and self-similarity.

IndianaUniv. Math. J. 30, 713-747, 1981.

[9] H. von Koch: Une méthode géométrique élémentaire pourl’étude de certaines questions de la théorie des courbes planes. Acta Math. 30, 145-174, 1906.

[10] H. von Koch: Sur une courbe continue sans tangente,obtenue par une construction geóme´trique e´le´mentaire. Ark. Mat. 1, 681-704, 1904.


7 © 2014 UTA

Abstract-- This paper presents solution of optimal power flow (OPF) problem of a power system via Differential Evolution (DE) algorithm. The purpose of an electric power system is to deliver real power to the greatest number of users at the lowest possible cost all the time. So the objective is to minimize the total fuel cost of the generating units and also maintaining an acceptable system performance in terms of limits on generator reactive power outputs, bus voltages, Static VAR Compensator (SVC) parameters and overload in transmission lines. CPU times can be reduced by decomposing the problem in two subproblems, the first subproblem minimize the fuel cost of generation and the second subproblem is a reactive power dispatch so optimum bus voltages can be determined and reduce the losses by controlling tap changes of the transformers and the static Var Compensators (SVC). To verify the proposed approach and for comparison purposes, we perform simulations on the Algerian network with 114 buses, 175 branches (lines and transformers) and 15 generators. The obtained results indicate that DE is an easy to use, fast, robust and powerful optimization technique compared to the other global optimization methods such as PSO and GA.

Keywords: Economic Power Dispatch, Optimal Power Flow, Differential Evolution, FACTS Device, Optimization, Algerian Network.

I. INTRODUCTION The optimal power flow (OPF) can be defined as a typical

flexible nonlinear programming problem with many objectives. Because of the increasing s cale and co nstraint n umber o f electric power system, the OPF has been a complicated large-scale m athematic p rogramming p roblem. Th e op timal p ower flow (OPF) calculation optimizes the static operating condition of a power generation-transmission system. The main benefits of optimal power flow are (i) to ensure static security of quality of ser vice b y im posing lim its on generation-transmission system’s o peration, (ii) t o optimize reactive-power/voltage scheduling and (iii) to improve economy of operation through the full utilization of the system’s feasible operating range and by the accurate co ordination o f tran smission losses in th e scheduling process. The OPF h as b een usually con sidered as the m inimization o f an o bjective fu nction representing the generation cost and/or the tr ansmission lo ss. Th e co nstraints involved are the physical laws governing the power generation-

This wor k w as supported in par t by the Algerian Ministry of Higher Education and Scientific Research, grant number J0201220100003.

transmission system s an d the o perating li mitations of th e equipment.

The optimal power f low has been f requently so lved using classical optimization methods. Effective optimal power flow is limited by (i) the high dimensionality of power systems and (ii) the incomplete domain dependent knowledge of power system engineers [1][2] [3].

In recent years, energy, environment, deregulation of power utilities have d elayed th e con struction o f bo th gen eration facilities and new t ransmission li nes. Better u tilization o f existing p ower system cap acities by installing flexible A C transmission s ystems FACTS dev ices has beco me imperative. The ap plication o f Flexib le Alternative Cur rent Transmission Systems (FACTS) in electric power system, such as Thyristor Controlled Series Compensations (TCSC), Thyristor controlled phase angle Regulators (TCPR), Un ified Po wer Flow Controllers ( UPFC) an d Stat ic Var Co mpensator (SVC), i s intended for t he c ontrol of pow er flow, improvement of stability, voltage profile management, power factor correction, loss minimization, and reduced co st o f p roduction. Th e OPF becomes even more complex when FACTS devices are taken into consideration as control variables.

It can be seen that the generalised OPF is a non-linear, no-convex, larg e-scale, s tatic o ptimization p roblem with both continuous and discrete con trol v ariables. Ap plications o f conventional optimisation t echniques su ch as the gradient-based algorithms are n ot g ood en ough to s olve th is pr oblem. Because it depends on the existence of the first and the second derivatives of the objective function and on the well computing of these derivative in large search space.

A new flo ating po int en coded ev olutionary alg orithm fo r global optimization an d n amed it Diff erential Ev olution (DE) was p roposed by Stor n and Price [4], and since t hen the DE algorithm has been used in many practical cases. The original DE was modified, and many new versions proposed.

Generally DE is characterized as a simple heuristic of well-balanced m echanism with flex ibility to en hance an d adapt to both global an d lo cal ex ploration ab ilities. Th e ef fectiveness, efficiency and robustness of the DE algorithm are sensitive to the settings of the control parameters. The best settings for the control parameters depend o n th e fu nction and r equirements for co nsumption time an d accu racy. It has g ained a lot of attention in var ious po wer system ap plications. It is a population based method and an improved version of GA using similar operators: mutation, crossover and selection. The main

Application of Differential Evolution Algorithm to Optimal Power Flow incorporating FACTS: a

case study Linda Slimani and Tarek Bouktir

Department of Electrical Engineering, Setif University, Algeria [email protected], [email protected]

8 © 2014 UTA

difference in constructing better solutions is that GA relies on crossover while DE relies on mutation operation. The mutation operation is used as a search mechanism, which is based on the differences of randomly sampled p airs of so lutions in th e population. The algorithm uses selection operation to direct the search towards the prospective regions in the search space [5].

This paper proposes a simple app roach b ased o n DE algorithm implemented in C++ Build er to minimize the t otal fuel cost o f the thermal generating un its and also maintaining an acceptab le system p erformance in ter ms o f lim its o n generator reactive p ower outputs, bu s v oltages, Stati c VAR Compensator ( SVC) p arameters and overload i n transmission lines. CPU times can be red uced by decomposing the problem in two subproblems, the f irst su bproblem m inimize the fuel cost of g eneration an d the seco nd su bproblem is a reactiv e power dispatch so opt imum bu s v oltages c an be de termined and reduce the losses b y co ntrolling tap r atio o f th e transformers and the static Var Compensators (SVC).

To v erify the p roposed ap proach and fo r comparison purposes, w e perform simulations on t he Alg erian network with 114 buses, 175 branches (lines and transformers) and 15 generators. The obtained results indicate that ED is an easy to use, fast, robust and powerful optimization technique compared to other global Optimization methods such as PSO, and GA.

II. PROBLEM FORMULATION In OPF, the generators are modelled as voltage controlled

buses an d lo ads as l oad bu ses. On e gen erator s erves as the slack bus. The standard OPF problem can b e formulated as a constrained optimisation problem as follows:

min ( , ). . ( , ) 0

( , ) 0

f x us t g x u

h x u=≤

(1)

where f(x,u) is the objective function, g(x,u) represents the equality co nstraints, h(x,u) rep resents t he inequality constraints, x is the vector of the dependent variables such us the voltage and angle of load buses and u is the vector of the control v ariables su ch as g enerator real p ower Pg, ge nerator voltages Vg, t ransformer t ap se tting T, an d the reactance of dynamic shunt capacitors/reactors SVCB . Therefore, u can be expressed as

[ ], , , Tg g SVCu P V t B= (2)

A. Objective Function The Opt imal po wer flow pr oblem with consideration of

FACTS d evices can be decomposed in two sub -problems which a re the E conomic p ower Di spatch and the R eactive Power Flow combined with FACTS devices..

A.1 Economic Objective Function The essence of the optimal power flow problem resides in

reducing the objective function and simultaneously satisfying

the load f low eq uations ( equality co nstraints) witho ut violating the inequality constraints

The m ost com monly u sed o bjective in t he OPF problem formulation is the minimisation o f the to tal operation cost o f the fuel co nsumed for producing electric p ower within a schedule time interval (one hour). The individual costs of each generating unit are assumed to be function, only, of real power generation and are represented by quadratic curves of second order. The objective function for the entire power system can then b e ex pressed as t he s um o f th e q uadratic co st model at each generator [6-7].

( ) ( )2

1

ngec i i i i i

iF x Pg Pgα β γ

== + +∑ $/h (3)

where iα , iβ and iγ are the cost coefficients of generator at bus i.

a) Active Power Transmission Losses and Voltage Deviation Objective Function

The objective is to minimise the active power losses in the transmission network and/or the voltage deviations at the load buses inv olving r eactive p ower co ntrols, while f ixing activ e power controls.

The tap changers of the transformers and SVC can control the reactive p ower flo w so optimum bus voltages can be determined and reduce the lo sses. T he s hunt FACTS d evice should be placed on the most sensitive buses. The insertion of SVC enhances t he vol tages a t v arious bu ses, a nd r eduction power loss of t he s ystem. Fo r S VC, i t c an pr ovide r eactive power and v oltage su pport. A s a r esult, the r eactive power generation of SVC becomes one of the control variables. One of t he i mportant i ndices of po wer s ystem se curity i s t he bu s voltage magnitude. The voltage magnitude deviation from the desired value at each load bus must be as small as possible.

The active power transmission losses ( lossP ) is given by:

( )2 2

12 cos

lNloss k k i j k i j ij

kP g t V V t V V θ

=⎡ ⎤= + −∑ ⎣ ⎦ (4)

where lN is number of branch on the network, t equal =1 if th e b ranch is a tran smission line and t equ al the tap ratio value if t he br anch is a transformer, . k is a br anch with conductance g connecting the ith bus to the jth bus.

The deviation of voltage is given as follows:

1

PQNdes

k kk

V V V=

Δ = −∑ (5)

where PQN i s the number of load buses and deskV is the

desired or target value of the voltage magnitude at load bus k.

b) The total objective function of OPF problem The e quation of t he t otal obj ective fu nction using into

account the Economic Pow er Disp atch (ED) objective

L. Slimani and T. Bouktir Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 8-15

9 © 2014 UTA

function; active pow er t ransmission losses ( lossP ); an d th e sum of the normalized vio lations of vo ltages ( ViF ) is as follow:

ED l loss V Vf F P Fω ω= + + (6)

where

( ) ( )lim max min

1

PQN

V PQj PQj PQj PQjj

F V V V V=

= − −∑

lω and Vω constants a re re lated to l ine loss and voltage deviation. These constants were found as a result of trials.

A.2 Types of Equality Constraints While minimising the objective function, it is necessary to

make sure that the gen eration sti ll supplies the load demands plus losses in transmission lines. Th e equality constraints are the power flow equations d escribing bus in jected acti ve an d reactive may be defined as follows:

( )1

cos sinnb

i i i i j ij ij ij ijj

P Pg Pd VV g bθ θ=

= − = +∑ (7)

( )1

sin cosnb

i i i i j ij ij ij ijj

Q Qg Qd VV g bθ θ=

= − = −∑ (8)

where iPg , iQg ar e th e active an d reactive power

generation at bus i; iPd , iQd are the real and reactive power

demands a t bus i ; iV , jV , the voltage m agnitude at bu s i,j ,

respectively; ijθ i s the ad mittance an gle, ijb an d ijg are the real an d imaginary p art o f th e ad mittance an d n b i s the total number of buses.

The equality constraints are sati sfied by ru nning Newton-Raphson algorithm.

A.3 Types of Inequality Constraints The inequality constraints of the OPF reflect the limits on

physical devices in the power system as well as t he li mits created t o en sure system secu rity. Th e m ost u sual types o f inequality constraints are u pper b us voltage lim its at generations and load b uses, lower bus vo ltage l imits at l oad buses, var. limits at generation buses, maximum active power limits cor responding to l ower lim its at some generators, maximum lin e lo ading lim its and limits o n t ransformer tap setting.

The i nequality con straints on the p roblem variables considered include:

• Upper and lower bou nds on t he a ctive ge nerations a t generator buses Pgi

min≤ Pgi ≤ Pgimax , i = 1, ng.

• U pper and l ower bo unds on the reactive power generations at generator buses Qgi

min≤ Qgi≤ Qgimax , i = 1, ng

• Upper an d lower b ounds o n reactive power injection at buses with VAR compensation Qci

min≤ Qci≤ Qcimax, i= 1, nc

• Upper and lower bounds on the voltage magnitude at the all buses . Vi

min≤ Vi ≤ Vimax , i = 1, nb.

• Upper and lower bounds on the bus voltage phase angles θi

min≤ θi ≤ θimax , i = 1, nb.

• f or secu re o peration, the transmission line load ing Sl is restricted by its u pper limit as:Sli≤ Sli

max , i = 1, nl, where Sli, Sli

max are stand for the power of transmission line and limit of transfer capacity of transmission line and nl is the number of transmission lines.

The constraints o n the stat e v ariables can b e t aken i nto consideration by adding pe nalty f unction t o the obj ective function.

B.1 Application of FACTS in Electric Power System The purpose of the transmission network is to pool power

plants an d lo ad centres in order to su pply th e load at a required reliability an d m aximum ef ficiency at a lo wer co st. As p ower tr ansfer g row, the p ower system can beco me increasingly more difficult to operate, and the system becomes more i nsecure w ith u nscheduled po wer flows and hi gher losses. I n this co ntext, a co ncept call ed a flexib le alternative current tran smission system w as introduced. The conception of f lexible ac tran smission syst ems (FA CTS) as a total network c ontrol philosophy was f irst i ntroduced by N. G. Hingorani [8] f rom t he El ectric po wer research institute (EPRI) i n t he U SA i n 1 988, a lthough t he pow er electronic controlled devices had been used in the transmission network for many years before that.

The ap plication of FA CTS in electric po wer syste m i s intended for t he c ontrol of pow er f low, improvement of stability, voltage profile management, power factor correction, and loss minimization [9-12]. Power flow through an ac line is a fu nction o f ph ase ang le, line and voltages and line impedance. The co nsequences o f lack co ntrol over any of these variables are problems with stab ility, undesirable power flows, undesirable Var fl ows, hi gher losses, hi gh or l ess voltage a nd a mong t he ot hers; wi th F ACTS devices we can control the phase angle, the magnitude at chosen bus and line impedance.

Thyristor Controlled Series Capacitors (TCSC) and Static Var Compensators (SVC) are the most popular devices of the FACTS [13]. The main functionality of the SVC is to regulate the voltage at a chosen bus by controlling the reactive power injection at th e location. Main taining the rated voltage levels is im portant f or p roper op eration an d utili zation of loads. Under voltage causes deregulation in the performance of loads such as in duction m otors, l ight bu lbs, et c. W hereas ov er voltage c auses m agnetic s aturation a nd re sultant ha rmonic generation, a s we ll a s e quipment failures due to insulation breakdown. Th ese d evices ar e characterized by rapid response, wide operational range and high reliability.

B.2 Modeling of Static VAR Compensator Thyristor controlled Static V AR Co mpensators ( SVCs)

were d eveloped in the 19 70s to act as com pensation for arc furnaces, these devices are one of the earliest types of Flexible AC T ransmission System (FACTS) co ntrollers. T he ty pical shunt connected SVC consists of thyristor controlled reactors and thyristor switched capacitors. The full continuous range of


10 © 2014 UTA

the SVC can b e accessed by coordinating the switching of the discrete capacitor block and t he c ontinuous re actor c ontrols [14].

The SVC is usually operated in a voltage regulating mode, which ad justs its s usceptance to maintain the local transmission network voltage to a voltage setpoint value. The SVC c an a lso ope rate i n a constant MVAr mode, which maintains a f ixed v alue o f susceptance under steady state conditions. The effect of the SVC controller on the economic operation and voltage stability of the network is the p rinciple motivation be hind i ncorporating t he S VC i nto various formulations. In th is stud y, w hen th e SV C is installed in the transmission lin e, it can be treated a s a PV bu s with the generation o f r eal power as 0. Th e following algebraic equation gives the reactive power injected at the SVC bus i:

2/svcQ B V= (9)

The reactance svcB is locked if one of its limits is reached.

B. Application of DE Algorithm on OPF Problem DE is a d irect search method using operators: mutation,

crossover a nd s election. T he a lgorithm r andomly c hooses a population vector of fixed si ze. Dur ing e ach iteration of algorithm a new population of same size is generated. It uses mutation operation as a sear ch m echanism. This o peration generates ne w pa rameter vector by a dding a weighted difference vector between two population members to a third member. In o rder to incr ease the d iversity o f the parameter vectors, the crossover operation produces a trial vector which is a combination of a mutant vector and a parent vector. Then the s election o peration directs th e search toward the prospective r egions in t he search sp ace. In addition, the best parameter vector is evaluated for every generation in order to keep tr ack o f the p rogress that is m ade during the minimization process. Th e ab ove i terative process of mutation, crossover an d selection o n t he p opulation will continue u ntil a u ser-specified sto pping criterio n, normally, the maximum number of generations or the maximum number of function ev aluations allowed is m et. Th e pr ocess i s assumed to have converged if the dif ference between the b est function v alues i n t he ne w a nd ol d pop ulation, a nd the distance between the new best point and the old best point are less than the specified respective tolerances. The other type of stopping c riterion c ould be if t he g lobal minimum of the problem is k now a- priori. Then DE will be terminated i f the difference b etween the best fu nction value in th e n ew population and t he kn own gl obal m inimum i s l ess t han t he user defined tolerance level [5].

Initialisation of Vectors

Difference vector based mutation

Crossover/ Recombination

Selection

Figure 1. Main stages of the DE algorithm.

DE is a sim ple r eal param eter opt imization algo rithm. I t works through a simple cycle of stages, presented in Fig. 1.

B 1) Differential Evolution optimization process Differential Evolution uses a population P o f size NP that

evolves over G generations to reach the optimal solution. Each individual Xi is a vector that contains as many parameters as the problem decision variables D.

( ) ( ) ( )1 ,........,G GG

NpP X X⎡ ⎤= ⎣ ⎦ (10)

( ) ( ) ( )1, ,,........, 1, ,

TG G Gpi i D iX X X i N⎡ ⎤= =⎣ ⎦ … ( 11)

The population size N P is an alg orithm control parameter selected by t he us er whi ch re mains constant throughout the optimization process. The optimization process in Differential Evolution i s carried out using t he t hree basic operations: Mutation, Crossover and Selection.

The main steps of the DE algorithms are given below: Initialization Evaluation Repeat Mutation Crossover Evaluation Selection Until (termination criteria are met)

• Initialization

At the early stag e o f DE search, i.e., t = 0, the algorithm starts by creating an initial population of NP vectors.

The pr oblem i ndependent variables are initialized somewhere in their feasible numerical range in every vector as follows.

( )(0) min max min, (0,1)j j jj iX X rand X X= + ⋅ − (12)

Where 1, ....., Pi N= and 1,.....,j D= ; minjX a nd max

jX are the lower and upper bounds of the jth decision parameter; and ( 0,1)rand is a un iformly distributed r andom n umber

within [ 0, 1 ] gen erated fo r each v alue o f j . )0(,ij

X is th e j th

parameter of the ith individual of the initial population.

• Mutation The mutation o perator creates m utant v ectors ( )'

iX by perturbing a randomly selected vector Xa with the difference of two other randomly selected vectors Xb and Xc

( )'( ) ( )( ) ( )G GG Ga ci bX X F X X= + − 1,..., Pi N= (13)

Where Xa Xb and Xc are randomly chosen vectors among the Np population, and a b c i≠ ≠ ≠ . The scaling constant F is an algo rithm co ntrol parameter u sed to ad just the perturbation size i n the mutation o perator a nd t o i mprove algorithm convergence. Typical value of F is in the range of 0.4–1.0.

• Crossover

Two t ypes of c rossover s chemes c an be used by D E algorithm. T hese a re e xponential c rossover and binomial crossover. Although the exponential crossover was presented


11 © 2014 UTA

in the or iginal work of Storn a nd P rice [4], t he b inomial variant is much more used in recent applications.

In exponential type, the crossover operation generates trail vectors ( )''

iX by mixing the parameters of t he mutant vectors

( )'iX with t he targ et vecto r ( )iX according to a selected

probability distribution,

( )

⎪⎩

⎪⎨⎧ =≤

=otherwiseX

qjorCifXX

Gij

RjGijG

ij ,

,)(

,

')(',"

,

η

(14 )

Where PNi ,.....,1= and Dj ,.....,1= ; q is a r andomly chosen index PN,.....,1∈ that guarantees that the trail vector

gets at least o ne p arameter fr om the m utant vector; 'jη is a

uniformly distributed random number within [0, 1] generated for each value of j.The crossover constant CR is an algorithm parameter that controls the diversity of the population and aids the alg orithm t o escap e fr om local minima. )('

,)(

, , Gij

Gij XX and

)('',GijX ar e t he jth parameter of the ith target v ector, m utant

vector and trail vector at generation G, respectively.

• Selection

To keep th e population siz e con stant ov er subsequent generations, the selection process is car ried out to d etermine which one of the child and the parent will survive in the next generation

The selection operation forms the population by choosing between th e tr ail vectors an d th eir predecessors (target vectors) th ose ind ividuals that p resent a better fitness or are more optimal according to (18).

( ) ( )PG

i

Gi

Gi

GiG

i NiotherwiseX

XfXfifXX ,......,1,

)(

)()"()"()!( =

⎪⎩

⎪⎨⎧ ≤

=+

(15) This p rocess is rep eated for s everal gen erations al lowing

individuals to improve their fitn ess as they ex plore th e solution space in search of optimal values.

DE has thr ee essential c ontrol par ameters: t he scaling factor (F), the crossover constant (CR) and the population size (NP). The s caling factor is a value in the ran ge [ 0, 2 ] th at controls th e am ount of per turbation in the m utation process. The cr ossover co nstant is a v alue in the r ange [0 ,1] that controls the di versity of t he po pulation. T he pop ulation size determines t he num ber of i ndividuals i n the population and provides the algorithm enough diversity to search the solution space.

Proper selection of con trol par ameters i s ver y im portant for al gorithm su ccess an d performance. The optimal con trol parameters are problem specific. Therefore, the set of control parameters th at best fit each p roblem h ave to b e chosen carefully. The most common method used to select the control parameter i s par ameter tu ning. Param eter t uning adjusts the control p arameters th rough testi ng u ntil t he b est setti ngs ar e determined. Typ ically the fo llowing ran ges ar e good initial

estimates: [15]: F= [0.5, 0.6], CR= [0.75, 0.90] and NP= [3D, 8D].

In order to avoid premature convergence, F o r N P should be incr eased, or C R should be decreased. Larger values of F result in larger perturbation and better probabilities to escap e from local optima, while lower CR preserves more diversity in the population thus avoiding local optima.

B 2) 5. DE Implementation for OPF While applying DE to s olve the OPF p roblem, th e

following issues need to be addressed.

1. Representation of the problem variables and

2. Formation of the evaluation function.

These two issues are described in this section.

a) Problem Representation Each v ector in the DE p opulation represents a can didate

solution o f the giv en pr oblem. The e lements of that solution consist o f al l th e op timization variables of t he problem. For the case o f minimization o f cost the generator active powers are the optimization variables. For the reactive power planning problem under consideration, gen erator term inal v oltages ( )giV t he t ransformer t ap po sitions (t k) an d th e Capacitor settings (Q Ci) are t he op timization variables. Gen erator bus voltage is represented as floating point numbers, whereas the transformer tap position and r eactive p ower g eneration o f capacitor are represented as integers.

b) Evaluation Function Differential evolution searches for the o ptimal solution by

maximizing a g iven f itness f unction, and therefore an evaluation function which provides a measure of the quality of the problem solution must b e p rovided. Th e o bjective is to minimize the total co st while satisf ying all con straints. Th e equality constraints are sa tisfied by runni ng t he Newton Raphson power flow algorithm. The inequality constraints on the cont rol variabl es are taken into ac count in t he problem representation itself, and the constraints on the state variables are tak en into consideration by a dding a quadratic p enalty function to the objective function. W ith t he i nclusion of penalty function the new objective function becomes,

1

1 1 1

PQN N N

j j jj j j

Min F f SP VP QP LPτ

= = =

= + + + +∑ ∑ ∑ (16)

Here, SP , VP j ,QP j a nd L Pj are the p enalty terms f or th e reference bus generator active power limit violation, load bus voltage li mit violation; r eactive p ower g eneration lim it violation a nd line fl ow l imit violation respectively. T hese quantities are defined by the following equations:

SP =

( )( )

⎪⎪⎩

⎪⎪⎨

⎧

<−

>−

otherwisePPifPPK

PPifPPK

sssss

sssss

0

min2min

max2max

(17)


12 © 2014 UTA

VPj =

⎪⎪⎩

⎪⎪⎨

⎧

<−

>−

otherwiseVVifVVK

VVifVVK

jjjjv

jjjjv

0

)(

)(min2min

max2max

(18)

QPj =

⎪⎪⎩

⎪⎪⎨

⎧

<−

>−

otherwiseQQifQQK

QQifQQK

jjjjq

jjjjq

0

)(

)(min2min

max2max

(19)

LPj = ⎪⎩

⎪⎨⎧ >−

otherwiseLLifLLK jjjjl

0

)( max2max

(20)

Where, Ks, Kv, Kq and Kl are the p enalty factors. Since DE maximizes the fitness f unction, th e m inimization objective function f is transformed to a fitness function to be maximized as,

Fitness = k F (21)

Where k is a large constant.

III. APPLICATION STUDY The OP F us ing DE m ethod has be en de veloped and

implemented by the use of C++Build er2009 so ftware, tested with I ntel Pe ntium D ual C PU 2220, 2. 4 GH z, 2GB RAM. Consistently acceptable r esults wer e observed. I nitially, several ru ns are d one wit h d ifferent v alues of DE key parameters su ch as d ifferentiation (o r m utation) con stant F, crossover constant CR, size of po pulation NP, and maximum number of generations GEN wh ich is used here as a stopping criteria. In this p aper, t he fo llowing v alues ar e selected as: F=0.9: C R=0.9; NP =30; GEN=50. T he pro posed method is applied to two test systems.

Application on the Algerian Network: The DE-OPF has also been tested on t he Alg erian

network. It consi sts of 114 buses, 15 generators, 159 transmission li nes an d 16 tran sformers (F ig. 1 ). Th e table I shows the technical and eco nomic p arameters o f 15 ten generators of the Algerian electrical networ k. Kn owing th at the generator of the bus of N°=04 is the slack bus. The voltage of generator buses and load buses in the system are between [1, 1.1] and [0.90, 1.1], respectively.

In this test, in o rder to r educe th e CPU tim e because the Algerian network i s relati vely larg e, the OPF problem is decomposed i n t wo s ubproblems, t he fi rst subproblem minimize t he fu el cos t o f g eneration and environmental pollution and th e seco nd subproblem is a reactive power dispatch so optimum bus vo ltages can be determined and reduce the l osses by con trolling gen erator voltages, tap ratio of the transformers and the static Var Compensators (SVC).

Figure 2. : the topologies of the Algerian Network

TABLE I. . POWER GENERATION LIMITS AND COST COEFFICIENTS FOR ALGERIAN Network.

B u s N u m b e r

P m in [M W ]

P m a x [M W ]

a [$ /h r ]

b [$ /M W h r]

c [$ /M W 2 h r ]

4 1 3 5 . 0 0 0 0 1 3 5 0 0 1 .5 0 0 0 0 .0 0 8 5 5 1 3 5 . 0 0 0 0 1 3 5 0 0 1 .5 0 0 0 0 .0 0 8 5 1 1 1 0 . 0 0 0 0 1 0 0 0 2 .5 0 0 0 0 .0 1 7 0 1 5 3 0 . 0 0 0 0 3 0 0 0 2 .5 0 0 0 0 .0 1 7 0 1 7 1 3 5 . 0 0 0 0 1 3 5 0 0 1 .5 0 0 0 0 .0 0 8 5 1 9 3 4 . 5 0 0 0 3 4 5 0 2 .5 0 0 0 0 .0 1 7 0 5 2 3 4 . 5 0 0 0 3 4 5 0 2 .5 0 0 0 0 .0 1 7 0 2 2 3 4 . 5 0 0 0 3 4 5 0 2 .5 0 0 0 0 .0 1 7 0 8 0 3 4 . 5 0 0 0 3 4 5 0 2 .5 0 0 0 0 .0 1 7 0 8 3 3 0 . 0 0 0 0 3 0 0 0 2 .5 0 0 0 0 .0 1 7 0 9 8 3 0 . 0 0 0 0 3 0 0 0 2 .5 0 0 0 0 .0 1 7 0 1 0 0 6 0 . 0 0 0 0 6 0 0 0 2 .0 0 0 0 0 .0 0 3 0 1 0 1 2 0 . 0 0 0 0 2 0 0 0 2 .0 0 0 0 0 .0 0 3 0 1 0 9 1 0 . 0 0 0 0 1 0 0 0 2 .5 0 0 0 0 .0 1 7 0 1 1 1 1 0 . 0 0 0 0 1 0 0 0 2 .5 0 0 0 0 .0 1 7 0

TABLE II. : COMPARISON OF THE RESULTS OBTAINED BY GLOBAL METHODS OF 114 ALGERIAN ELECTRICAL NETWORK

Pmin

[MW] GA PSO DE Pmax

[MW]Pg4(MW) 135.0000 515.11 515.8825 462.3908 1350 Pg5(MW) 135.0000 241.9 441.411 1 459.5589 1350 Pg11(MW) 10.0000 99.9 100.0000 99.9431 100 Pg15(MW) 30.0000 135.07 186.9059 192.5196 300 Pg17(MW) 135.0000 674.04 449.1401 453.0142 1350 Pg19(MW) 34.5000 163.76 206.6362 196.6569 345 Pg52(MW) 34.5000 211.16 190.3105 189.0239 345 Pg22(MW) 34.5000 277.06 177.8684 193.9372 345 Pg80(MW) 34.5000 228.37 224.2734 192.1215 345 Pg83(MW) 30.0000 182.49 188.7075 188.1283 300 Pg98(MW) 30.0000 153.95 192.8819 189.0847 300 Pg100(MW) 60.0000 598.41 600.0000 599.9752 600 Pg101(MW) 20.0000 197.54 200.0000 199.9703 200 Pg109(MW) 10.0000 98.11 99.7997 99.9909 100 Pg111(MW) 10.0000 39.46 100.000 0 99.9415 100 Ploss(MW) 89.345 87.9052 89.2570 Cost[$/hr] 19203 19235 19203.34 Time (sec.) 290 70 75

The comparisons of the results ob tained b y the p roposed approach DE, wit h th ose found b y GA and PSO algorithms are reported in the Table II. In this case we m inimize the fuel cost g eneration us ing into account t he c ontrol ve ctor composed only o f th e act ive po wers o f t he g enerators. Th e results obtained with the pr oposed a pproach are be tter t han


13 © 2014 UTA

those obtained by PSO and are very comparable to the results obtained by GA. The DE gives a more important profit in fuel cost of 19203,34$/h compared to the result obtained from PSO (19235 $/h) and are equal to the results of GA. The optimum value has been o btained at a co mparable tim e (7 0 sec) compared to the execut ion tim e of PSO (75 sec) with 250 iterations. T he DE is v ery fas t than GA since the GA execution tim e is ab out 2 90 s ec. A b etter co st value can be found by DE with 300 iterations which is 19202.86 $/hr.

Fig. 3 shows the typical convergence characteristics for the

best so lutions, wor st so lutions an d the average solutions obtained fo r each g eneration. It can b e seen that the convergence is fast for the proposed DE. The deviation is little between the worst and the best value of the optimum.

19000

20000

21000

22000

23000

24000

25000

26000

1 26 51 76 101

126

151

176

201

226

generation

Fuel

Cos

t ($

Worst SolutionAverage SolutionBest Solution

Figure 3. . Convergence of DE-based OPF Solutions Algorithm for the

Algerian Network.

In this test system acco rding to res ults o btained from th e continuation load flow to enhance the reactive power planning for th e Al gerian Networ k, the SVC Compensators can b e installed at these critical buses 12, 53 , 54, 55, 66 , 67, 68, 69, 70, 73 , 91, 92 , 93. The DE algorithm based on the objective function which takes i nto account the power losses with only one SVC installed provides bus 68 as the optimum location

TABLE III. COMPARISON OF THE RESULTS OBTAINED BY DE WITH & WITHOUT REGELATION OF TAP CHANGE & SVC CONTROL

DE (w/o) Tap & SVC control

DE with Tap &SVC control

Pg4(M W ) 462.3908 434.68 Ploss(M W ) 89.2570 61.550 Cost[$/hr] 19203.34 18950.514

Based on the Table III, i f we don’t use the regulation of

the t ransformer tap change and the SVC dev ice the cos t was 19203.34 $ /MWh and the loss es value was 89.257 MW compared with the case of compensation the cost was reduced to 18950.514 $/MWh and t he lo sses v alue was red uced to 61.550 MW. Th e act ive p ower g eneration f or the slack b us was reduced from 462.3908 MW to 434.68 MW. The optimal values o f t he vo ltage gen erators b y DE are shown in the Figure 4. The opt imal t ap changes of t he 16 t ransformers of

the Algerian network after optimisation by DE ar e shown in the figure 5.

0.96

0.98

1

1.02

1.04

1.06

1.08

1.1

1.12

4 5 11 15 17 19 22 52 80 83 98 100 101 109 111

Voltage

Figure 4. Optimal Values of Voltages of Generators of 114 Algerian electrical network by the DE-based OPF

0.8

0.9

1

1.1

1.280

-88

81-9

086

-93

42-4

158

-57

44-4

360

-59

64-6

372

-71

17-1

821

-20

27-2

628

-26

31-3

048

-47

74-7

6

Optimal Tap change Value

Figure 5. : The optimal tap change values of transformers of the Algerian

Network

0.85

0.9

0.95

1

1.05

1.11

2 3 4 5 6 7 8 9 10 11 12 13 14151617181920212223

2425262728293031323334

3536373839404142

43444546474849505152535455565758

59606162

6364

6566

6768

6970

7172

7374

7576

7778

7980

818283848586

878889909192

939495

9697

9899

100101

102103

104105

106107

108109

110111

112113

114

DEOPF (Pg) DEOPF(Pg, Vg, Tap, SVC)

Figure 6. : Voltage profile of all buses for the Algerian Network with &

without FACTS device


14 © 2014 UTA

0.8

0.85

0.9

0.95

1

12 53 54 55 56 66 67 68 69 70 73 91 92 93

Bus N°

Vo

lta

ge

Ma

gn

itu

de

(p

.u.)

DEOPF (Pg) DEOPF(Pg, Vg, Tap, SVC)

Figure 7. Voltage magnitude in the critical buses with and without the

include of SVC in the bus 86

The s ecurity co nstraints ar e also ch ecked for voltage magnitudes a nd a ngles. All vol tage m agnitudes of t he Algerian Networ k are b etween their minimum and the maximum values (Fig. 6). No load bus was at the lower limit of the voltage magnitudes (0.9 p.u). Th e f igure 7 s hows th e voltage profile magnitude improvement at critical buses.

IV. CONCLUSION In th is p aper, DE Op timization h as b een p resented an d

applied to economic p ower disp atch. Th e DE al gorithm fo r solving th e OPF has the c ontrol o ver the g lobal and local exploration capabilities. This improves the s earch eff iciency, overcomes premature convergence and avoids getting trapped into the local optima. It does not depend on the nature of the function i t m inimizes. Thus a pproximations m ade in traditional methods can be avoided. And it is insensitive to the initial searching points, thereby ensuring q uality s olution fo r different tr ial r uns. T o ver ify the pr oposed approach and for comparison purposes, we perform simulations on the Algerian Network system. The obtained r esults in dicate th at D E is an easy to use, fast, robust and powerful optimization technique compared with Parti cle Swarm Optimization (PSO) and Genetic Algorithms. Simulation results sh ow that th e DE i s able to minimize the total cost along with minimization of loss in the s ystem wi th m ixed con trol v ariables (discrete and continuous). Also, it is found that DE obtains a better solution in reduced time. T he result show also that i nstalling SVC in right location can significantly enhance the security of power system b y m inimizing t he o verloaded l ines, the bus voltage limit violations and power losses.

REFERENCES [1] A. Momoh, M. E. El-Haw ary a nd R . A dapa, “ A R eview of Sele cted

Optimal Power Flow Literature to 1993 Part I: Nonlinear and Quadratic Programming Approa ches”, IEEE Tra nsaction on Po wer Sy stems, vol. 14 (1), pp. 96-104, 1999.

[2] A. Mo moh, M. E. El-Hawary an d R. Ad apa, "A Review of Selected Optimal Power Flo w Liter ature to 199 3 Par t II: Newton, Linear Programming and Interior Points Methods", IEEE Transaction on Power Systems, vol.14 (1), pp. 105-111, 1999.

[3] H. W. Dommel, W. F. Tinney . “Optimal Power Flow Solution s”, IEEE Transactions on power a pparatus and sy stems, vol. PAS.87 (10), pp. 1866-1876, 1968.

[4] R. Storn, K. Pr ice, Dif ferential evolution—a s imple and e fficient adaptive sc heme fo r global o ptimization over c ontinuous s paces, in : Technical Report TR-95-012, ICSI, 1995.

[5] R. Storn Differential Evolution, A Simple and Ef ficient Heuristic Strategy f or Glo bal Optim ization o ver Continuo us Spaces. Journa l of Global Optimization, 11:341–359, 1997.

[6] A. J. Wood and B. F. Wolle nberg. Power Gener ation, Operation and Control, 2nd Edition, John Wiley, 1996.

[7] Glenn W. Sta gg, A hmed H . El Abiad. Com puter methods in pow er systems analysis, McGraw-Hill, 1981.

[8] Hingorani N.G., High Power Electronics and Flexible AC Transmission System, Power Engineering review, IEEE, 1988

[9] Ge. Shaoyun, T . S Chung, “Coupled a ctive dispatch with FACTS devices and spec ified power flow c ontrol c onstraints,” Proc eedings of the 4th International Conference on Advances in Power System Control, APSCOM 97, HONG KONG, vol.2, pp. 678-683, November 1997.

[10] Y. Lu, A. Ab ur, “Static security enhancement via optim al utilization of thyristor-controlled s eries capacitors,” IEEE Trans . Power System, vol. 17, no. 2, pp. 324-329, May 2002.

[11] M. A. Abdel-Moumen, N. P. Padhy , “ Power flow c ontrol a nd transmission loss m inimization m odel with TCSC fo r practical power networks,” IEEE Proceeding, pp. 880-884, 2003.

[12] C. R . Feu rt-Esquivel, E. Ach a, an d H. Am briz-Pérez, “ A thy ristor controlled se ries c ompensator m odel for the power f low solution of practical power networks,” IEEE Trans. Power Systems, vol . 15, no. 1, pp. 58-64, February 2000.

[13] W. D. Rosehart, C. A. Canizares, and V. H. Quintana, “Effect of detailed power system m odels in Traditio nal and v oltage stability constr ained optimal powe r flow problems,” IEEE Trans. Power Sy stems, vo l. 18, no.1, pp. 25-35, February 2003.

[14] M. Moghavvem i, M.O. Faruque, “Ef fects of FAC TS de vices on s tatic voltage stability,” IEEE Proceeding, pp.357-362, 2000.

[15] Differential Evolution Online], Available: http://www.icsi.berkeley.edu/∼ storn/code.html, 2008.


15 © 2014 UTA

A. Bendakir, D. Dib, T. Ruibah

Abstract— the development of the electrical supply networks of power reveals problems involved in the electromagnetic field; it is essential to take it into account in the installations design in order to avoid or decrease a few effects obstructing or damaging the agglomeration. In this paper we propose to expose to the designer and the owner of these works a calculation code based on the dipoles method for the representation of the electromagnetic fields in the switching station.

Keywords- Electric Station, Electromagnetic Fields, Dipole Method, Busbar.

I. INTRODUCTION

The electromagnetic transient environment appears in substations and air tanks. The closing and opening of circuit breakers create surge waves moving very fast inside the substations.

We present here an analytical method for quantification by calculating the transient electromagnetic radiation following a maneuver.

This method relies on the concept of calculation of transient currents and the electromagnetic field using Maxwell's general equations in the air. Note that their formalisms are based on the same model for the magnetic vector potential (approximating thin son). Also they can use stream functions chosen arbitrarily (lightning current, current maneuver ...) or calculated.

II. Propagation of overvoltageWhen closing switches, for example, suddenly are joined

two ends of drivers that, before closing, had not the same voltage to ground. This closure causes non-sinusoidal voltage waves that both propagate along the line and reflect and refract in areas where two lines with different characteristics are connected or several lines are connected to the same busbar, or at places where the line stops. The calculation of the spread of

these overvoltages is important for the selection of the necessary isolates. The study of these phenomena uses the distributed-constant model. The term surge is justified by the fact that this is an additional voltage which is added locally and temporarily with the voltage corresponding to normal operation [1].

III. Analytical approaches for calculating the transientelectromagnetic field emitted by the power transmission

network For the calculation of transient electromagnetic field

emitted by the VHV and HV air, with short radiating structures (busbar), two approaches are shown in the literature [2]:

- Dipole formalism, - Formalism antennas.

The latter formalism which is purely digital and devoted to calculation frequency requires the use of the Fourier transformation.

III-1. Theoretical overview on the analytical formalisms These formalisms are proposed in the approximation of

thin son (the radius of the antenna is considered very low compared to its length).

Based on a current element as presented in Figure.1, it is possible to deduce three approaches for calculating the electromagnetic field in free space. Under these conditions, the magnetic vector potential is parallel to the current single component.

x

∆ζ

Y

Az(t)

P(x,y,z)

i (ζ, t)

ρ

R(ζ)

0

Z

Figure 1. Basic configuration for calculating the electromagnetic field.

Calculation of transient electromagnetic field in an electrical Substation EHV / HV

Laboratory of Studies and Modeling in Electrical Engineering Department of Electrical Engineering, University of Tebessa E-mail: [email protected]

16 © 2014 UTA

III-2. Method of dipole The basic principle of this method is to discrete the

physical medium, i.e. the wire structure in small cells commonly referred to as "dipoles". The size of these cells depends on electrical and topological criteria [3]. These criteria include: the wavelength of the transmitted signal and the distance from the observation point.

In general, the dipole must be chosen so that the current passing through can be considered as constant from the observation point. In practice, we must observe the following two conditions:

Ld ≤ 20λ (1)

Ld ≤ 10R (2)

Where λ is the wavelength of the signal, R is the distance from the observation point, and Ld the length of the dipole.

Condition (1) permits to hide the spread and condition (2) can account for small variations of the current views of a point very close to the wire structure.

. For a dipole in free space, the expression of the magnetic

vector potential is:

ξξξ

ξξξ

π∆

−∂−∂++−−=∆

)/)(()/)((

)(1

)()/)((

4)( 23 cRt

cRti

RcR

cRtiytH i

x

………. (8)

ξξξ

ξξξ

π∆

−∂−∂++−=∆

)/)(()/)((

)(1

)()/)((

4)( 23 cRt

cRti

RcR

cRtixtH i

y

………. (9) 0)( =∆ tH i

z (10)

The electric field components are obtained from equations:

t

AdgraE

∂∂

−−=r

rrϕ (11)

et

0²

1 =∂∂+

tcAdiv

ϕr (12)

We will have then:

ξξξ

ξξτξτ

ξξ

ξπε

∆

−∂−∂+

−

+−=∆

∫)/)(()/)((

)(²1

)(

)/)((3

)()/)((3

)(4)(

30

220

cRt

cRti

RcR

dcRi

cR

cRti

R

xztE

t

ix

(13)

ξξξ

ξξ

τξτ

ξξ

πεξ

ξξ

ξ

ξ

τξτ

ξξ

ξπε

∆

−∂−∂

+−

+

−−∆

−∂−∂

+−

+

−=∆

∫

∫

)/)(()/)((

)(²1

)(

)/)((

)()/)((

41

)/)(()/)((

)(²1

)(

)/)((3

)()/)((3

)(4²)(

30

20

30

220

cRt

cRti

RcR

dcRi

cR

cRti

cRt

cRti

Rc

R

dcRi

cR

cRti

R

ztE

t

t

iz

…………. (14)

IV. Transient electromagnetic environment of airstations

The closing and reclosing of a transmission line of energy occurring in a vacuum is from an air station; a classic sequence which consists of a maneuvered and a disconnected circuit breaker is generally used.

The off-load operations are causing a severe transient confined to the air post. These transients are the cause of overvoltage’s and surges that, in turn, cause significant electromagnetic radiation.

Given the short length of busbar sections, it seems more appropriate to calculate the transient electromagnetic field using the theory of Hertzian dipoles [3].

IV-1. Taking into account the finite conductivity of the soil In the transmission of energy, land is the return

conductor; it is necessary to take into account its presence given the low height of the busbar. In the case of a perfectly conducting ground the method used to account for the soil is the conventional images.

In this work, to avoid the concept of plane wave, we use the method of images introduced by Takashima [4]. It offers a comprehensive study in which he deduces the current image according to the position of the source and observation point.

In the case where the source and the observation point are in the air, the electromagnetic field can shortly be estimated as the sum of the field due to source current (I) that is due to its image ( 'I ).

Interface

Air( 0ε )

Soil( ss ,εσ )

hp

h’

r

A. Bendakir et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 16-20

17 © 2014 UTA

Figure 2. Configuration for the calculation of the field in medium 1 (air).

IIs

s

0

0'εεεε

+−

= (15)

With: ω

σεε

js

ss +=

To account for the soil-air interface we use the following procedure:

I(t) I(ω)

R(ω)

T(ω) "I (ω)

'I (ω) 'I (t) FFT

IFFT ''I (t)

IFFT

Figure 3. The flow chart for calculating the current in the medium 1 (air).

FFT, IFFT: are the transform and inverse Fourier transform.

Field components produced are obtained by superimposing all dipolar contributions (real and images).

( )∑=

+=n

i

iimage

iréel EEE

1 (16)

( )∑=

+=n

i

iimage

iréel HHH

1 (17)

Where: n: is the number of dipoles

V. APPLICATIONS To validate our theoretical work, we consider the case of

application flexibility of a circuit breaker in an air post. The air station has only a portion of 410 kV (RMS voltage between phases). It is schematically represented in figure.4.

In a post, although three-phased, usually the maneuver

takes place phase by phase. Transients do not co-exist simultaneously on all three phases. For our simulation we proceed with the closure of the central arm breaker D0 and we calculate the field at point P. Usually in an aerial position there

is a grid of ground, which results in a perfectly conducting ground. In our application, we have examined the case of a perfectly conducting ground and another of average soil conductivity ( mSs /01.0=σ ).

Sectionneur S0 fermé

Disjoncteur D0 ouvert

Ueff = 410 kV (entre phases) f = 50 Hz

Ligne d’arrivée

Sectionneur S1 fermé

Disjoncteur D1 fermé

Ligne de départ

Jeu de barres

Self

P

x

y z

Figure.4. Descriptive diagram of the air station

To calculate the current distribution we use the utility in

Matlab Simulink [5]. In figure.5, we present the wiring diagram for the study of the transient with Sumilink.

Figure.5. Electric diagram for the calculation of transient

currents in Simulink

Figure.6.Variation of the current injected into the busbar.

Air( 0ε )

Air( 0ε )

I

p

h’

h

h r

r’

I’


18 © 2014 UTA

Figure.7. Variation of the current injected into the busbar (result published in [6] for an air mail EHV when closing a vacuum breaker.

In figure.6, we present the current injected at the entrance of busbar due to the closure of the vacuum breaker D0. Figure.7 shows the injected current measured [6] at the entrance to the bus bar for the same type of air mail, the authors publish only the measurement results without giving details of the measuring station. Note that the value of the inductor current limiting will enormously influence the magnitude and speed of the transient current.

From these results (figures.6 and 7) we can state that the modeling that we carried out with Simulink (Matlab) is qualitatively acceptable.

Figure. 8. Variation of the Hx component at point P

Figure .9. Variation of the component Hy at the point P

Figure.10. Variation of the component Hz at the point P

Figure.11. Variation of the component Hy point inside the air

station at different voltages (as published in [7]).

What we get as the general shape of the magnetic field for the three components at close range being calculated directly in the game bar turned on, is also confirmed by the measurement published in [7] (figure.11). We can compare the order of magnitude as the authors [7] do not show the calculation point and the length and the height of the busbars from the ground. Note, however, that the order of magnitude is respected.

Figure.12. Variation of the component Ex at point P


19 © 2014 UTA

Figure13. Variation of the component Ey at point P

Figure.14. Variation of the component Ez at point P

Figure.15. Variation of the component Ex point inside the air

station at different voltages (as published in [7]).

What we get as the general appearance of the electric field for the three components at close range being calculated directly in the game bar turned on, is also confirmed by the measurement published in [7] (Figure 14). The electric field (three components) is contrary to the uni-polar magnetic field that is dual-polar.

VI. CONCLUSION For the integrity of the substation, it is necessary to

characterize its own electromagnetic environment. On-site measurement is possible, but remains expensive and

sometimes inadequate. Modeling when it is well suited, is an effective way to relay and complete the measure. In this paper, we used an analytical model for calculating the electromagnetic field emitted transient based on the method of Hertzian dipoles [3]. Our calculation results show the nature dual-polar magnetic field and pole for the electric field (figures.8 to 15). This result is confirmed by the measurement made by Electric Power Research Institute [7]. Also the general shape is present in our results although we have not the right value of the self-limiting current in the busbar;

the self influences the amplitude and pseudo-period of transition. Finally, it should be noted that an important element was not taken into account in our modeling, it is the nonlinearity of the closing operation, which involves resistance of the arc nonlinear. Note that our simulation is to represent the maneuver by the all or none of the contact’s resistance of the circuit breaker. This important element appeared in the results of measurements made by EPRI [7].

REFERENCES [1] Michel Pays, ‘‘Câbles de Transport d’Énergie : Technologies.

Caractéristiques ’’, Technique de l’ingénieur, Vol. D4520, pp.1-35. [2] R. S. Shi, "Rayonnement Électromagnétique des Réseaux Électriques à

Topologie Complexe", Thèse de Doctorat de l’INPG. Grenoble 1992. [3] N. Ari, W, Blumer, ‘‘transient electromagnetic fields due to Switching

Operations in Electric Power Systems’’, IEEE Trans. On EMC, Vol. EMC-29, No.3, pp, 233-237, Aug., 1987.

[4] T. Takashima, T. Nakac, R. Ishibashi, « Calculation of Complex Fields in Conducting Media », IEEE Trans on electrical insulation Vol EI-15, N° 1, February 1980.

[5] Toolbox, Matlab 6.5, Simulink [6] C.M.Wiggins, and al, ‘‘Measurement of switching Transients in a

115kVSubstation’’, IEEE Trans. On PWRD, Vol. 4, No.1, pp.756-769, January, 1989.

[7] C.M.Wiggins, S. E. Wright, ‘‘Switching Transient Fields in a 115 kV Substation’’, IEEE Trans. on PWRD, Vol. 4. 6, No. 2, pp. 791-769, January, 1989.


20 © 2014 UTA

Design of a fractional order PID controller for a class of fractional order MIMO systems

Abdelhamid Djari and Toufik BoudenNDT-Laboratory, Automatic Control Department,

Jijel University, Jijel, Algeria [email protected] ; [email protected]

[email protected]

Abdesselem BoulkrouneLAJ-Laboratory, Automatic Control Department, Jijel

University, Jijel, Algeria [email protected]

Abstract— This paper deals with the design of a Fractional-order

proportional-integral-derivative (FOPID) by using the different

approximation methods, namely: Oustaloup, Matsuda, Carlson, for a class of MIMO fractional-order systems. A comparative

study between the FOPID implemented via the approximation methods above cited and the classical PID is carried out. In

addition, the results of a numerical simulation of a synchronous

machine are presented.

Keywords- MIMO systems, a fractional-order systems, PID,

fractional-order PID, approximation of the fractional operators.

I. INTRODUCTION

Fractional-order calculus is an area of mathematics that deals with derivatives and integrals from non-integer orders. In other words, it is a generalization of the traditional calculus that leads to similar concepts and tools, but with a much wider applicability. In the last two decades, fractional calculus has been rediscovered by scientists and engineers and applied in an increasing number of fields, namely in the area of control theory. The success of fractional-order controllers is unquestionable with a lot of success due to emerging of effective methods in differentiation and integration of non-integer order equations [1]-[14].

Fractional-order proportional-integral-derivative (FOPID) controllers have received a considerable attention in the last years both from academic and industrial point of view. In fact, in principle, they provide more flexibility in the controller design, with respect to the standard PID controllers, because they have five parameters to select (instead of three). However, this also implies that the tuning of the controller can be much more complex. In order to address this problem, different methods for the design of a FOPID controller have been proposed in the literature. However, the implementation of a FOPID has been generally made via an appropriate approximations of those fractional-orders, namely Oustaloup, Matsuda, Carlson, or via the so-called exact analytical formula (i.e. without any approximation) [15]-[22].

This paper deals with the design of a FOPID by using the different approximation methods above cited, for a class of MIMO fractional-order systems. A comparative study between the FOPID (by using the different approximation methods) and the classical PID will be carried out.

This paper is organized as follows. Section 2 includes basic concepts in fractional calculus and the different used approximations for FOPID implementation. In Section 3 the

Bode’s ideal loop (BIL) with temporal and frequency analysis is presented. In section 4 the determination of the representation of state of such a non-integer system is presented and the fractional order multiple input and multiple-output (MIMO) systems models to be controlled are studied in the state representation. The design of FOPID is detailed in section 5. The application of FOPID, made according to the approximations of: Oustaloup, Matsuda, Carlson and the exact analytical formula (without any approximation) and a comparison of the corresponding results using fractional MIMO system are presented. Finally, concluding remarks are drawn in Section 7.

II. FRACTIONAL ORDER OPERATORS

A fractional derivator term is given by:

𝑦 𝑡 = 𝜏𝑛𝐷𝑛𝑢(𝑡) (1)

τ indicates the differentiation time constant and

Cn

the complex order of the derivation (Re (n) which can be higher or lower than 0, the operator considered being then either a derivator, or an integrator).

Using the Laplace’s transform and under null initial conditions, we can write (1) as follows:

𝑌 𝑠 = 𝜏𝑠 𝑛𝑈(𝑠) (2)

Posing now ωu = 1/τ, called transitional frequency, we get:

𝑌 𝑠 = 𝑠

𝜔𝑢 𝑛𝑈(𝑠) (3)

Thus, the transmittance is given by:

𝐷 𝑠 =𝑠

𝜔𝑢

𝑛 (4)

The calculation of fractional derivatives and integrals of an order α for a specified function is generally very difficult by using the analytical methods. By consequence, a numerical approximation is necessary [3][6][7][12]-[14]. We use a function fotf (fractional order transfer function) of the Matlab toolbox to determine the so-called exact responses of an fractional order derivative or integral and to compare them with the various following approximations:

A. Oustaloup’s approximation

The method [7] is based on the approximation of a

function of the form:

𝑇 𝑠 = 𝑠q , q ∈ 𝑅+ ,

(5) By a rational function:

21 © 2014 UTA

mailto:[email protected]



𝑇 𝑠 = 𝐶1+

𝑠

𝑤𝑘

1+𝑠

𝑤′𝑘

𝑁𝑘=−𝑁 (6)

Where the parameters of this function can be determined via the following formulas :

𝑤0′ = 𝛼−0.5𝑤𝑢 ; 𝑤0 = 𝛼0.5𝑤𝑢;

𝑤𝑘+1′

𝑤𝑘′ =

𝑤𝑘+1

𝑤𝑘= 𝛼𝜂 > 1;

𝑤𝑘+1′

𝑤𝑘

> 0; 𝑤𝑘

𝑤𝑘′

= 𝛼 > 0; 𝑁 =log (𝑤 𝑁

𝑤 0)

log (𝛼𝜂 ); q = log (𝛼)

log (𝛼𝜂 ) (7)

With 𝑤𝑢 being the frequency of the unit gain and the center frequency of a band of the frequencies geometrically

distributed around it, i.e. : 𝑤𝑢 = 𝑤ℎ𝑤𝑏 where 𝑤ℎ and 𝑤𝑏

are respectively the high and low transient-frequencies.

B. Matsuda’s approximation

The method suggested by [8] is based on the

approximation of the fractional-order operator 𝑇 𝑠 = 𝑠α by a rational function identified by using its gain. The gain is

calculated by using M frequencies left again in a waveband

𝑤0 ,𝑤𝑀 in which the approximation is made. For a set of

selected points 𝑤𝑖 , 𝑖 = 0,1,2,… , 𝑀, the approximation takes

the following form:

𝑇 𝑠 = 𝑎0 +𝑠−𝑤0

𝑎1 +𝑠−𝑤1

𝑎2+𝑠−𝑤2𝑎3+⋯

(8)

𝑎𝑖 = 𝑣𝑖 𝑤𝑖 ; 𝑣0 𝑤 = 𝐺 𝑗𝑤 ; 𝑣𝑖+1 𝑠 =𝑠−𝑤𝑖

𝑣𝑖 𝑠 −𝑎𝑖 ;

For: i = 0,1,2,…M (9)

The approximated model is obtained by replacing each

fractional operator of the irrational explicit transfer function

by its approximation.

C. Carlson’s approximation

The method suggested by Carlson in [9], derived from a

regular Newton process used for the iterative approximation of

𝛼 roots, can be considered as a membership of this group. The

starting point of this method is the report of the following

relations:

𝐻(𝑠) 1

𝛼 − 𝐺(𝑠) = 0; 𝐻 𝑠 = 𝑠𝛼 (10)

Defining α=1/q , m=q/2, in each iteration, starting from

the initial value 𝐻0 𝑠 = 1, an approximated rational function

is obtained in the form:

𝐻𝑖(𝑠) = 𝐻𝑖−1(𝑠) 𝑞−𝑚 𝐻𝑖−1(𝑠) 2+ 𝑞+𝑚 𝐺(𝑠)

𝑞+𝑚 𝐻𝑖−1(𝑠) 2+ 𝑞−𝑚 𝐺(𝑠) (11)

III. BODE’S IDEAL LOOP (BIL)

The Bode's ideal loop (BIL) transfer function was proposed

for the first time by Bode in its work on the design of the

amplifiers with feedback in 1945 [6][10]. The diagram of

such an amplifier with unity-gain feedback is given by the

figure 1, whose transfer function in open loop is defined by an

integrator of a fractional nature of the form:

𝑇 𝑠 =𝐴

𝑆𝛼 , 1 < 𝛼 < 2 (12)

Where α is the slope of the ideal characteristic of the

gain.

Figure 1. Diagram of BIL

A. Temporal analysis of the BIL

The transfer function in closed-loop of BIL is the form:

𝐻 𝑠 =𝑌(𝑠)

𝑅(𝑠)=

𝐴

𝑆𝛼 +𝐴 (13)

Its step response is given by :

𝑦 𝑡 = 𝐴𝑡𝛼𝐸𝛼 ,𝛼+1(−𝐴𝑡𝛼) (14)

Where 𝐸𝛼 ,𝛼+1 is the function of Mittag-Leffler:

𝐸𝛼 ,𝛼+1(𝜃) =𝜃𝑘

Γ(1+𝑘.𝛼)

∞𝑘=0 (15)

The step responses of the system 𝐻 𝑠 are characterized by a

damping coefficient 𝛿 , a natural pulsation 𝜔𝑛 and an eigen

frequency 𝜔𝑝 given by the following formulas:

𝛿 = − cos 𝜋

𝛼 ; 𝜔𝑛 = 𝐴

1𝛼 ; 𝜔𝑝 = 𝐴

1𝛼𝑠𝑖𝑛

𝜋

𝛼 (16)

The maximum overshoot can be expressed according to the

order α by [12]:

𝑀𝑝 =ℎ𝑚𝑎𝑥 −ℎ(∞)

ℎ(∞)≈ 0.8 𝛼 − 1 𝛼 − 0.75 ; 1 < 𝛼 < 2 (17)

The time of the first overshoot and the boarding time

(2% to 90%) can be given in an approximate way by the

following expressions:

𝜔𝑢𝑡𝑑 ≈1.106 𝛼−0.255 2

𝛼−0.921 2 ; 1 < 𝛼 < 2 (18)

And

𝜔𝑢𝑡𝑚 =0.131 𝛼+1.157 2

𝛼−0.724 ; 1 < 𝛼 < 2 (19)

Where 𝜔𝑢 is the transitional frequency. The figure 2

represents the step responses of the system (12) for 𝛼 = 1.5 and for the different values of A. This figure shows that the

maximum overshoot is approximately of 29.62% and

independent of the gain A. This property is called iso-

damping (ξ = constante ≈ 0.35).

Figure 2. The step responses of the BIL for α = 1.5 and various values of A

B. Frequential analysis of the BIL

The Bode’s diagram of the direct chain of the ideal loop

of Bode is given by the fig. 3. The frequential response is

characterised by a slope of −20𝛼𝑑𝐵/𝑑𝑒𝑐 and a constant phase

0 5 10 15 20 25 30 35 400

0.2

0.4

0.6

0.8

1

1.2

1.4

A=0.036

A=1

A=31.62

A. Djari et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 21-27

22 © 2014 UTA

of –απ

2 rad . Thus, the phase margin in closed-loop is

independent of the gain A and equal to: Φm = (1 −α

2)π.

The factor of resonance Mr and the frequency of resonance ωr can be determined in the same way like in the case of integer-

order systems. Let H jω be the transmittance in closed-loop:

𝐻 𝑗𝜔 =𝐴

𝑗𝜔 𝛼 +𝐴=

𝐴

𝐴+𝜔𝛼 𝑐𝑜𝑠 𝛼𝜋

2+𝑗𝜔 𝛼 𝑠𝑖𝑛 𝛼

𝜋

2

(20)

Its module is given by:

𝐻(𝑗𝜔) =𝐴

𝜔2𝛼 + 2𝐴𝜔𝛼 𝑐𝑜𝑠 𝛼𝜋2

+ 𝐴2

(21)

The module (21) has a maximum for:

𝜔𝑟 = −𝐴𝑐𝑜𝑠 𝛼𝜋

2

1𝛼

, 𝛼 > 1 (22)

Corresponding to a factor of resonance given by:

𝑀𝑟 =1

𝑠𝑖𝑛 𝛼𝜋

2

(23)

Figure 3. Bode’s diagram in open-loop of BIL

IV. FRACTIONAL –ORDER MIMO SYSTEMS

Fractional-order (FO) systems have attracted increasing

interests, mainly due to the fact that many real-world physical

systems are better characterized by FO differential equation.

From the differential equations of a non-integer-order system,

it is possible to build a state model using the fractional

derivative of the system’s state variables. In order to determine

the state representation of such a non-integer –order system,

let’s consider the following differential equation obtained by

linearization [1]-[14][19][20][23]:

𝑎𝑛𝑑

𝑑𝑡

𝛼𝑛𝑁

𝑛=0

𝑦 𝑡 = 𝑏𝑚𝑑

𝑑𝑡

𝛽𝑚𝑀

𝑚=0

𝑢 𝑡

où 𝛼𝑛 ,𝛽𝑚 ∈ 𝑅, 𝛼𝑁 ≥ 𝛽𝑀 (24)

Where the vector 𝛼𝑛 represents the vector of the derivative orders of the output y(t) classified in the ascending order. The

vector βn represents the vector of the derivative orders of the

input u(t). The coefficients an and bm are defined according to

the parameters of the considered system.

To build the system of generalized states, one can carry out

the following variables change :

The first introduced variable, x1, is defined likethe derivative of order α0 of y(t);

The second variable x2 is defined like thederivative of order α1 of the output y(t);

The ith variable is defined like the derivative of

order αi of the output y(t).

In general case, α0 = 0; the variable x1 corresponds thus to

the function y(t).

The new variables and the relations between them are given

by :

𝑑

𝑑𝑡

(𝛼0 )𝑦 𝑡 = 𝑥1 𝑡

𝑑

𝑑𝑡

(𝛼1 )𝑦 𝑡 =

𝑑

𝑑𝑡

(𝛼1−𝛼0+𝛼0)𝑦 𝑡 = 𝑥1

(𝛼1−𝛼0) 𝑡 = 𝑥2 𝑡

⋮𝑑

𝑑𝑡

(𝛼𝑖−1)𝑦 𝑡 =

𝑑

𝑑𝑡

(𝛼𝑖−1−𝛼𝑖−2+𝛼𝑖−2)𝑦 𝑡 = 𝑥𝑖−1

(𝛼𝑖−1−𝛼𝑖−2) 𝑡

= 𝑥𝑖 𝑡 ⋮

𝑑

𝑑𝑡

(𝛼𝑁−1)𝑦 𝑡 =

𝑑

𝑑𝑡

(𝛼𝑁−1−𝛼𝑁−2+𝛼𝑁−2)𝑦 𝑡 = 𝑥𝑁−1

(𝛼𝑁−1−𝛼𝑁−2) 𝑡 𝑡

= 𝑥𝑁 𝑡

𝑑

𝑑𝑡

(𝛼𝑁 )𝑦 𝑡 =

𝑑

𝑑𝑡

(𝛼𝑁−𝛼𝑁−1+𝛼𝑁−1)𝑦 𝑡 = 𝑥𝑁

(𝛼𝑁−𝛼𝑁−1) 𝑡

(25) Taking into account of this variables change, the equation

(24) becomes:

𝑥𝑁(𝛼𝑁−𝛼𝑁−1) 𝑡 = −

𝑎0

𝑎𝑁𝑥1 𝑡 −

𝑎1

𝑎𝑁𝑥2 𝑡 − ⋯−

𝑎𝑁−1

𝑎𝑁𝑥𝑁 𝑡 +

𝑏𝑚

𝑎𝑁

𝑑

𝑑𝑡

𝛽𝑀𝑢 𝑡 𝑀

𝑚=1 (26)

From (25) and (26), we can write the state model as follows:

𝑥1(𝛼1−𝛼0 ) 𝑡

𝑥2(𝛼2−𝛼1 ) 𝑡

⋮𝑥𝑁−1

(𝛼𝑁−1−𝛼𝑁−2 ) 𝑡

𝑥𝑁(𝛼𝑁−𝛼𝑁−1 ) 𝑡

=

0 1 0 ⋯ 00 0 1 ⋱ ⋮⋮ ⋱ ⋱ ⋱ 00 ⋯ ⋯ 0 1

−𝑎0

𝑎𝑁

−𝑎1

𝑎𝑁

⋯ ⋯ −𝑎𝑁−1

𝑎𝑁

.

𝑥1 𝑡

𝑥2 𝑡

⋮𝑥𝑁−1 𝑡

𝑥𝑁 𝑡

+

0⋯00⋯00⋯00⋯0

𝑏1

𝑎𝑁⋯

𝑏𝑀

𝑎𝑁

. 𝑢(𝛽1

) 𝑡

⋮

𝑢(𝛽 ) 𝑡

(27)

Thus, the system (27) can be generally written in the

form:

𝑥 𝑛 𝑡 = 𝐴. 𝑥 𝑡 + 𝐵. 𝑢 (𝑡)… . 𝑦 𝑡 = 𝐶. 𝑥 𝑡 + 𝐷. 𝑢 (𝑡)

(28)

Thus, as in the integer-order case, a representation of

non-integer-order state comprises two equations; an

equation of state generalized in which the vector of

state is not any more the object of a unit derivation but of a derivation of non-integer real order and an


23 © 2014 UTA

equation of observation.

We also see appearing the vectorial operator 𝑛 =

(𝛼1−𝛼0, … , 𝛼𝑁−𝛼𝑁−1) steady to the vector of state. Thus, this vector is not inevitably composed of identical terms.

Nevertheless, in the studies of stability, we will systematically

seek an arrangement of the state variables so that the

components of this vector (n) are the same ones. It could then

be comparable with a real number n [14][15][20][23].

V. DESIGN OF A FRACTIONAL ORDER PID CONTROLLER

Recently, FOPID controller design has been increasingly

used in the control area by more and more researchers.

Compared with the classical PID controller, FOPID controller

have the potential to improve the control performance,

because extra real parameters 𝛼 and 𝛽 are involved. However, to the best of our knowledge, there are two common

assumptions for the considered plant, which are (i) The plant

can be modeled as first or second-order systems, even plus a

time delay item. (ii) The plant is of the form of single input

single output system. And few results have been obtained for

multivariable FOPID controller design for the FO systems,

especially when the parameters of the FO system is interval

uncertainties. Therefore, it is highly make sense that

developing methods to determine the parameters of the FOPID

controllers for FO multi-inputs multi-outputs high order

systems. The above all motivate this present study.

Podlubny proposed a generalization of the classical PID

controller, called fractional-order PID controller defined by

the following transfer function [15]:

𝐶 𝑆 =𝑈(𝑆)

𝐸(𝑆)= 𝐾𝑝 +

𝐾𝑖

𝑆𝛼+ 𝐾𝑑𝑆

𝛽 , 𝛼, 𝛽 ∈ 𝑅+ (29)

Several methods were proposed for the design of this type of

order [11][15]-[18][21][23]. Moved by the remarkable

performances characteristic in quality of robustness of the

ideal loop of Bode. In what follows, we will propose, the

design of a control PIαDβ which ensures the same frequentialand temporal behavior that found while being based on the

ideal loop of Bode in closed loop.

After having fixed the fractional orders α and β starting

from the behavior frequential of the system of open loop

control, which must be equivalent to that of the ideal transfer

function of Bode. We established a simple design of a control

system of a fractional nature elementary based on the

parameters 𝑚 and 𝜔𝑢. Consequently, in this section we use the

fractional integrator of order of the equation (12) like a

function of reference 𝐺𝑚 (𝑠) for the control device PIα Dβ [15]. We start by considering the system in closed loop shown in

the figure 4, where C(s) is the controller PIα Dβ and 𝐺𝑝(𝑠) is

the transfer function of the process, characterized by an

asymptotic order at low frequency 0 ≤ 𝑛′ ≤ 2 and high

frequency 2 ≤ 𝑛 ≤ 4 with 𝑛′ < 𝑛 (fig. 4).

Figure 4. Closed loop system

The ideal transfer function of the controller is the

following form:

𝐶 𝑠 = 𝑘𝑝 1 +𝑇𝑖

𝑆𝛼+ 𝑇𝑑𝑠

𝛽 , 𝑘𝑝 , 𝑇𝑖 𝑒𝑡 𝑇𝑑 ∈ 𝑅+ (30)

With α and β are positive real numbers, 𝑘𝑝 is the

proportional gain, 𝑇𝑖 is a constant of integration and 𝑇𝑑 is a constant of derivation.

The cut-off frequency 𝜔𝑢 is considered to be higher 10

times than the transitional frequency of the process.

The transfer function of the model of reference in open

loop has the form:

𝐺𝑚 𝑠 =1

𝑆

𝜔𝑢

𝑚 1 < 𝑚 < 2 𝑒𝑡 𝜔𝑢 ∈ 𝑅+ (31)

Where 𝜔𝑢 and 𝑚 are fixed according to the

performances desired in closed loop.

The method of synthesis of PIαDβ based on theinterpretation of the transfer function in open loop T(s)

which can be written in the form:

𝑇 𝑠 = 𝐶(𝑠)𝐺𝑝(𝑠) (32)

Transmittance T(s) can be here regarded as the approximation

of the transfer function in open loop of the model of

reference 𝐺𝑚 𝑠 , and then we can write:

𝑘𝑝 1 +𝑇𝑖

𝑆𝛼+ 𝑇𝑑𝑠

𝛽 𝐺𝑝(𝑠) ≈𝑘𝑢

𝑠𝑚; 𝑘𝑢 = 𝜔𝑢

𝑚 (33)

We give the waveband: 𝜔𝑚𝑖𝑛 ≪ 10𝜔0, 𝜔𝑢 ≪ 𝜔𝑚𝑎𝑥

Where 𝜔0 is the cut-off frequency of the process,

transmittance T(s) should present:

An asymptotic slope of −20𝑑𝐵/𝑑𝑒𝑐 with low

and the high frequency of the limited bandwidth

𝜔𝑚𝑖𝑛 ,𝜔𝑚𝑎𝑥 which makes it possible to calculate the parameters α and β.

A same gain with the reference 𝐺𝑚(𝑠) into high

and low frequencies, thus, the initial values of

the parameters 𝑇𝑖 ′ and 𝑇𝑑 ′ can be estimated

with an initial value of 𝑘𝑝 considered 𝑘𝑝 = 1.

A cut-off frequency equal 𝜔𝑢, the parameter 𝑘𝑝

can be deduced by using the initial values 𝑇𝑖′and 𝑇𝑑 ′.

Finally, we make the adjustment of the

parameters 𝑇𝑖′ and 𝑇𝑑 ′ to obtain the parameters

values: 𝑇𝑖 = 𝑘𝑝𝑇𝑖′ And 𝑇𝑑 = 𝑘𝑝𝑇𝑑 ′.

Taking as n and n’ an asymptotic order of the

process 𝐺𝑝(𝑠) at high and low frequencies respectively,

the parameters of the controller can be given by the following

equations [15]:

𝛽 = 𝑛 − 𝑚 , 𝛼 = 𝑚 − 𝑛′ (34)


24 © 2014 UTA

𝑇𝑑′ =

𝐾𝑢

𝐺𝑝(𝑗𝜔𝑚𝑎𝑥 ) 𝜔𝑚𝑎𝑥𝑛 (35)

𝑇𝑖′ =

𝐾𝑢

𝐺𝑝 (𝑗𝜔𝑚𝑖𝑛 ) 𝜔𝑚𝑖𝑛𝑛′ (36)

𝑘𝑝 =𝐺𝑝(𝑗𝜔𝑢 )

−1

1+𝑇𝑖′ 𝑗𝜔𝑢 −𝛼 +𝑇𝑑

′ 𝑗𝜔𝑢 𝛽 (37)

𝑇𝑖 = 𝑘𝑝𝑇𝑖′ and 𝑇𝑑 = 𝑘𝑝𝑇𝑑 ′ (38)

VI. APPLICATION

In this section, we have focused on synchronous machine

modeling (The synchronous machine used here is a Darlington

type). It is then necessary not only to supply accurate models

of machines over a wide range of frequencies, but also to

include knowledge in models. Moreover, model order has to

be reduced as robust control techniques lead also to high-order

controllers [23]. The system order is then so high that classical

dynamic studies like stability become difficult. In most

generators, as frequency increases, induced currents can no

longer be neglected. Equivalent circuits of synchronous

machines are then improved by including ladder elements

with constant parameters [24][25]. However, as this effect is a

distributed phenomenon described by partial differential

equations, classical improved equivalent circuits must include

in theory an infinite number of lumped and constant

parameters. Finally, half-order linear impedances can be

included in the Park equivalent circuits of synchronous

generators (Fig. 5) following some physical conditions [24]-

[26]. Then, for d-axis modelling, no induced currents expand

in armature windings or field windings for low frequencies;

they are then modelled by constant parameters. After modeling synchronous machine we will stabilize their

currents id and iq using both PID and FOPID (with different

approximation methods).

Figure 5. Non integer order d and q-axis equivalent circuits of a synchronous machine

The system of equations governing the electric circuits is

given in the field of Laplace by:

Rd,ch . id = −rsd . id − lsd . s. id − lad . s. id + if + i1d + i2d

Vf = −rf . if − lf . s. if − lad . s. id + if + i1d + i2d

−l12d . s. if + i2d

0 = −R2d . 1 +s

ω2d

1

2 . l2d − lad . s. id + if + i1d + i2d

− l12d . s. if + i2d

0 = −s.L1d .i1d

1+ s

ω1d

12

− lad . s. id + if + i1d + i2d

Rq,ch . iq = −rsq . iq − lsq . s. iq − laq . s. iq + i1q + i2q

0 = −r2q . i2q − l2q . s. i2q − laq . s. iq + i1q + i2q

0 = −s. L1q . i1q − 1 + s

ω1q

1

2

. laq . s. iq + i1q + i2q

(39)

A system of generalized state space is then built by using the

following change of variables:

x1 = id ; …………… ..

x2 = x1(1/2)

= id(1/2)

;

x3 = if ;……………… .

x4 = x3(1/2)

= if(1/2)

;

x5 = i2d ;

x6 = x5(1/2)

= i2d(1/2)

;

x7 = i1d(1)

;

x8 = iq ;

x9 = x8(1/2)

= iq(1/2)

;

x10 = i2q ;

x11 = x10(1/2)

= i2q(1/2)

;

x12 = i1d(1)

; (40)

We can write (40) as a following state space model:

x (1

2) = A. x + B. u ; (41)

Where:

A =

0A2.1

0A4.1

0A6.1

A7.1

00000

100000

A7.2

00000

0A2.3

0A4.3

0 A6.3

A7.3

00000

001000

A7,4

00000

00000

A6.5

A7.5

00000

000

A4.6

1A6.6

A7.6

00000

000

A4.7

0A6.7

A7.7

00000

00000000

A9,8

0A11,8

A12,8

00000001000

A12,9

00000000

A9,10

0A11,10

A12,10

00000000010

A12,11

00000000

A9,12

0A11,12

A12,12 (42)

𝐵 =

0𝑘2

0𝑘4

0𝑘6

𝑘7

00000

000000

𝑘7′

00000

; 𝑢 = 𝑉𝑓 𝑉𝑓

(1/2) (43)

The 05 parameters of 𝑃𝐼𝛼𝐷𝛽 will be calculated using the

approximations of Oustaloup, Matsuda, Carlson and the exact

analytical formulas to control and stabilize synchronous

machine’s currents id and iq using d and q-axis in the


25 © 2014 UTA

waveband [10-3,103] rad/s. We have already chosed the order

of approximation n = 4 in the simulation. The figures 6-9 respectively represent the step responses and the Bode’s diagrams of the studied system (synchronous machine currents id and iq in the reference axis of Park) controlled by PID and FOPID(with various approximations) in open loop as illustrated in fig. 4.

Figure 6. Step responses of the current id (state x1) controlled by classical PID and FOPID with various approximations

Figure 7. Bode’s diagrams of the current id (state x1) controlled by classical PID and FOPID with various approximations

Figure 8. Step responses of the current iq (the state x8) controlled by classical PID and FOPID with various approximations

Figure 9. Bode’s diagrams of the current iq (state x8) controlled by classical PID and FOPID with various approximations

Figures 6 and 8, show that the FOPID stabilize the currents and ensure the best results of robustness by comparing them with those of an integer order (a better speed).

The performances of PID and FOPID with various

approximations are summarized in the tables below:

TABLE 1. PID and FOPID PERFORMANCES FOR CURRENT id

Tm (s) Tr5% (s) D% Vfinale MP(°)

FOPID (without approximation)

0.002 0.003 00 01 45

FOPID with Oustaloup

approximation 0.001 0.004 05 01 45

FOPID with Matsuda

approximation 0.007 0.021 16 01 47.5

FOPID with Carlson

approximation 0.001 0.003 18 01 50

Classical PID 0.017 0.062 25 01 43.8

TABLE 2. PID and FOPID PERFORMANCES FOR CURRENT iq

Tm (s) Tr5% (s) D% Vfinale MP(°)

FOPID (without approximation)

0.005 0.007 00 01 45

FOPID with Oustaloup

approximation 0.005 0.007 02 01 45

FOPID with Matsuda

approximation 0.01 0.026 07 01 47.5

FOPID with Carlson

approximation 0.005 0.007 01 01 50

Classical PID 0.03 0.054 08 01 43.8

With regard to the comparison of the performances of robustness (Tab. 1 and Tab. 2), they reveal clearly that the robustness is much better in the case of the FOPID and its approximations, in favor of the approximation suggested by Oustaloup.

VII. CONCLUSION

In this paper, we stabilize and control a fractional MIMO

system (synchronous machine) using PID and FOPID with

0 0.05 0.1 0.15 0.2 0.250

0.2

0.4

0.6

0.8

1

1.2

1.4

Step Response

Time (sec)

Am

plit

ude

Matsuda

Carlson

Oustaloup

PID entier

PID frac-exact

-50

0

50

100

150

200

Mag

nitu

de (

dB)

10-3

10-2

10-1

100

101

102

103

-540

-360

-180

0

180

Pha

se (

deg)

Bode Diagram

Frequency (rad/sec)

PID frct-exat

Carlson

Oustaloup

PID entier

Matsuda

0 0.05 0.1 0.15 0.2 0.25 0.30

0.2

0.4

0.6

0.8

1

1.2

1.4

Step Response

Time (sec)

Am

plitu

de Mtsuda

Carlson

Oustaloup

PID entier

PID frac-exact

-100

-50

0

50

100

150

200

Magnitu

de (

dB

)

Bode Diagram

Frequency (rad/sec)

10-3

10-2

10-1

100

101

102

103

-540

-360

-180

0

180

Phase (

deg)

PID frac-exact

Carlson

Oustaloup

PID entier

Matsuda


26 © 2014 UTA

approximations (Oustaloup, Carlson and Matsuda). The

simulation results show that the FOPID is preferment and

more robust than a traditional PID in term of degree of

stability. Therefore, we can control a such fractional system

defined by its form of generalized state by a FOPID starting

from a simple loop of regulation, but the calculation of the 05

parameters of FOPID will be very difficult in the case of the

non-integer systems not defined by its generalized state form.

REFERENCES

[1] K. B. Oldham, and J. Spanier, “The fractional calculus”, Academic Press, New York et Londres, 1974.

[2] K. S. Miller, and B. Ross, “An introduction to the fractional calculus and fractional differential equations”, John Wiley and Sons, 1993.

[3] S. G. Samko, A. A. Kilbas, and O. I. Marichev,” Fractional integrals and

derivatives:Theory and applications”, Gordon and Breach, 1987. [4] D. Matignon, and G. Montseny (dir.), “Fractional differential systems:

Models,methods, and applications”, SIAM, décembre 1998 (disp. : http://www.emath.fr/Maths/Proc/Vol.5).

[5] G. Montseny, J. Audounet, and D. Matignon, ”Diffusive representation for pseudo-differentially damped non-linear systems”, dans Isidori A.,

Lamnabhi-Lagarrigue F., Respondek W. (dir.), Nonlinear control in the year 2000 (Paris, France), Springer-Verlag, vol. 2, pp. 163-182, 2000.

[6] A. Charef et al., “Fractional System as represented by singularity function”, IEEE Transaction on Automatic Control, vol. 37, n° 9, 1992.

[7] A. Oustaloup et al., ”Frequency-band complex non integer differentiator: characterization and synthesis”, IEEE transaction on Circuit and Systems, 2000.

[8] K. Matsuda and H. Fuji, “Optimized wave-absorbing control: Analytical and experimental results”, Journal of Guidance, Control, and Dynamics, 16, No. 6, 1146-1153, 1993.

[9] G. E. Carlson and C. A. Halijak, “Approximation of fractional capacitors (1/s)1/n by a regular Newton process”, IRE Transactions on Circuit Theory, 1964.

[10] H. W. Bode, Network Analysis and Feedback Amplifier Design, Van Nostrand. New York, 1945.

[11] R.S. Barbosa, J. A Tenreiro Machado. and I. M. Ferreira, “A Fractional Calculus a Perspective of PID Tuning”, Proceedings of DETC. 03ASME. 2003.

[12] Y. Q. Chen, I. Petras and D. Xue, “Fractional Order Control” - A

Tutorial”, American Control Conference, 2009.

[13] C. A. Monje, Y. Q. Chen, B. M. Vinagre, D. Xue and V. Feliu, Fractional-Order Systems and Controls: Fundamentals and Applications. Springer, Heidelberg (2010).

[14] J. Sabatier and C. Farges, “On stability of fractional order systems”, International Journal of Bifurcation and Chaos (2011, in press).

[15] I. Podlubny, “Fractional-order systems and PIαDβ controllers”, IEEE

Transactions z on Automatic Control , 1999. [16] I. Petráš, L. Dorčák and I. Koštial, “Control quality enhancement by

fractional ordes controllers”, Acta Montanistica Slovaca Ročník 3, 2, pp. 143-148, 1998.

[17] A. Djouambi, A. Charef and T. Bouktir, “Fractal robustness and

parameter tuning PIµDm controllers”, proceedings of the 5th int. conf. on signal, speech and image processing , Corfu, Greece, pp. 158-159. August 17-19, 2005.

[18] C. A.Monje, B.M. Vinagre, Chen Y. Q., and V. Feliu, “ Proposals for fractional PID-tuning”, In Proceedings of the first IFAC Symposium on Fractional differentiation and its Application (FDA04), Bordeaux,

France, 2004. [19] H. S. Ahn and Y. Chen, “Necessary and sufficient stability condition of

fractional-order interval linear systems”, Automatica, vol. 44, pp. 2985–2988, 2008.

[20] H. S. Ahn, Y. Chen, and I. Podlubny, “Robust stability test of a class of linear time-invariant interval fractional-order system using Lyapunov

inequality”, Appl. Math. Comput., vol. 187, pp. 27–34, 2007. [21] D. Xue, C. Zhao, and Y. Q. Chen. Fractional order PID control a DC-

motor with elatic shaft: A case study. In Proceedings of the 2006 American Control Conference, pages 3182–3187,Minnesota, USA, 2006.

[22] Y. Luo, Y. Q. Chen, C. Y. Wang, and Y. G. Pi. “Tuning fractional order

proportional integral controllers for fractional order systems”, Journal of Process Control, vol. 20, pp. 823–831, 2010.

[23] S. Skogestad, I. Postlethwaite, Multivariable feedback control. Analysis and design. John Wiley &sons. 2nd edition. 2005.

[24] I. Kamwa, P. Viarouge, H. Le-Hui, J. Dickinson,”A frequency maximum likelihood of synchronous machine high order models using

SSFR data”, IEEE Trans. on Energy Conversion, Vol. 7, No.4, pp. 525-536, 1992.

[25] I. M. Canay, “Causes of discrepancies on calculation of rotor quantitiesand exact equivalent diagrams of synchronous machine”, IEEE Trans. On Power Apparatus, Vol. 88, pp. 1114-1120, 1969.

[26] D. Riu, N. Retière, M. Ivanès, “Induced currents modelling by half-order

systems. Application to hydro- and turbo-alternators”, IEEE Trans. On Energy Conversion, Vol. 18, pp. 94-99, 2003.


27 © 2014 UTA

Effects of Different Parameters on Power System Transient Stability Studies

M. Amroune Department of Electrical Engineering,

University of Setif 1, Algeria [email protected]

T. Bouktir Department of Electrical Engineering,

University of Setif 1, Algeria [email protected]

Abstract— The transient stability studies plays a vital role in providing secured operating configurations in power systems. This paper shows an analysis of the effects of various parameters on the transient stability studies of power system. The various parameters for which the analysis is presented include the Fault Clearing Time (FCT), Fault location, load increasing, machine damping coefficient D, and Generator Armature Resistance GAR. Under the condition that the power system is subjected to a three-phase short-circuit fault, the Critical Clearing Time (CCT) is calculated using numerical integration method. The analysis has been carried out on the IEEE 30-bus test system. From this analysis, we can conclude the importance of these different parameters on power system transient stability studies.

Keywords- transient stability analysis; numerical integration method; critical clearing time; damping coefficient; generator armature resistances

I. INTRODUCTION The transient stability is one of important items in the

planning and maintaining security of power system operation. A transient stability is concerned with the ability of the power system to maintain synchronism when subjected to a severe disturbance. These disturbances can be faults such as: a short circuit on a transmission line, loss of a generator, loss of a load, gain of load or loss of a portion of transmission network [1][2].

One of the requirements of transient stability analysis is to compute a transient stability index (TSI) for the contingencies, which is used to assess the stability of single contingency and furthermore rank the severity of different contingencies [3].

The Critical Clearance Time of a fault is generally considered as the best measurement of severity of a contingency and thus widely used for ranking contingencies in accordance with their severity [4]. In this paper Critical Clearing Time (CCT) is employed as a transient stability index to evaluate test system. In IEEE report [5], the Critical Clearing Time is defined as “the maximum time between the fault initiation and its clearing such that the power system is transiently stable”.

The CCT is efficient factor for estimation of transient stability limits of large power to avoid any cascading outages

which may lead to black out. The transient stability limits refers to the amount of power that can be transmitted through some point in the system with stability when the system is subjected to sever disturbance. The transient stability limits depends on duration and location of fault, construction parameters of the network and generators, and dynamic characteristics of loads [6][7][8]. In this order the main objective of this paper is to know the effects of various parameters on the transient stability studies of power system i.e. Fault Clearing Time (FCT), Fault location, load increasing, machine damping coefficient and Generator Armature Resistance GAR. The analysis has been carried out on the IEEE 30-bus test system.

Many methods for transient stability analysis and assessment have been proposed and improved over the years, such as equal area criteria, numerical integration and Lyapunove method [9][10][11], in this study the numerical integration method is required in order to get the exact CCTs. The numerical integration method is the most reliable and accurate method for transient stability assessment [12].

II. MATHEMATICAL MODEL OF POWER SYSTEM

This section gives a mathematical model for the power system network which includes modeling of synchronous machines and the network.

A. Swing equation The equation governing the motion of the rotor of

synchronous machine is based on elementary principal in dynamics which states that the accelerating torque is the producer of the moment of inertia and angular acceleration [6]:

mm e DT T TJα= + +

(1) Where Tm [Nm] is the mechanical energy input at the rotor shaft; Te [Nm] is the torque equivalent of the generator electrical output power; J [kg m2] is the combined polar moment of inertia of the rotor masses; αm [rd/sec2] is the acceleration of the rotor masses; TD [Nm] is the damping torque of the generator. The damping term TD in equation (1) is a very small percentage of Te and thus equation (1) is sometimes approximated by [13]:

mm eTJT α= + (2)

28 © 2014 UTA

This equation can be represented in terms of electrical power [10]:

( )2

2 m ed f P P

Hdtδ π= − (3)

Where δ is the electrical power angle in radians and [ ]

[ ]kinetic energy in MJ at rated speed

machine rating in MVAH = (4)

f is the system frequency; H is the inertia constant of machine expected on the common MVA base; Pm is the mechanical input power and Pe is the electrical output. It is often desirable to include a component of damping not accounted for in the calculation of Pe separately. This is accomplished by adding a term PD proportional to speed deviation in the above equation as follows [14]:

( )2

2 m Ded f P P

HdtPδ π

= − − (5)

Where

DdP Ddtδ

= (6)

D is the generator damping coefficient.

B. Electrical network modelling The machine classical electromechanical model is

represented by the following differential equations:

( )2

2

ii s

i ii

Dm e iid f P P

Hd

d

Pt

dtδ

ω ω

δ π= −−

= −

(7)

The electrical power of the ith generator is given by [6]:

( )2

1

sin 90m

ei i ii ij i j ijj

P E R C δ δ θ=

⎡ ⎤= + − + −⎣ ⎦∑ o (8)

Where i = 1, 2,3,…m is the number of synchronous machines. Cij = |Ei||Ej||Yij| is the power transferred at bus ij. E is the magnitude of the internal voltage. Yij are the internal elements of matrix Y. Rii are the real values of the diagonal elements of Y.

III. TRANSIENT STABILITY EVALUATION

Transient stability analysis is used to investigate the stability of power system under sudden and large disturbances, and plays an important role in maintaining security of power system operation. The transient stability analysis is performed by combining a solution of the algebraic equations describing the network with numerical solution of the differential equations. However, due to the non-linearity of the differential equations, the solving process is tedious and complicated. Thus, numerical integration methods have been applied to examine a system’s stability.

In order to reduce the complexity of the transient stability analysis for the considered test systems, the following assumption are accepted [10]: (i) Each synchronous machine is represented by a constant voltage source behind the direct axis

transient reactance. (ii) The governor’s action are neglected and the input powers are assumed to remain constant during the entire period of simulation. (iii) Using pre-fault bus voltage, all loads are converted to equivalent admittances to ground and are assumed to remain constant. (iv) The mechanical rotor angle of each machine coincides with the angel of the voltage behind the machine reactance. (v) Machines belonging to the same station swing together and are said to be coherent. A group of coherent machines is represented by one machine.

A. Solution steps The algorithm for the transient stability studies involves the

following steps:

• Reads the line and bus data. It includes the data forlines, transformers and shunt capacitors.

• Form admittance matrix, Ybus.

• Solve the initial load flow.

• Reads generator data.

• Modify Ybus by adding the generator and loadadmittances.

• Compute the pre-fault admittance matrix Ypre-fault byeliminating all nodes except the internal generatornodes.

• Solve the generator swing equation for the pre-faultperiod.

• Set t = 1s a three-phase fault occurs at any line in thesystem, and fault bus voltage equal to zero.

• Compute the new faulted admittance matrix Yfault.

• Solve the swing equation for the fault period.

• Isolate the line witch fault occurred.

• Compute the post-fault system admittance matrix Ypost-

fault.

• Solve the swing equation for the post fault period.

• Plots the swing curves for all generators.

In this paper, we define the CCT as the small lest from all CCTs values for different generators.

IV. SIMULATION AND RESULTS

This section presents computer simulation studies with programs developed in the environment of MATLAB software. The analysis has been carried out on the IEEE 30-bus test system shown in Fig. 1. It consists of 6 synchronous machines, 24 loads and 41 transmission lines. Detailed parameters of this system can be found in [15].

M. Amroune and T. Bouktir Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 28-33

29 © 2014 UTA

Figure 1. Single line diagram of the IEEE 30-bus system

The test system is analyzed using optimal power flow (OPF). A standard OPF problem can be formulated as follows [16][17]:

( ) 2

1( )

ng

i i gi i gii

F x a b P c P=

= + −∑ (9)

( ), 0i di giP V P Pθ + − = (10)

( ), 0i di giQ V Q Qθ + − = (11) min max

gi gi giP P P≤ ≤ ;i=1,...ng (12) min maxgi gi giQ Q Q≤ ≤ ;i=1,...ng (13)

Where F(x) is a cost function; ai, bi, ci are cost coefficients shown in the Table 1; Pgi and Qgi are the active and reactive power generations at ith bus; Pdi and Qdi are the active and reactive power demands at ith bus; Pi and Qi are the active and reactive power injections at ith bus. The Optimal Power Flow results obtained with used of the Lambda Iteration Method [10] are listed in Table 2.

TABLE 1. GENERATOR COST COEFFICIENTS

Cost Coefficients BUS a b c

PGMIN

(MW) PG

MAX

(MW) 1 0 2 0.00375 50 200 2 0 1.75 0.0175 20 80 5 0 1 0.0625 15 50 8 0 3.25 0.0083 10 35

11 0 3 0.025 10 30 13 0 3 0.025 12 40

TABLE 2. OPF RESULTS

BUS PG (MW)

1 176.9202 48.4515 20.9728 22.432

11 12.14613 12.000

P loss 9.522 Total cost 802.35 ($/h)

The analysis of effects of various parameters on power system transient stability studies is presented here. The various parameters for which the analysis is presented include the Fault Clearing Time (FCT), fault location, load increasing, generator damping coefficient D and generator armature resistances GAR.

A. Effect of Fault Clearing Time (FCT) In order to know the effect of Fault Clearing Time (FCT)

on transient stability a disturbance in the form of a three phase to ground fault is occurs at t = 1 second at bus 1, cleared by opening the line connecting the nodes 1–2. The rotor angle differences are shown in Figure 2. If the fault is cleared rapidly the angular deviation is less and subsequently the system may become stable. This angular deviation increases if the fault clearing time increases and ultimately if the fault is cleared after Critical Clearing Time (CCT) the system will lose synchronism. In this case the CCT is equal to 166 ms.

Figure 2. Rotor angle differences with fault at bus 1

B. Effect of fault location In this sub-section the effects of fault location in transient

stability are analyzed. A three-phase fault is located at two different locations, one closer to the generating stations (at bus 1 with opening the line 1–2), in this case the CCT is equal to 166 ms (Fig. 1), the other one far from the generating stations (at bus 6 with opening the line 4–6). Fig. 2 shows the angular positions of the machines in the system for a fault on line 4–6. It is found that the CCT is equal

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

20

40

60

80

100

120

140

160

180

200

t, sec

Ang

le ro

toriq

ue re

latif

, deg

ree

δ1-δ2δ1-δ3δ1-δ4δ1-δ5δ1-δ6

FCT = 100 ms

FCT = 166 ms

FCT = 167 ms***


30 © 2014 UTA

to 620 ms, which is higher then the CCT of the fault at bus 1 (on line 1–2). Table 3 gives the CCTs for different fault locations in IEEE 30-bus system. It is clear that the fault that is nearer to the generating station must be cleared rapidly than the fault on the line distant from the generation station.

Figure 3. Rotor angle differences with fault at bus 6

TABLE 3. CCTS FOR DIFFERENT FAULT LOCATION IN IEEE 30-BUS

Faulted bus Open line CCT (ms) 1 1 – 2 166 1 1 – 3 219 2 1 – 2 236 2 2 – 4 299 2 2 – 5 306 2 2 – 6 300 3 3 – 4 537 3 1 – 3 518 4 2 – 4 505 4 3 – 4 513 4 4 – 6 514 6 2 – 6 607 6 4 – 6 620

C. Effect of load increasing The man objective of this sub-section is to know the impact

of load increasing on the power system Critical Clearing Time. For this reason, active load at all buses in the IEEE 30-bus system are increased from base case by 10%, 20%, 30%, and 40%. Real example of this case is electrical peak load of energy consumption.

Figure 4 and Figure 5 shows respectively the power generation and system voltage magnitudes for total loading change. It is observed that more then allowed level of load increasing, power generation increased and voltage at all buses dropped.

Figure 4. Power generation

Figure 5. Voltage magnitude values

In order to evaluate the effects of load variation in transient stability analysis, a three-phase fault occurs at bus 1 with opening the line 1–2. The obtained CCTs for different load values are listed in Table 4. It is very clear that the effect of load increasing in decreasing the CCT.

TABLE 4. CCTS FOR TOTAL LOADING CHANGE

Total loading change (%)

Base case 10 20 30 40

Total loading change (MW) 283.4 311.74 340.08 368.24 396.76

CCTs (ms) 166 133 110 107 104

D. Effect of damping coefficient D The machine damping coefficient D represents the natural

damping of the system. In this sub-section, the effect of this coefficient on the transient stability evaluation has been investigated.

Fig.6 and Fig.7 show the rotor angle differences with and without of damping coefficient for a short circuit at 1, the FCT in this case is set at 100 ms.

From figures it can be seen that under the fault occurred with consideration of damping coefficient, the oscillation amplitudes and shapes are different. The damping coefficient prevents the growth of oscillations; when the FCT = 110 ms

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

50

100

150

200

250

300

350

400

t, sec

Ang

le ro

toriq

ue re

latif

, deg

ree

δ1-δ2δ1-δ3

δ1-δ4

δ1-δ5

δ1-δ6

FCT = 620 ms

FCT = 621 ms Total loading change

Pow

er g

ener

atio

n


31 © 2014 UTA

(Fig. 7) the system with D remains stable and can return to steady state finally. However, the system with D neglected is unstable. Table 5 shows how the CCTs increase with including the damping coefficient D for the different fault locations.

Figure 6. Rotor angle differences with fault at Bus 1 (FCT = 100 ms)

Figure 7. Rotor angle differences with fault at Bus 1 (FCT = 110 ms)

TABLE 5. CCTS WITH AND WITHOUT OF DAMPING COEFFICIENT

Faulted bus Open line CCT without D

(ms) CCT with D

(ms) 1 1 – 2 105 166 1 1 – 3 166 219 2 1 – 2 156 236 2 2 – 4 219 299 2 2 – 5 222 306 2 2 – 6 220 300 3 3 – 4 384 537 3 1 – 3 372 518 4 2 – 4 365 505 4 3 – 4 371 513 4 4 – 6 380 514 6 2 – 6 365 607 6 4 – 6 362 620

E. Effect of Generator Armature Resistances (GAR) The influence of generator armature resistances on transient

stability limits is presented in this sub-section. Two scenarios were analyzed. In the first one the GAR is included; in the second situation the GAR is neglected. The rotor angle differences with and without considering of GAR are shown in Figure 8. The FCT is set at 168 ms, the system without armature resistances is go out of step (FCT > CCT). However, the system with armature resistances is stable.

Figure 8. Rotor angle difference with fault at Bus 1 (FCT = 168 ms)

Another’s simulations have been performed for different fault locations IEEE 30-bus system, in order to compare accurately CCTs with and without generator armature resistances. The results from the cases study are presented in Table 6 and Figure 9. The comparative results have shown that the impact of generator armature resistances in transient stability analysis. From the obtained results it is investigated that the generator armature resistance has an effect on the transient stability analysis. In some cases the ΔCCT is expected values 5 and 6 ms for example the fault at bus 2 with openings of the circuit breakers at both ends of line (2–5), faults at bus 4(2–4), 4(3–4), 4(4–6) and 6(4–6). It is very clear that the effect of generator armature resistances in power system Critical Clearing Time. For this reason the transient stability analysis of power system can be accurately represented by including the armature resistances of the synchronous machines.

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0

50

100

150

t, sec

Ang

le ro

toriq

ue re

latif

, deg

ree


With D

Without D

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

0

50

100

150

200

250

300

t, sec

Ang

le ro

toriq

ue re

latif

, deg

ree


With D

Without D

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50

20

40

60

80

100

120

140

160

180

200

t, sec

Ang

le ro

toriq

ue re

latif

, deg

ree


With GAR

Without GAR


32 © 2014 UTA

TABLE 6. COMPARATIVE ANALYSIS OF CCTS FOR IEEE 30-BUS SYSTEM

Faulted bus Open line

CCT without

GAR (ms)

CCT with GAR

(ms)

ΔCCT (ms)

1 1 – 2 166 170 4 1 1 – 3 219 224 5 2 1 – 2 236 239 3 2 2 – 4 299 303 4 2 2 – 5 306 311 5 2 2 – 6 300 304 4 3 3 – 4 537 539 2 3 1 – 3 518 519 1 4 2 – 4 505 500 5 4 3 – 4 513 508 5 4 4 – 6 514 508 6 6 2 – 6 607 603 4 6 4 – 6 620 614 6

Figure 9. CCTs with and without GAR for IEEE 30-bus system

V. CONCLUSION In this study, the effects of various parameters on the

transient stability studies of a power system is presented and discussed. The parameters investigated were the Fault Clearing Time, fault location, different load levels, generator damping coefficient and generator armature resistances. A numerical integration method is used to compute CCTs, and IEEE 30-bus test system is required.

According to the simulation results, some preliminary conclusions and comments can be summarized as follows:

Under three-phase short-circuit fault, the rapid clearing of the fault promotes power system stability;

The fault that is nearer to the generating station must be cleared rapidly than the fault on the line distant from the generation station;

More then allowed level of load increasing a power generation increased, voltage at all buses dropped and Critical Clearing Time decreased.

The damping coefficient prevents the growth of oscillations and improves the power system Critical Clearing Time; Generator armature resistances have an effect on results of the transient stability analysis, for this reason the transient stability analysis of power system can be accurately

represented by including the armature resistances of the synchronous machines.

REFERENCES [1] P. Kundur and J. Paserba, “Definition and classification of power system

stability,” IEEE Trans. on Power Systems, 19, 2, pp. 1387-1401, 2004.

[2] Abdul Malek Miah , “A new method of transient stability assessment by using a simple energy margin function,” Second International Conference on Electrical and Computer Engineering Dhaka, Bangladesh, pp. 24–27, 26-28 December 2002.

[3] Ancheng Xue, Chen Shen, Shengwei Mei, “A New Transient Stability Index of Power Systems Based on Theory of Stability Region and Its Applications,” IEEE 2006.

[4] R. E br ahimpour , E. K. Abharian, S. Z. Moussavi & A. A. MotieBirjandi, “Transient stability assessment of a power system by mixture of experts,” International Journal of Engineering (IJE), 4 (1), pp. 93–104, 2010.

[5] IEEE Committee Report, “Proposed terms and definitions for power system stability,” IEEE Trans. Power App. Syst, vol. PAS-101, pp. 1894–1898, 1982 .

[6] P.K. Iyambo, R. Tzoneva, “Transient stability analysis of the IEEE 14-bus electric power system,” IEEE 2007.

[7] M. R. Agha mohammadi, A. Beik Khormizi, M. Rezaee, “Effect of Generator Parameters Inaccuracy on Transient Stability Performance,” IEEE 2010.

[8] Elmer Sorrentino, Orlando Salazar, Daniel Chavez “Effect of Generator Models and Load Models on the Results of the Transient Stability Analysis of a Power System”, IEEE 2010 .

[9] Y. Xue, Th. Van Cutsem, M. Ribbens-Pavella "Extended equal area criterion justification, generalization, applications", IEEE Transactions on Power Systems, 4, 1, February 1989.

[10] H. Saadat, Power system analysis. Second Edition. McGraw-Hill. USA, 1999.

[11] GERD A. LUDERS, "Transient stability of multi-machine power systems via the direct method of Lyapunov", IEEE Transaction on Power Apparatus and Systems, Vol. PAS-90, No. 1, January/February 1971

[12] P. Kundur, Power System Stability and Control. New York: Mc Graw-Hill, pp 104-120, 1994.

[13] G. W. Stagg and A. H Fl-Abiad, Computer Methods in Power System Analysis. New York: McGraw-Hill, pp. 150-164, 1968.

[14] Naoto Yorino, Yoshifumi Kamei, Yoshifumi Zoka “a new method for transient stability assessment based on critical trajectory”, 15th PSCC, Liege, 22-26 August 2005

[15] E.DeTuglie ,S.M.Iannone ,F.Torelli “Acoherency-based method to increase dynamic security in power systems”, Electric Power Systems Research 78, 1425 – 1436, 2008.

[16] Tarek Bouktir , Linda Slimani , M. Belkacemi, “A Genetic Algorithm for Solving the Optimal Power Flow Problem”, Leonardo Journal of Sciences, 4, p. 44-58, January-June 2004.

[17] P. Somasundaram and, N.B. Muthuselvan, “a modified particle swarm optimization technique for solving transient stability constrained optimal power flow”, Journal of Theoretical and Applied Information Technology, pp. 154-164, 2010 .


33 © 2014 UTA

Energy Conversion System Performance and Analysis

Abstract— in this paper, an energy conversion system

consisting of a DFIM powered by a three-level inverter is

considered. For this, we present the models of the studied

system and the control strategy used for the tree level

inverter. Using simulation analysis, we proceed to analyze

the characteristics and the performance of the doubly fed

induction machine fed by the multilevel inverter and to test

the rate of harmonic distortion of the rotor stator and rotor

currents obtained for the considered configuration. The

simulation results obtained by Matlab/simiulink are shown

and analyzed.

Keywords- DFIM; multilevel inveter; performances; modeling; simulation.

I. INTRODUCTION

The three-phase induction motor fed by a voltage inverter is a drive system possessing many advantages: simple structure, robust and cheap. This set converter-machine remains restricted to the lower limit of the range of high power (up to several MW), due to the constraints faced by electric semiconductors and their low operating frequency. To ensure an electric drive for high power applications, such as rail traction or marine propulsion, It is often necessary segmented power. For this, we can act at the converter using multilevel techniques or paralleling converters. Literature attests to the great interest now attaches a doubly fed machine for various applications: as a renewable energy generator or as a driver for some industrial applications such as rolling, rail traction or marine propulsion. Many applications use this type a machine as: hyposynchrone

cascade, the variation of the rotor resistance, the generator operation with a variable speed, power to the stator and the rotor by a converter, or the stator by a fixed network and the rotor by a variable power supply which may be a voltage source or current source, the latter presents many advantages compared to that in tension, as it provides great flexibility and greater ease of operation [1]. DFIM with its dual power offers several possibilities for reconfiguring the operating mode of the machine. Indeed, with the renewed interest in the renewable energy, variable speed wind systems with DFIM experiencing a great boom and a large number of applications accompanying this development. Configuration, widely used in variable speed wind systems with DFIM, is shown in Figure 1, It consists in feeding the rotor by a converter and the stator to bind directly to the network.

In this paper, we present the modeling of DFIM and simulation results for its association with PWM tree level inverter, allowed to determine and analyze the different characteristics torque, speed, current and flows for different values of frequency and load torque.

II. MODELING OF THE DOUBL FED INDUCTION MACHINE

The DFIM with the distributions of its windings and its own geometry is very complex. Therefore to analyze it, its exact configuration is taken into account. It is then necessary to adopt the following simplifying assumptions in order to develop a simple model [2],[3],[4] . The machine is symmetrical and the air gap is constant; the magnetic circuit is not saturated and it is perfectly laminated, with the result that the iron losses and hysteresis are negligible and only the windings are driven by currents; the mmf created in one phase of stator and rotor are sinusoidal distributions along the gap. Thus, the electrical equations and the flux linkage expressions of the DFIM in the synchronous reference frame (d_q) are given by:

Y. Soufi T. Bahi S. Ghoudelbourk H. Merabet S. Lekhchine

Electrical Engineering Department Electrical Engineering Department Electrical Engineering Department Tebessa University

Algeria [email protected]

Badji Mokhtar University Algeria.

[email protected]

Skikda University Algeria

[email protected]

Badji Mokhtar University Algeria

[email protected]

Badji Mokhtar University Algeria

[email protected]

34 © 2014 UTA

drsqr

qrrqr

qrsldr

drsdr

dseqs

qssqs

qseds

dssds

.-dt

d+.iR=V

.-dt

d+.iR=V

.-dt

d+.iR=V

.-dt

d+.iR=V

l

(1)

qrrqssqr

qrqssqs

drdssdr

drdssds

iLiLiMiLiLiLiMiL

..

......

(2)

With: Vds, ids, Φds are respectively the « d » components of the stator voltages, current and flux linkage; Vqs, iqs, Φqs are respectively the « q » components of the stator voltages, current and flux linkage; Vdr, idr, Φdr are respectively the « d » components of the rotor voltages, current and flux linkage; Vqr, iqr, Φqr are respectively the « q » components of the rotor voltages, current and flux linkage; Rs and Rr are respectively the per phase stator and the per phase rotor resistance; we is the speed of the synchronous reference frame; Ls, Lr, M are respectively the per stator self inductance and the per phase rotor self inductance and the mutual inductance between stator and rotor.

The electromagnetic torque is evaluated as: )..(. qrdsdrqsem iiiiMpT

(3) With: p is the number of pole pairs.

The dynamical equation of machine is described as:

flem kTTdt

dJ (4)

So, the global model of the DFIM is presented by the

following expression:

lrf

dsqrqsdrr

qrqsqrdrqsdsrq

drdsqrdrqsrds

drqsqrdrrqsds

drdsqrrdrqsds

TJp

Jk

iiiiJMp

dtd

VBVBiAiAiAiA

VBVBiAiAiAiA

VBVBiAiAiAiA

VBVBiAiAiAiA

..)...(.

......dt

di

.......dt

di

......dt

di

......dt

di

2

123133343132r

323134333231dr

121113141112qs

121114131211ds

(5)

Where,

srssr L

MATL

MAAA.

;);.1(;.T1- 141312

s11

.)(;.T1-;

.; 34

r333231

r

ssss

AAL

MATL

MA

r323112

s11 .L

1;-;;.L1

B

LLMB

LLMBB

rsrs

III. MODELING AND CONTROL

A. Structure of NPC three level inverter

Figure 1 shows the structure of a three-phase three level inverter. The DC voltage source is formed by the series connection of two groups of capacitors provides at point (0) a half-voltage E/2.

This structure outputs three voltage levels -E/2, 0, and E/2 according to the configurations defined in Table 1.

Table 1: Output voltages of the three level inverter according to the statements of switches.

Sequence Ki1 Ki2 Ki3 Ki4 Ci Uio

1 0 0 1 1 -1 -E/2

2 0 1 1 0 0 0

3 1 1 0 0 1 E/2

The three-level inverter output voltages are obtained using the following expression [5]:

Fig. 1. Structure of NPC three-level inverter

Y. Soufi et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 34-38

35 © 2014 UTA

(6) With: Ci = -1; Ci = 0 or Ci = 1;

(7) Now if you consider all the switches, all phases, we

obtain 83 = 64 possible combinations [6] with 27 are presented in Table .2.

Table .2. Output voltages of the three levels inverter

N0 Ka1 Ka2 Kb1 Kb2 Kc1 Kc2 Ca Cb Cc Uao Ubo Uco

1 0 0 0 0 0 0 -1 -1 -1 -E/2 -E/2 -E/2

2 0 0 0 0 0 1 -1 -1 0 -E/2 -E/2 0

3 0 0 0 1 0 0 -1 0 -1 -E/2 0 -E/2

4 0 0 0 1 0 1 -1 0 0 -E/2 0 0

5 0 1 0 0 0 0 0 -1 -1 0 -E/2 -E/2

6 0 1 0 0 0 1 0 -1 0 0 -E/2 0

7 0 1 0 1 0 0 0 0 -1 0 0 -E/2

8 0 1 0 1 0 1 0 0 0 0 0 0

9 0 1 0 1 1 1 0 0 1 0 0 E/2

10 0 1 1 1 0 1 0 1 0 0 E/2 0

11 0 1 1 1 1 1 0 1 1 0 E/2 E/2

12 1 1 0 1 0 1 1 0 0 E/2 0 0

13 1 1 0 1 1 1 1 0 1 E/2 0 E/2

14 1 1 1 1 0 1 1 1 0 E/2 E/2 0

15 1 1 1 1 1 1 1 1 1 E/2 E/2 E/2

16 1 1 0 0 0 1 1 -1 0 E/2 -E/2 0

17 1 1 0 1 0 0 1 0 -1 E/2 0 -E/2

18 0 1 1 1 0 0 0 1 -1 0 E/2 -E/2

19 0 0 1 1 0 1 -1 1 0 -E/2 E/2 0

20 0 0 0 1 1 1 -1 0 1 -E/2 0 E/2

21 0 1 0 0 1 1 0 -1 1 0 -E/2 E/2

22 1 1 1 0 0 0 1 -1 -1 E/2 -E/2 -E/2

23 1 1 1 1 0 0 1 1 -1 E/2 E/2 -E/2

24 0 0 0 1 0 0 -1 1 -1 -E/2 E/2 -E/2

25 0 0 0 1 1 1 -1 1 1 -E/2 E/2 E/2

26 0 0 1 0 1 1 -1 -1 1 -E/2 -E/2 E/2

27 1 1 1 0 1 1 1 -1 1 E/2 -E/2 E/2

B. Three level inverter PWM Algorithm control The control signals of the switches of the NPC three level

inverter are obtained from the intersections of three sinusoidal reference signals out of phase with each other by 120° and the carrier. It uses a single carrier with three reference signals to generate the control of each phase see

fig. 3. The control algorithm triangular-sine for the three levels can be summed up in one arm according to the following:

Table. 3. Control algorithm three level inverter

Test Uref>0 Uref<=0 Urefi >=Uport Urefi < Uport Urefi >=Upor Urefi< Uport

Ui0 E/2 0 0 -E/2

IV. SIMULATION RESULTS AND DISCUSSION

The modeling and simulation of DFIM were used to determine and analyze the different characteristics of the torque, speed, flow and currents for different values of frequency and load torque. We tested the operation of the stator and rotor machine supplied by a three-level inverter. This is to evaluate the performance of machine operation. In order to study the behavior of the machine at starting, followed by applying a load torque. Figures 3 shows the electrical and mechanical characteristics and Figure 4 shows the stator and rotor currents and their spectral analysis, obtained by numerical simulation. Figure 3.a and 3.b shows the voltage at the output of the inverter supplying respectively the stator and rotor of the machine.The DFIM fed by tree level inverter (fs = 50 Hz) with short circuit rotor at startup and when t = 1 s are supplied with the rotor Ur = 10V and fr = 5Hz, and there was a decrease in the speed of rotation of the rotor to food (see Fig 3c), and the couple reacts to this state and its oscillations increase (see Fig 3d) from t = 2s until the machine was loaded with a torque of 25 Nm. The frequency analysis of the DFIM stator and rotor current of the DFIM fed by a three-level inverter-controlled PWM are presented and analyzed in Figure 4. The stator and rotor currents are shows in Figure 4.a and 4.c. The THD of the stator and rotor currents ( see Fig 4.b and Fig 4.d) are respectively 20.87% and 39.48%.

Fig. 2. Control signals of switches


36 © 2014 UTA

Fig. 3. Characteristics of DFIM fed by tree level inverter

0 0.5 1 1.5 2 2.5 3 3.5 4-20

-15

-10

-5

0

5

10

15

20

Time (s)

Rot

or v

olta

ge (V

)

0 0.5 1 1.5 2 2.5 3 3.5 4-400

-300

-200

-100

0

100

200

300

400

Time (s)

Sta

tor v

olta

ge (V

)

0 0.5 1 1.5 2 2.5 3 3.5 4-20

0

20

40

60

80

100

120

140

160

Time(s)

Spee

d (rd

/s)

0 0.5 1 1.5 2 2.5 3 3.5 4-60

-40

-20

0

20

40

60

80

100

120

140

Time (s)

Torq

ue (N

.m)

Fig. 4. Stator and rotor currents analysis

0 0.5 1 1.5 2 2.5 3 3.5 4-80

-60

-40

-20

0

20

40

60

Time (s)

Sta

tor c

urre

nt (A

)

0 0.5 1 1.5 2 2.5 3 3.5 4-60

-40

-20

0

20

40

60

80

Time (s)

Rot

or c

urre

nt (A

)

-a-

-b-

-c-

-d- -d-

-c-

-b-

-a-


37 © 2014 UTA

V. CONCLUSION The developed model of DFIM and its power via an inverter three levels have been validated by numerical simulation. This model is nonlinear and strongly coupled. To overcome this difficulty, we turned to Park transformation. The structure of the considered system has allowed to visualize and analyze the characteristic of the considered machine. The simulation of DFIM fed by three level inverter controlled by PWM control strategy, shows that the speed and torque have good performance and more analysis of stator and rotor currents are presented and analyze.

REFERENCES

[1] M. Cherkaoui. "Contribution a la Modelisation, a l’Etude et a la Commande des Machines Alternatives Application à une Machine Asynchrone à Double Alimentation". Thèse de Doctorat, Institut National Polytechnique de Lorraine. E.N.S.E.M., France,1990.

[2] C.R Kelber, W. Schumcher, "Frequency energy generation with Doubly Fed Induction Machine", Proc. VSSHy European Conference on Variable Speed in small hydro (Grenoble), 2000.

[3] G. Salloum, " Contribution à la commande robuste de la machine asynchrone à double alimentation ". Thèse de Doctorat de l’INP de Toulouse, France, 2007.

[4] T. Bahi, A. Dekhane, H. Merabet, Y. Soufi, " Modeling and control of a doubly fed induction generator, "Ecological vehicles and renewable energies" International Conference and Exhibition, Ever Monaco, 2011.

[5] S. Drid, S.M-S.Nait, M. Tadjine, "The Doubly Fed Induction Machine Modeling In The Separate Reference Frames", Journal of Electrical Engineering, JEE , 4, 1, pp. 11-16, 2004.

[6] F. Bonnet, P Vidal and M pietrzk-david, "Dual Direct Torque control Of Doubly fed Induction Machine", IEEE Trans Industrial Electronices, 54, 5, 2007.

APPENDIX MACHINE PARAMETERS

Power motor 7.5KW

Stator phase résistance Rs =0.455Ω

Rotor phase résistance Rr =0.62Ω

Inductance phase stator Ls =0.084H

Inductance phase rotor Lr =0.081H

Mutual Inductance Lm 0.078H

Moment of inertia J= 0.3125 Kg.m2

Friction coefficient Kf =6.73.10-3 N.m.s-1

Number of pole pair 2


38 © 2014 UTA

Face localization using neural networks trained with geometric and skin characteristics

M. Saaidia, A. Gattal and M. Maamri Dept. of electrical Engineering

University of Tebessa Tebessa, Algeria

[email protected], [email protected] and [email protected] .dz

M. Ramdani Dept. of electrical

Engineering University of Annaba Annaba, Algeria

Ramdani.m@y ahoo.fr

Abstract- Geometric and skin face characteristics were

used, in this paper, to localize face in images. In first

stage, well known horizontal symmetric characteristic

of face is exploited to localize the vertical symmetry axis of the face, then to delimit its vertical position in

the image. In second stage, a neural network, trained

to recognize pixel with skin characteristics, is used to

completely localize the face by delimiting the region

surrounding the vertical symmetry axis which colour

characteristics agree with skin ones according to TSL

colour space. At this stage, we explore three strategies

to perform region delimitation. Finally, a quantitative

measurement criterion is used to record localization

quality. Experiments of the proposed method were

carried out on images of XM2VTS database and also on a set of non standard images.

Keywords-component; face localization, face detection,

autocorrelation, neural networks, TSL

I. INTRODUCTION

Face localization or more generally face detection is

being of great importance not only in research aspects but

also in human daily life. This was possible mainly

through the great technological developments essentially

in information sciences were the machines are faster,

more complex and more efficient. Thus, for face

recognition (identity check), face expressions analysis or

to take movements of face parts into account (gesture

communication); localization of face in image or in video

acquired by various peripherals (cameras, scanner, infra-

red...) is necessary to the achievement of these operations.

Several ways were explored by the researchers which

gives a great number of methods. Classification of all

these methods depends on the parameter’s classification.

According to Hjelmas et al. [1], two principal approaches

can be distinguished; the global approach which consists

in entirely seeking the face and the components approach

which consists in finding the face through the localization

and the regrouping of its components (eyes, nose...). For

Yang et al. [2] there are four major classes; knowledge-

based methods, feature invariant approaches, template

matching methods and appearance based methods.

According to one or the other of these classifications,

each developed method exploits one or more

characteristics of face like colour, shape, movement etc, to perform face detection.

In this proposed work we exploit two principal

characteristics of the face to perform its localization in the

image. Vertical symmetry, of the face is used in first to

determine the principal vertical symmetry axis in the

image which is probably the vertical symmetry face’s

axis. To delimit the vertical zone of the face,

autocorrelation function is then applied to pixels on both

sides of this axis. In second stage, colour and appearance

skin characteristics were used to search the face region

within the previously delimited zone. In this goal a neural

network trained to recognize pixel with skin

characteristics was applied to pixels within the delimited

region. An eleptical delimitation strategie was explored to

perform the scan of that region. Offline measurement of

method performance was done according to a quantitative

measurement criterion [3]. The proposed method was

experienced on XM2VTS database and a set of non-

standard images.

In section 2 we explain the proposed localization

method then in section 3, the technique to determine

method performances will be presented. Section 4 will

contain the experiments and comparison results and in section 5 we will conclude.

II. FACE LOCALIZATION METHOD

Symmetry characteristic of the face is the most apparent

geometric particularity. It was usually used in

researchers’ works for face's detection [4] and [5] or for

the detection of feature's face [6] and [7]. The advantage

of this characteristic is its universality (independent of the

ethnicity of the person) and its robustness against the

different conditions and also perturbations produced by

image’s acquisition process like illumination, presence or

absence of structural elements (glasses, moustache, …),

presence or absence of facial expressions, image quality,

neutral or complex background, etc , (figure 1).

39 © 2014 UTA

Figure 1. Images from different databases (Yale, JAFFE and XM2VTS) with different quality and acquisition

conditions

However, due to the fact that it’s not restrictive to the

face, this characteristic is still largely insufficient to

resolve face’s detection problem. Researchers uses it in

second phase of the detection process to confirm the first

phase in which a region was localized as candidate to be a

face [8] or to extract the eyes position on a region

which is considered as a face [9].

In our proposed method we exploit this characteristic first

to determine the principal symmetrical axis of the image

using the fact that if it contains a face, the image’s axis

symmetry will be the same as the face one. Then, pixels

on both sides of this axis are correlated to extract the

vertical region which contains the face.

A. Image principal symmetrical axis

The presence of a face in an image, affects directly its

symmetrical properties; essentially when it contains only one. Based on this observation, we propose an efficient

and fast procedure to determine the principal symmetrical

axis, of the treated image, using correlation function

according to equation (1).

)),(2),,(1(C(j) jIMatjIMatCor= (1)

where:

jpNjMIwith

pjIMatjIMat

and

jIMat

:1),2/(,...,3,2,1,:1:

)1*2,(),(2

),(j)Mat1(I,

===

+−=

=

Mat is an ( MxN) matrix containing treated image values;

Cor, the well known correlation function and C the

obtained curve representing the energy evolution of the

compiled correlation between the matrices Mat1 and

Mat2 .On figure 2, we give an example of images

representing some steps of the compilation of C

according to equation (1).

(a)

(b)

(c)

Figure 2. Image’s symmetrical axis compilation (a) original image from Yale database, (b) sample’s of pair

images representing Mat1 and Mat2 evolution, (c)

Correlation energy compilation.

The curve of figure 2.c represents the evolution of the

correlation energy which indicates the level of similarity

between the two matrices Mat1 and Mat2. The maximum

level indicates the greatest similarity and therefore the

position of the symmetrical axes of the image which will

be probably the principal symmetrical axis of the face.

Figure 3. Symmetrical axis determination

The examples above are chosen to demonstrate the

robustness of the symmetry characteristic of the face for

different situations of image capture, like illumination and background nature; face gender, face expression and face

position. These images were taken from the well known

databases Yale, JAFFE, ORL and XM2VTS and from a

set of non standard images used here to demonstrate more

generalization.

M. Saaidia et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 39-44

40 © 2014 UTA

B. Vertical face zone

The face’s symmetrical axis being found, the vertical

delineation of the face will be easy to compile according

to equation (2)

)),(),,(((j)lr jICjICCorC rl= (2)

where:

1,...3,2,1

,:1:

),(),(

),(j)(I,

=−⋅⋅⋅⋅⋅⋅=+⋅⋅⋅=

=

+=

−=

jJaorNjJauntilj

MIwith

jJaIMatjIC

and

jJaIMatC

r

l

Vertical delineation of the face is based on the

compilation of the correlation function, Clr, between the

columns on both sides of the symmetrical axis,

determined here by index Ja, found in first stage.

(a)

(b) (c)

Figure 4. Example of vertical delimitation of the face,

(a) original image from XM2VTS database, (b) facial

symmetrical axis determination, (c) vertical delineation of

the facial region.

Figure 4 shows an example of vertical delineation of

the face according to the compilation of Clr.

To demonstrate performances obtained in this step

we show in figure 5 some examples from the databases

sited earlier.

Figure 5. Examples of face vertical region delimitation

C. Face region localization

In this phase we will explore the way to determine the

exact area of the face inside or closest outside areas to the

vertical boundaries obtained in the first phase.

Proposed approach is based on the use of a neural network unit preliminarily trained to recognize pixel

which colour characteristics agree with those of a skin one

according to the TLS colour space representation.

Equation 3 gives TSL colour space formulation:

[ ]

BGRL

g

ggr

ggr

T

grS

114.0587.0299.0

0,2/1

0,4/32/)/arctan(

0,4/12/)/arctan(

(5/92/122

++=

=′

<′+′′

>′+′′

=

′+′=

KKKKKKKKK

K

K

π

π

Where r’ and g’ are the normalized components of R and

G in the RGB colour space.

Indeed, Skin colour has proven to be a useful and robust

cue for face detection, localization and tracking [10]. Skin

colour modelling problem was widely studied for

computer graphics, video signal transmission and

compression, etc. These studies have led to the emergence

of several colour spaces like RGB, HSI, YCrCb, YUV,

TSL, …

TSL colour space is a transformation of the normalized

RGB one. Compared to other colour spaces, TSL

appears more adapted for skin modelling [11]. This

modelling way was mainly used for face detection

[12], [13] and [14].For the method presented in this paper we choose to

save time compilation by limiting the use of the

colour characteristics, for skin recognize, only on the

pixels of the region preliminarily delimited in first stage.

Figure 6. Examples of rectangular face delimitation

General shape of the face is very similar to that of an

ellipse and even in the case where it is different; the

components of the face can be gathered by an elliptical

form. Therefore, elliptical model was largely used in

(3)


41 © 2014 UTA

algorithms for face detection and recognition [15] and

[16]. In our strategy, we choose to exploit this characteristic to enhance the delimitation procedure of the

face’s zone. To do so, the delimitation procedure exploits

the principal symmetrical axis which was found at the

first phase. Skin colour characteristic is then used to lead

the delimitation procedure, according to two steps:

(i) First we apply the neural network unit to the pixels,

one by one, of the symmetrical axis in the image (in the

two directions: pixel 1 to pixel N then pixel N to pixel 1)

until we obtain the first skin’s pixel on both sides. These

two first pixels will be used to delimit the major axis of

the ellipse which will be compiled to encompass the face

or at least the main features of the face. This will be done

according to equation 4.

Where :

y the ordinates of the point

),( yx

),(cc

yx the focus coordinates

a and b respectively the minor and major axis

Having the major axis b , we can directly determine the

set of ordinates )(

iy

and the coordinates of the focus

point ),(

ccyx

. The relation between the minor and major axis, given in the second term of equation 4, is a

supposition made according to the characteristics of the

human faces [17]. So we use equation 4 to compile

abscises set )(

ix

and therefore the ellipse. On figure 7, we give results of applying this algorithm to the set of

images given on figure 6.

White portion line indicates the major axis. The red curve

is the ellipse or potion of ellipse compiled according to

Eq.4 with some variations on the value of a to obtain real

values for )(

ix

.

The preliminary delimitation is not very effective. In

several cases, the ellipse is very large compared to the

face and therefore it includes areas that are not part of the

face and sometimes important parts are lost due to too

tight delimitation.

Figure 7. Elliptical delimitation on examples of figure 6.

(ii) In this second step, we will try to adjust the resulting

contour, previously obtained, to properly bond to the

contour of the face. This will be done by applying a

neural network unit on each point of the elliptical contour

obtained previously to determine whether or not this pixel

is a skin one. In the case of a non-skin pixel we push this

contour point to go towards the skin pixels according to

the side on which the point is positioned. If not, the

contour point is pushed to go towards the limits of the

skin zone (figure 8).

Figure 8. Elliptical face zone delimitation on examples of figure 6.

III. EXPERIMENTAL RESULTS

In order to check the validity of the proposed method

studied here, experimental studies were carried out on the

XM2VTS images database. This extended database

contains 4 recordings of 295 subjects taken over a period

of 4 months with rotating head shot in vertical and horizontal directions. Images are coloured and in ppm

format. To determine principal symmetrical axes and

vertical face zone we brought some transformations to

ba *3

2≈

2

2

22

)(* cc yya

bbxx −−±=

(4)


42 © 2014 UTA

original images like change to GIF format (more

compressed) and the use of luminance information only

(grey scale images). Instead, coloured images were used

for other steps.

A. General results

In this section we give, discuss and compare the results

obtained through experiments made according to the

strategy described above.

Figure 9. Gdr measure for Elliptical delimitation

procedure applied on 259 images of XM2VTS.

Curve given in figure 9 is recorded for Gdr (Good

Detection Rate) [3] measure applied to the 259 first

images of XM2VTS database. The average rate is also

given. This curve shows that the presented strategy is

valid for almost all images processed. Indeed, only 16

images have rates below 80% which represents about 6%

of fail responses.

For the Qdr (Quality Detection Rate) [3] curve shown on

figure 10, the same remarks can be done. We have only

21 images which have a Qdr below than 70%. This,

represents a failure rate less than 9%. This success is

based on correct determination of the major axis of the

ellipse.

Figure 10. Qdr measure for Elliptical delimitation procedure applied on 259 images of XM2VTS.

Figure11. Processing time (Pt) for Elliptical procedure applied on 259 of XM2VTS images.

On figure 11, we represent the time taken to process each

image. This time is only given show that the processing

time is different for each image. This is due, firstly to the

size of the face in the image, and secondly to the quality

of the initial detection of the face area.

Figure 12 is given to appreciate more clearly the quality

of the obtained results by applying the proposed technique

on a small sample of images (48 first images of the

XM2VTS database).

Figure 12. Gdr (first plot), Qdr (second plot) and Pt (third plot) on the 48 first images of XM2VTS database.

B. Out of XM2VTS

In this section we give some examples of applying the

proposed method on a set of non-standard images.

Figure13 gives four images from a set of non-standard

images which we use to test developed algorithms. The

most important characteristics which differ from the

XM2VTS standard database images are the non-neutral

background, the large variability in face’s dimensions and

in face’s position in the image.

Figure 13. Sample of results obtained by applying each

one of the three proposed algorithms to a set of non-

standard images.

M. Saaidia et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, NÁ 01 (2014) 39-44

43 É 2014 UTA

IV. CONCLUSION

A fast and efficient face localization method using geometric and skin characteristics was presented in this paper. Firstly we search a vertical symmetry on the image to determine the principal symmetrical axis of the face than a neural network trained to recognize skin pixels is used to delimit the face zone. To do so, an elliptical delimitation strategy was adopted. Recorded results demonstrate that elliptical face’s model delimitation gives good results in terms of quantitative measurement adopted. Compiled processing time was different for each image according to the size of the face in the image and to the quality of the initial face delimitation step. Simplicity of the presented method increasingly promising to become very efficient when implemented on a hard support which will permit parallel processing.

REFERENCES

[1] E. Hjelmas and B. K. Low, “Face detection: A survey” Computer Vision and Image Understanding, 83 (3), pp. 236-274, 2001.

[2] M-H. Yang, D. J.Kriegman and N. Ahuja, “Detecting faces in images: a survey”, IEEE transaction on pattern analysis and machine intelligence, 24 (1), 2002.

[3] M. Saaidia, A. Chaari, S. Lelandais, V. Vigneron a n d M . Be d d a ,“Face localization by neural networks trained with Zernike moments and Eigenfaces feature vectors. A comparison”, AVSS2007, pp. 377-382, 2007.

[4] Q. B. Sun, W. M. Huang, J. K. Wu, “Face detection based on colour and local symmetry information“, Third IEEE Int. Conf. on Digital Object Identifier, 2 (1), pp.130-135, 1998.

[5] J. S. Hyun , H. K. Mi, S. C. Yun, C. K. Nam, “Face detection using sketch operators and vertical symmetry”, Int.

Conf. on Flexible Query Answering Systems, 4027, pp. 541-551, Milan 2006.

[6] S. H. Yea, Y. C. Hao, F. C. Po and Y. T. Cheng, “Face Detection with High Precision Based on Radial-Symmetry Transform and Eye-Pair Checking“, AVSS'06, p. 62, Sydney 2006.

[7] A. Hamzah, A. Fauzan and M. S. Noraisyah, “Face localization for facial features extraction using a symmetrical filter and linear Hough transform“, Artificial Life and Robotics, Springer Japan, 12, Num.1-2, 2008.

[8] A. Hadid, M. Pietikäinen and B. Martinkauppi, “Color- based face detection using skin locus model and hierarchical filtering“, 16th Int. Conf. on Pattern Recognition, pp.196-200, Quebec City, Canada 2002.

[9] W. Qiong, Y . Jingyu and Y. Wankou, “Face Detection using Rectangle Features and SVM“, Int. Journal of Intelligent Systems and Technologies, 2006.

[10] V. Vezhnevets, V . Sazonov and A. Andreeva, “A Survey on Pixel-Based Skin Color Detection Techniques”, Proc. Graphicon, 2003.

[11] J. -C. Terrillon, M. N. Shirazi, H . Fukamachi and Akamatsu, “Comparative performance of different skin chrominance models and chrominance spaces for the automatic detection of human faces in color images”, Int. Conf. on Face and Gesture Recognition, pp. 54–61, 2000.

[12] J . Kovac, P . Peer and F . Solina, “Human skin color clustering for face detection”, EUROCON 2003, Computer as a Tool. The IEEE Region 8, 2, pp. 144-148, 2003.

[13] D . Brown, I . Craw and J . Lewthwait, “A SOM based approach to skin detection with application in real time systems”, British Machine Vision Conference, 2001.

[14] N . Guerfi, J.-P. Gambotto and S. Lelandais, “Ligne de Partage des Eaux pour l'extraction de visage dans l'espace de couleurs TLS”, CORESA 2005.

[15] N. –K. Kim, H. Choi and S. –H. Chien, “Face detection using scan-line based Hough transform and MLP“, IAPR workshop on machine vision applications, pp. 314- 317,

Nara , JAPON 2002. [16] A. X. Han, R. S. Ma and Li. Yibin, “Face detection and

recognition with SURF for human-robot interaction“, IEEE Int. Conf. on Automation and Logistics, ICAL '092009, pp. 1946 – 1951, 2009.

[17] M . Paula, R. Leonardo, J. Luciano and B. Maria, “Facial dimensions, bite force and masticatory muscle thickness in preschool children with functional posterior crossbite“, Brazilian Oral Research, 22 (1), 2008.


44 © 2014 UTA

Fuzzy Regression Analysis using Quadratic and Support Vector Machines Approaches

Riadh Djabri Fayçal Megri Noureddine Guerfi Electrical Engineering Laboratory, Electrical Engineering Laboratory, Electrical Engineering Laboratory, Department of Electrical Engineering, Department of Electrical Engineering, Department of Electrical Engineering, University of Tebessa, Algeria, University of Tebessa, Algeria, University of Tebessa, Algeria [email protected] [email protected] [email protected]

Abstract—This paper compares two approaches of fuzzy regression analysis, the first one, it is a conventional technique of fuzzy regression proposed by Tanaka, and the second based upon fuzzy support vector regression.

Keywords-fuzzy arithmetic; regression analysis; support vector machines.

I. INTRODUCTION

Since Tanaka et al. [5] introduced the fuzzy regression model with symmetric fuzzy parameters, the properties of fuzzy regression have been studied extensively by many researchers.

Fuzzy regression model can be simplified to interval regression analysis which is considered as the simplest version of possibilistic regression analysis with interval coefficients. Some coefficients in interval linear regression model tend to become crisp because of the characteristic of linear programming (LP) . To alleviate the issue of LP, Tanaka and Lee [4] propose an interval regression analysis with a quadratic programming (QP) approach which gives more diverse spread coefficients than a LP one.

The support vector machine (SVM) has been widely used in pattern recognition, regression and distribution estimation for crisp data [3]. Recently, using SVM to solve the interval regression model becomes an alterative approach. Hong and Hwang introduce SVM for multivariate fuzzy regression analysis [4] and evaluate fuzzy svm regression models [1]. The rest of this paper is organized as follows. In Section II, model of the conventional approaches is described. Simple support vector regression (SVR) is detailedly explained in Section III. The forecasting procedure using fuzzy support vector regression approach in Section IV. The experimental part is showed in Section V. Finally, conclusions are drawn in Section VI.

II. CONVENTIONAL APPROACHES

In this section we review fuzzy regression analysis with linear and quadratic programming approaches proposed by Tanaka et al. [4].

A. Linear programming (LP) An interval linear regression model can be written as

0 1 1(x) . ... . .xn nY A A x A x A= + + + = (1)

where 1x (1, ,..., )tnx x= is a real input vector, A is an interval

coefficient vector, and (x)Y is the corresponding estimated interval. An interval coefficient iA is denoted in his Midpoint/Radius form as ( , )i i iA m r= where im is the midpoint and ir is the radius. Thus it can also be expressed in the bounds form as

[ ], ,i i i i iA a a m r m r− += = − +⎡ ⎤⎣ ⎦ (2)

By interval arithmetic, the regression model (1) can be expressed as

0 0 1 1 1

0 1 1 0 1 1

(x ) ( , ) ( , ). ... ( , ).( . ... . , . | |

... . | |)

( x , | x |)

j j n n jn

j n jn j

n jn

t tj j

Y m r m r x m r xm m x m x r r x

r x

m r

= + + +

= + + + +

+ +

=

(3)

where 0( ,..., )tna m m= , 0( ,..., )t

nr r r= , and

1|x | (1,| |,...,| |)tj j jnx x= . Here xt

jm , and | x |tjr represent the

midpoint and the radius of the estimated interval (x )jY , respectively [6]. In general the LP approach is a linear optimization problem with linear constraints as follows

,

1 0

n |i |mi i

M

im rj i

n

ijr x= =∑∑ (4)

Subject to:

| |

| | , 1,...,0, 1,...,

j

j

t tj j

t tj

i

j

m x r y

m x r y

x

x j Mr i n

+ ≥

=

≤

=

−

≥

(5)

By solving the optimization problem (4) under constraints (5) we obtain the mid and radius of the estimated fuzzy output.

45 © 2014 UTA

B. Quadratic programing (QP) Here we introduce a basic formulation of QP in interval regression analysis corresponding to the former LP based regression. A QP approach is an optimization problem, which involves minimizing a quadratic objective function subject to linear constraints. To formulate interval regression by QP, the followings, should be assumed

1)- The input-output data are given as 1(x , ) (1, ,..., ; ), 1,...,j j j jn jy x x y j M= =

2)- The data can be represented by by the interval linear model (1)

3)- The given output jy should be included in the estimated output (x)Y

(x ), 1,...,j jy Y j M∈ = (6) 4)- The objective function is defined by

1 1

2| |) |( | ||t tj j

M

jj

Mt

j

x r rJ x xr= =

⎛ ⎞= ⎜

⎝= ⎟

⎠∑ ∑ (7)

which is the sum of squared spreads of the estimated outputs

and 1

| || |tjj

j

M

x x=∑ is an ( 1) ( 1)n n+ × + symmetric positive

definite matrix. Thus, the basic formulation can be expressed as the following QP problem under the same linear constraints as the LP one.

,1

| || |

subject to:

min

| |

| | , 1,..., 0, 1,...,

i i

M

m rj

t tj j

t

t t tj j

tj j j

j

i

m x

r x

r y

m x r y

x r r r

x

x j Mr i n

ξ=

⎛ ⎞+⎜ ⎟

⎝ ⎠

+ ≥

≥ =

− ≤

=

∑ (8)

where ξ is a significantly small positive number. Here it is necessary to add tr rξ to the objective function (7) so that (8) becomes a QP problem.

III. SUPPORT VECTOR REGRESSION

The basic principle of support vector regression (SVR) is to map data in the input space to a high dimensional feature space by using a nonlinear mapping. Then, a linear mapping is made in the high dimensional space [3].

Suppose we are given a training data set (x , ), i 1,..., i iy n= , x n

i ∈ℜ are given as input, niy ∈ℜ are

the corresponding output. The main objective in -SVMε is to find a function (x)f that has at most ε deviation from the actually obtained targets jy for all the training data.

An ε -insensitive loss function is used in this study. | ( ) | if | ( ) | 0 otherwisey f x y f x

Lε

ε ε− − − ≥⎧= ⎨⎩

(9)

SVM regression theory is to find a nonlinear map from input space to output space and map the data to a higher

dimensional feature space through the map, then the following estimate function is used to make linear regression:

(x) . (x)f w bφ= + (10) where (x)φ denotes the high-dimensional feature space, w denotes the weight vector and b denotes the bias term. The problem of the function approximate is equivalent with the minimizing the following problem:

2 2

1

1( ) ( , ( )) n

empi

J w Risk f w C L y f xn ε

=

= + = + ∑ (11)

According to statistic theory, the regression function is determined by minimizing the following convex optimization problem;

( )1 2

2

, , , 1 21

1

2

1 2

1minimize ,2

( . ( ) ) subject to: . ( )

, 0

n

w b i ii

i i

i i

i i

w C

y w x bw x b y

ξ ξ ξ ξ

φ ε ξφ ε ξ

ξ ξ

=

+ +

− + ≤ −⎧⎪ + − ≤ +⎨⎪ ≥⎩

∑

(12) C is used to equalize the complicated item of the model and the parameters of the item of training error. and 1iξ , 2iξ are relaxation factors and ε is insensitive loss function. Using the Lagrange multiplier method, this quadratic programming problem can be formulated as the dual form shown in (13), where 1iα , 2iα are the nonnegative Lagrange multipliers.

( )( )

( ) ( )

( )

[ ]

1 2 1 21 1

1 2 1 21 1

1 21

1 2

1maximize ( ). ( )2

0subject to:

, 0,

n n

i i j j i ji j

n n

i i i i ii i

n

i ii

i i

x x

y

C

α α α α φ φ

α α ε α α

α α

α α

= =

= =

=

− − −

+ − − +

⎧ − =⎪⎨⎪ ∈⎩

∑∑

∑ ∑

∑

(13)

Solving the above dual quadratic programming problem, we obtain the Lagrange multipliers 1iα and 2iα , which give the weight vector w as a linear combination of (x )iφ .

( )1 21

(x )n

i i ii

w α α φ=

= −∑ (14)

Knowing w, we can determine the bias term by exploiting the Karush–Kuhn–Tucker (KKT) conditions. Hence can be computed as follows:

( )( )

1

2

( . (x )) for 0,

( . (x )) for 0,i i i

i i i

b y w C

b y w C

φ ε α

φ ε α

= − − ∈

= − + ∈ (15)

Hence, the regression function is:

( )

( )

1 21

1 21

(x) (x ). (x)

(x , x)

n

i i ii

n

i i ii

f b

K b

α α φ φ

α α

=

=

= − +

= − +

∑

∑ (16)

R. Djabri Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 45-49

46 © 2014 UTA

where (x , x)iK is called the kernel function, the value of the kernel function equals the inner product of (x )iφ and (x)φ which are produced by mapping two vectors x i and x into the higher dimensional feature space. In SVM, four common kernels function types are given as follows:

( )

2 2RBF kernel (x , x) exp( x x / )

Sigmoid kernel (x , x) tanh( x x )

Polynomial kernel (x , x) 1 x x

Linear kernel (x , x) x x

i i

Ti i

dTi i

Ti i

K

K k

K

K

σ

θ

= − −

= +

= +

=

IV. FUZZY SUPPORT VECTOR REGRESSION In modeling some systems where available information is uncertain, we must deal with a fuzzy structure of the system considered. This structure is represented as a fuzzy function whose parameters are given by fuzzy sets. The fuzzy functions are defined by Zadeh’s [6], extension principle. Basically, the fuzzy function provides an effective means of capturing the approximate, inexact natural of real world. In this section, we incorporate the concept of fuzzy set theory into the SVM regression model. The parameters to be identified in the SVM regression model, such as the components of weight vector and bias term, and the desired outputs in training samples are set to be the fuzzy numbers. For computational simplicity, we assume the fuzzy parameters to be identified are symmetric triangular fuzzy numbers [1]. First, the components in weight vector and bias term used in the function regression model are fuzzy numbers. Given the fuzzy weight vector ( , )W w c= and fuzzy bias term ( , )B b d= where 1( ,..., )t

nw w w= and 1( ,..., )tnc c c= are the midpoints and

the radius vectors, respectively. The fuzzy function:

1 1(x) .x +...+ .x .xn nY W W B W B= + = + (17) is defined by the following membership function

( .x )1 , if x 0.x

( ) 1, if x=0,y=00, if x=0,y 0

Y

y w bc d

yμ

− +⎧ − ≠⎪ +⎪= ⎨⎪ ≠⎪⎩

(18)

The steps for getting the regression function are the same as the simple support vector regression. So, our optimization problem is the following:

( )1 2

2 2

1 2, , , , ,1

1

2

1 2

1 1 ,2 2

( . ) (1 )( . )(1 )

( . ) (1 )( . ) subject to:

(1 )0, 0, , 0

1: , 1:

i i

n

i iw c b di

i i

i i i

i i

i i i

j i i

minimize w K c d P

w x b H c x dy H e

w x b H c x dy H e

d ci n j M

ξ ξξ ξ

ξ

ξξ ξ

=

⎛ ⎞+ + + +⎜ ⎟⎝ ⎠+ + − +⎧

⎪ ≥ + − −⎪⎪− + + − +⎨ ≥ − + − −

≥ ≥ ≥

= =

∑

⎪

⎪⎪⎪⎪⎩

(19)

Where 2w is the term that characterizes the model

complexity, and 212

c d⎛ ⎞+⎜ ⎟⎝ ⎠

is the term that characterizes the

vagueness of the model. More vagueness in the fuzzy linear regression model means more inexactness in the regression result. K is a tradeoff parameter chosen by the user. The value of H determines the low bound for the degree of fitting of the fuzzy linear model, and the given fuzzy desired output data

( , )i i iY y e= , where P is a fixed penalty parameter chosen by the user.

To solve the optimization problem (19), we must go to the dual problem as follows:

( )( )

( ) ( )( )

( ) ( )

1 2, 1 2 1 21 1

2

1 2 1 21 1

1 2 1 21 1

1maximize .2

1 .

2

i i

n n

i i j j i ji j

n n

i i j j i ji j

n n

i i i i i ii i

k x x

Hk x x

y e

α α α α α α

α α α α

α α α α

= =

= =

= =

− − −

−− + +

+ − + +

∑∑

∑∑

∑ ∑ (20)

Subject to:

( )

( )

[ ]

1 21

1 21

1 2

0,

, 1

, 0, .

n

i ii

n

i ii

i i

KH

P

α α

α α

α α

=

=

− =

− ≤−

∈

∑

∑ (21)

By solving this optimization problem, we obtained the optimal lagrange multipliers 1iα and 2iαFinally,

( )

( )

1 21

1 21

x

1 |x |

n

i i ii

n

i i ii

w

HcK

α α

α α

=

=

= −

−= +

∑

∑

V. EXPERIMENTAL RESULTS We present a data set of type, crisp input-fuzzy output, to

give an illustrative point of view of the proposed fuzzy support vector regression machine. First, we present a numerical example that can be visualized easily.


47 © 2014 UTA

TABLE 1. CRISP INPUT – FUZZY OUTPUT

In this section, we use the data set, shown in Table 1. These data are taken from Tanaka and Lee [4]. The fuzzy support vector regression approaches are applied to this data set from an illustrative point of view. We first consider fuzzy linear regression analysis. The fuzzy linear model for this data is assumed as

(x) .xY W B= +By applying fuzzy svr approach we obtain the model

(x) (8,0.5).x (1.1525,2.931)Y = + By applying Tanaka’s QP approach we obtain the model

(x) (8.6174,0.6079).x (1.3087,2.2158)Y = +We can see, that the fuzzy coefficients of the fuzzy linear regression are similar for both techniques, but in the nonlinear case the fuzzy SVR approach using RBF kernel (see Figure4) is better than Tanaka’s LP and QP approaches. As in Figures.1-5, the solid line explains the fitted regression line for center, while the two dashed lines explain the upper and lower bounds of the fuzzy model, respectively. Figure 1, shows the simple regression model by SVM, for imprecise data, we observe that the obtained model does not cover all training data set.

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.81

2

3

4

5

6

7

8

9

10

X

Y

Figure1. Simple SVM regression linear model

And the main advantage of the fuzzy support vector regression is the efficiency when dealing with large fuzzy data set, and the implementation of fuzzy nonlinear regression is more practical because of the kernel approach.

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8-2

0

2

4

6

8

10

12

X

Y

Figure 2. Fuzzy SVM regression linear model

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8-2

0

2

4

6

8

10

12

14

X

Y

Figure 3. Fuzzy SVM regression model by polynomial Kernel

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

2

4

6

8

10

12

X

Y

Figure 4. Fuzzy SVM regression model using RBF Kernel


48 © 2014 UTA

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

2

4

6

8

10

12

X

Y

Figure 5. Fuzzy regression model by LP- Tanaka

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8-2

0

2

4

6

8

10

12

X

Y

FIGURE 6. FUZZY REGRESSION MODEL BY QP- TANAKA

VI.CONCLUSION

In this paper we have shown a little overview about a very effective method for regression analysis, it is the fuzzy support vector regression (FSVR). In the study, some subsequent work needs to be addressed in the near future. And also dealing with large data set and multi inputs systems.

REFERENCES

[1] Pei-Yi Hao and jung-Hsien Chiang, Senior Member, IEEE. Fuzzy Regressions Analysis by Support Vector Learning Approach, VOL. 16,NO. 2, APRIL 2008.

[2] Samprit Chattefuee et Alis Hadi, Regression Analysis by Example. fourth edition 2006.

[3] S. Gunn, Support Vector Machines for Classification and Regression, ISIS Tech. Report, University of Southampton, 1998.

[4] H. Tanaka, and H. Lee, Fuzzy Linear Regression Combining Central Tendency and Possibilistic . Properties. IEEE. Fuzzy Sets and Systems, 24:063-068, 1997.

[5] H. Tanaka, and H. Lee, Interval Regression Analysis by Quadratic Programming Approach, . IEEE.Transactions on Fuzzy Systems, Vol. 6, pp. 473--481, 1998.

[6] L. A. Zadeh, Fuzzy Sets as a Basis for a Theory of Possibility, Fuzzy Sets and Systems, . Vol. 1, 1, pp.3-28, 1978.


49 © 2014 UTA

Nonlinear model-based coagulant dosingcontrol at water treatment plants

Toufik Benzaraa, Messaoud Ramdani, Khaled Mendaci and Abdenabi AbidiLaboratoire LASA, Universite Badji-Mokhtar de Annaba

BP. 12, Annaba 23000, AlgérieE-mail: [email protected]

Abstract— The optimal coagulant dosage in drinking watertreatment has become an issue of particular concern in modernwater industry. This paper presents the application of fuzzysystems to model the prediction of coagulant doasage. Themodel is constructed using actual process data for a watertreatment plant located at the north-east of algeria. The studyis used to describe how the introduction of fuzzy systems hasresulted in more reliable system measurement and consequentlyimproved coagulation control.

Keywords: Water treatment, coagulation control, nonlinearmodeling

I. I NTRODUCTION

With the increased demand on water supply over the last century due to population growth, the adoption of new technology to ensure water quality at lower cost is essential. In a typical process of water treatment for drinking purposes, raw water from various sources in conveyed by pipelines to the waterworks where it is chemically treated, filtred and disinfected. The type of treatment it then undergoes depends on the source and the quality of its water. In general, the poorer the qualities of the raw water the more expensive it is to treat.A water treatment plant consists of a complex group of in-terconnected physical and chemical processes, and it is well known that a problem with one process, if not addressed, will quickly result in a much larger problem in one or more of the subsequent stages. Traditional drinking water treatment that has surface water usually include four important processes: flocculation, sedimentation, filtration, and disinfection, as shown in Fig. 1.The addition of chemicals to the water is arguably the most critical process within a surface treatment works. The key unit operation is chemical coagulation in which chemicals, typically aluminium or iron salts, are added to water for the purpose of producing flocs from colloidal particules and precipitating other contaminants.The key to efficient coagulation is the addition of just sufficient coagulatant chemical to the process. Under dosing of coagulant can lead to poor quality water which may fail consent whilst too much coagulant leads to less efficient filtration and sedimentation at great cost. The various ap-proaches used for the control of coagulant dosage fall into to groups: the feed-back auto-coagulation control based on streaming current detector and the feed-forward based on mathematical model [5], [7], [10]. Nowadays, most of water

supply plants determine coagulant dosage by Jar-Test and rules of thumb. However, jar tests take a long time to analyse samples in laboratory, which means that they are reactive, rather than proactive, as coagulant dosage is continuously changing when responding to the occurence of water quality changes.To improve drinking water quality while reducing operating costs, many projects were initiated by water companies to study the potential capacity of advanced process control and automation in WTP. The use of soft computing, specifically artificial neural networks [6], [9], [11], and fuzzy systems [8], is increasing in the drinking water treatment industry. In all the above studies, the model inputs consists of raw water parameters, whereas the model output was the optimal coagulant dosage to achieve the desired treated water quality. This paper focuses on the development of an auto-coagulation control based on the raw water parameters to calculate the required dosage.

The rest of this paper is organized as follows: Section2 gives an overview of the Water treatment and clarificationprocess. Section 3 describes the nonlinear modeling based onfuzzy approach and the optimization algorithm is outlined insection 4. Finally, some concluding remarks as well as somepossible improvements are given in section 5.

II. WATER TREATMENT AND CLARIFICATION PROCESS

The coagulant dose is a complicated process which isinvolved in many physical and chemical reactions, andthe effect factors to dosage are numerous, such as rawwater turbidity, temperature, flowrate, pH value and so on.Therefore, it is not an easy tast to design a model that canbe used in control and predict the effect of various factors.In this study, the water treatment plant (WTP) of Mexa(El-Kala, Wilaya of El-Taref) which situated in the east ofAlgeria was chosen as a suitable site on which to evaluatethe nonlinear model-base automatic coagulation control.The plant is equiped with many sensors and a SCADAsystem allowing on-line data acquisition of many processvariables such as raw water turbidity, temperature andpH, and treated water parameters. The treatment consistsmainly of preliminary disinfection, coagulation-flocculation,setlling, filtration and final disinfection. Then, the wateris stored in tanks and transported to the drinking waternetwork, as shown by Fig. 1.

50 © 2014 UTA

Tank

− pH− Temperature− Conductivity− Turbidity Coagulation−floculation

Settling Filtration

Preliminary disinfection

Raw waterparameters

Coagulation

Treated water

Final disinfection

Fig. 1. The typical drinking water processing units

The schematicdescription of the automatic control strate-gies for coagutant dosing in drinking water treatment isgiven in 2. In the feed-forward scheme sensors are used tomonitor the raw water variables and from this informationthe controller sets an appropriate dosing rate. Whereas thefeed-back control uses settled water parameters to determinethe dosing rate.

Fig. 2. The common control strategies for coagutant dosing in drinkingwater treatment [5]

III. F UZZY MODELING

Fuzzy modeling has proved to be a powerful techniquefor the develpment of intelligent systems, especially processcontrol, fault detection and isolation, time series analysisand forecasting. Assume we have a complex nonlinearmulti-input and multi-output (MIMO) relationship. Letx= [x1, . . . ,xr ]

T ∈ X ⊂ Rr be the vector of input variablesand y ∈ Y ⊂ Rm is the vector of ouput variables. Thetechnique of fuzzy rule-based systems allows us to representthe nonlinear process by a collectionK of rules obtainedafter the partitioning of the input space.

In the present work, the multi-input and multi-output(MIMO) neuro-fuzzy network is based on Takagi-Sugeno(TS) type fuzzy model with Gaussian membership functions

(GMF), weighted average defuzzifier and product inference.The overal output is defined as:

y j (x) =K

∑i=1

f ijφ i

/

K

∑i=1

φ i (1)

where, j = 1,2, . . . ,m; i = 1,2, . . . ,K

f ij =

(

θ i0 j +

r

∑l=1

θ il j xl

)

(2)

with l = 1,2, . . . ,r; j = 1,2, . . . ,m; i = 1,2, . . . ,K;

φ i = α ir

∏l=1

ail exp

−(

xl − pil

)2/

(

σ il

)2

(3)

Here, we assume thatail = 1, pi

l ∈ Xi ,σ il > 0 and f i

j ∈ Y j ,where Xi , and Y j are the input and output universes ofdiscourse respectively. Theith rule of the TS model can bewritten in the following form:

Ri : IF x1 isAi1and x2 isAi

2and. . .and xr isAir

THEN fij = θ il0+θ i

l1x1+ . . .+θ ilr xr

(4)

which defines a locally valid model on the support of thecartesian product of fuzzy sets constituting the premise parts.The variablesxl , with l = 1,2, . . . ,r are the r number ofinputs to the system, where, ˆy j , with j = 1,2, . . . ,m; arethe mnumber of outputs from the system, andAi

l , with l =1,2, . . . ,r and i = 1,2, . . . ,K, are the Gaussian membershipfunctions (GMFs) with corresponding mean and varianceparameters aspi

l , σ il respectively.f i

j is the output consequentof the ith rule. The Gaussian membership function is delib-erately chosen because it is continuously defferentiable atall points, that is an essential requirement to apply gradientbased optimization procedure.The functional form described previously can be mappedinto a feedforward network as shown in Fig 3. Because ofthe connectionnist implementation of TS type multi-inputmulti-output neuro-fuzzy (NF) network where, instead ofconnecting weights and biases as in neural network-we have

0

here the mean (pil ) and variance (σl

i) parameters of the GMFs, along with (θ i j , θl

ij) i.e, f ji from the rules consequent

as the equivalent adjustable parameters of the network. Now if the ajustable parameters are optimized for a given input-output data the neuro-fuzzy system can approximate the underlying nonlinear relationship between the input and output variables. Note that the ith rule is weighted by the factor (α i) to allow an autostructuration of the fuzzy systems by a sensitivity analysis [4].

IV. H YBRID TRAINING ALGORITHM FOR NEURO-FUZZY

NETWORK

Once the fuzzy logic system has been mapped into anequivalent multi-input multi-output feedforward neuronal ar-chitecture (3), the learning task can be achieved using anysuitable training algorithm such as the standard backpropaga-tion algorithm (BPA). However, because of slow convergencespeed of pure BPA, in the following a more efficient trainingmethod, namely the combination of gradient descent withleast squares optimization procedure will be used.

T. Benzaraa Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 50-54

51 © 2014 UTA

α1

x1

x2

xr

L1

f 2(x)

f 1(x)

Π

Π

Π

N

N

N Π

Π

Π

∑

L2 L3

ψ1

ψ2

µ1

µ2

µK

y

f K (x)

ψKαK

α2

Fig. 3. Architecture of the Fuzzy Additive model.

A. Structure identification

From the available training data that containN input-output samples, a regression matrixX and an output vectory are constructed

X = [x1, · · ·,xN]T ,y= [y1, · · ·,yN]

T (5)

once the number of rulesK is given, the antecedent fuzzysetsAi

j , and the parameters of the rule consequencesθ i fori = 1,2, . . . ,K can be identified.

Now it is important to notice that for the regressionapplications, the fuzzy clustering in the Cartesian product-space X × Y is a useful method in order to capturethe interaction between the input and output variablesby a. This is based on the fact that the different clustersrepresent operating regions, where the system behaviour isapproximated by local linear models. The data setZ to beclustered is formed by combiningX andy

Z = [X;y]T (6)

Once the training data Z and the number of clusters K are given, the Gustafson-Kessel (GK) clustering algorithm [1, 2],[3] is used to discover the potential regions of the rules.

From the the partition matrixU , whose ikth elementµ i

k → [0,1] is the membership degree of the datazk, the kthrow of Z in cluster i, it is possible to extract the fuzzy setsin the antecedent parts. One-dimentional fuzzy setsAi

j areobtained from multidimentional fuzzy clusters (given byU)by point-wise projection onto the space of the input variablex j :

µAij (xk j)

= pro j j(µ ik) (7)

whereµ ik is the level of belonging of thekth sample (vector

xk) to the ith cluster, while µAij (xk j)

is the value of themembership of thejth input variable of thekth sample (jthco-ordinate of the vectorxk) to the fuzzy setAi

j . Since all thefunctionsµAi

j (xk j)under consideration represent membership

function, the conditionµAij (xk j)

: R → [0,1] must hold true.

B. Parameter optimization procedure

The parameters obtained by the identification procedurecan be optimized or fine tuned by a variant of gradientdescent optimization techniques. This is achieved by theprocedure outlined in the appendix B which requires thecomputation of the derivatives of the objective functionto be minimized with respect to all the parameters. Theoptimization algorithm uses a self step adaptation. Given aset D = (xp,dp)N

p=1, such thatxp ∈ X ⊂ Rr , dp ∈ Y ⊂Rm; the objective is to find fuzzy logic systems ˆy j(xp) in theform of 1, such that the mean squared error (MSE) function

E =12

m

∑j=1xp∈D

(

y j −dpj

)2(8)

is minimized. Theproblem is reduced to the adjustment ofthe f i

j i.e, (θ i0 j ,θ

il j parameters fro the rules consequent, and

the mean(pil ) and varianceσ i

l of the GMFs, so that the MSEis minimized.Now it can be seen that the network output ˆy j , and henceE, depends onθ i

0 j , and θ il j only through f i

j . Similarly, thenetwork output ˆy j and hence S also depend onpi

l and σ il

only throughφ i , where ˆy j , f ij ,b and φ i are represented by

the following equations:

y j =K

∑i=1

f ijψ i (9)

f ij = θ i

0 j +θ i1 jx1+ . . .θ i

r j xr (10)

ψ i =(

φ i/b)

and b=K

∑i=1

φ i (11)

The ith rule is weighted by a factorα i verifying 0< α i < 1.This constraint can be taken into account by introducing anew parameter such that

α i =1

1+exp(−ζ i)(12)

Derivatives of E w.r.t pil , σ il and ζ i

∂E

∂ pil

=∂E∂φ i

∂φ i

∂ pil

=m

∑j=1

(

∂E∂ y j

∂ y j

∂φ i

)

∂φ i

∂ pil

(13)

∂E

∂σ il

=∂E∂φ i

∂φ i

∂σ il

=m

∑j=1

(

∂E∂ y j

∂ y j

∂φ il

)

∂φ i

∂σ il

(14)

∂E∂ζ i =

∂E∂φ i

∂φ i

∂α i

∂α i

∂ζ i =m

∑j=1

(

∂E∂ y j

∂ y j

∂α i

)

∂φ i

∂α i

∂α i

∂ζ i (15)

Finally, the results of the chain rules are written as follows:

∂E

∂ pil

= A·

2·φ i ·(

xl − pil

)

/

(σ)2

(16)

∂E

∂σ il

= A·

2·φ i ·(

xl − pil

)2/

(σ)3

(17)

∂E∂ζ i = A·

(

1−α i)φ i (18)


52 © 2014 UTA

with A=

(

m∑j=1

(y j −d j)·(

f ij − y j

)/

b

)

For the estimation of the consequent parameters, the globalLS or the WLS can be used because the outputs of the fuzzymodel are linear in the regression variables.

V. THE RESULTS OF NUMERICAL EXPERIMENTS

The task is to design a suitable controller that can copewith the wide variability of raw water quality. The schemeis a feed-forward control as shown in Fig. 2. Raw waterparameters; pH, turbidity, temperature and conductivity aresampled by a data acquisition unit (SCADA system). Thecontroller calculates the predicted coagulant dosing rate forthe underlying raw water conditions. The predicted dosingrate is then supplied as the external set point to a lowerlevel PID controller. Raw water and coagulant dosing ratedata spanning a period of 570 days is illustrated in Fig. 4.Data points 1-285 were used to develop the fuzzy model.The complete data set was used to validate the resultingcontroller model. The prediction accuray of the nonlinearfuzzy model is illustrated in Fig. 5. The model is able topredict the coagulant dosage with good accuracy over theperiod considered.

0 100 200 300 400 500 6000

10

20

30

Sample # (Days)

Tem

pera

ture

(°C

)

0 100 200 300 400 500 6007

8

9

10

Sample # (Days)

pH

0 100 200 300 400 500 6000

100

200

300

Sample # (Days)

Tur

bidi

ty (

NT

U)

0 100 200 300 400 500 600200

400

600

800

Sample # (Days)

Con

duct

ivity

(µ

S)

Fig. 4. Raw water data

VI. CONCLUSIONS

This paper has presented somme preliminary resultsconcerning the challenging task of controling the coagulantdosing rate at a water treatment plant using fuzzy modeling.It has been demonstrated that process data can be used todevelop and train a feed-forward controler in the form ofan adaptive fuzzy model to accurately predict a suitablecoagulation dosing rate.It is worth noticing that the performances of the predictionmethod for coagualant dosage in water supply treatmentprocess depend on the quality and completness of the rulebase available. Thus, an adaptation scheme is expected toimprove the accuracy of the control system. Furthermore,

0 100 200 300 400 500 6000

20

40

60

80

100

120

140

160

180

Sample # (Days)

Coa

gula

nt d

ose

(mg/

l)

Fig. 5. Actual (solid line) versus predicted dosing rate (dashed line).

in the future we will take into accound the treated waterparameters.

APPENDIX A

The GK clustering algorithm :Given Z, choose 1< K < N,m> 1 andε > 0. InitializeU (0)

Repeat for l = 1,2, . . . .Step 1)Compute Cluster Means

v(l)i =

N∑

k=1

[

µ(l−1)ik

]mzk

N∑

k=1

[

µ(l−1)ik

]m, i = 1,2, . . . ,K

Step 2) Compute Covariance Matrices

Fi =

N∑

k=1

[

µ(l−1)ik

]m[

zk−v(l)i

][

zk−v(l)i

]T

N∑

k=1

[

µ(l−1)ik

]m, i = 1,2,K

Step 3) Compute Distances

D2ik =

[

zk−v(l)i

]T [

det(Fi)1/r)F−1

i

][

zk−v(l)i

]

i = 1,2, . . . ,K, k= 1,2, . . . ,N

Step 4)Update Partition MatrixIf Dik > 0 for 1≤ i ≤ K,1≤ k≤ N

µ(l)ik =

1K∑j=1

(

Dik/D jk)2/(m−1)

otherwiseµ(l)ik = 0 if Dik = 0, andµ(l)

ik ∈ [0,1]

withK∑

i=1µ(l)

ik = 1

until∥

∥

∥U (l)−U (l−1)

∥

∥

∥< ε


53 © 2014 UTA

APPENDIX B

Let us assume thatthe task is to find an optimal vector ofparametersw which minimizes some objective functionJ(w).In the case of the generalized fuzzy additive system underconsideration, all the parameters defining the membershipfunctions and the weights of the rules are lumped inw.The optimization algorithm is a variant of gradient descentin which each parameterw j has its own step sizeη j , andthe step sizes are adapted during the optimization process,depending on the learning performance and more specificallyon the evolution of the objective function and on the signof the derivatives at successive iterations. Lett be theiteration counter. In the case where the objective function hasdecreased between iterationt −1 andt. Then, the followingrule is applied to update each step sizeη j :

η j(t) =

βη j(t −1), i f ∂J∂w j

(t −1)· ∂J∂w j

(t)> 0

γη j(t −1), otherwise

whereβ > 1 andγ < 1 are two coefficients. Hence, the stepsize is increased if the derivatives have kept the same signduring two iterations, and it is increased if the sign of thederivative has changed, because a jump over a minimum hasoccured. The parameters are then updated by

w j(t +1) =w j(t)−η j(t)∂J

∂w j(t)

If now the objective function has increased between iterationst −1 andt, all step sizes are decreased simultaneously

η j(t) = δη j(t −1) ∀ j

w j(t +1) =w j(t −1)−η j(t)∂J

∂w j(t)

We have used the following numerical values of thecoefficients:

β = 1.2γ = 0.8δ = 0.5

REFERENCES

[1] T. Takagi,and M. Sugeno, Fuzzy identification of systems and itsapplications to modeling and control,IEEE Trans. Syst., Man, Cybern.,vol. 15, pp. 116-132, 1985.

[2] D. E. Gustafson and W. C. Kessel, Fuzzy clustering with a fuzzycovariance matrix, in Proc. IEEE CDC, San Diago, CA, pp. 761-766, 1979.

[3] M. Setnes, R. Babuska, and H. B. Verbruggen, Rule-Based Modeling:Precision and Transparency,IEEE Trans. Fuzzy, Syst., Man., Cybnern.C, vol. 28, No. 1, pp. 165-169, 1998.

[4] A. Fiordaliso, Autostructuration of fuzzy systems by rules sensitivity analysis. Fuzzy Sets and Systems, 118, pp 281-296, 2001.

[5] J. Evans, C. Enoch, M. Johnson and P. Williams, Intelligent basedauto-coagulation control applied to a water treatment works, UKACCinternational Conference on CONTROL’98, 1-4 September 1998.

system for optimization of coagulant dosing in water treatment plant, Proceeding of international joint conference on neural networks (IJCNN’99), Washington, 1999.

[7] Chris Cox, Ian Fletcher, Adam Adgar, ANN-based sensing and controldevelopments in the water industry: a decade of innovations, Interna-tional Symposium on Intelligent Control, September 5-7, México City, México, pp. 298-302, 2001.

[8] C. L. Chen, P.-L. Hou, Fuzzy model identification and control systemdesign for coagulation chemical dosing of potable water, Water Science & Technology : Water Supply, Vol. 6, No 3, IWA Publishing pp. 97-104, 2006.

[9] H.Bae, S. Kim, Y. J. Kim, Decision algorithm based on data mining forcoagulant type and dosage in water treatment systems, Water Scienceand Technology 53(45), pp. 321329, 2006.

[10] Guan-De Wu,Shang-LienLo, Predicting real-time coagulant dosage inwater treatment by artificial n eural n etworks a nd a daptive network-based fuzzy inference system, Engineering Applications of Artificial Intelligence 21, pp. 11891195, 2008.

[11] B. Lamrini, El-K. Lakhal, M. V. Le Lann, L. Wehenkel, Data validationand missing data reconstruction using self-organizing map for watertreatment, Neural Comput & Applic, DOI 10.1007/s00521-011-0526-5,Springer 2011.

[6] N. Valentin, T. Denoeux, F. Fotoohi, A hybrid neural network based


54 © 2014 UTA

PSO Optimization with Autoregressive Modeling and Support Vector Machines for Bearing Fault

Diagnosis

Abstract— As an effective tool in pattern recognition and ma-chine learning, support vector machine (SVM) has been adoptedabroad. In developing a successful SVM classifier, extractingfeature is very important. This paper proposes the applicationof Autoregressive Modeling to SVM for feature extraction. Toimprove the classification accuracy for bearing fault prediction,particle swarm optimization (PSO) is employed to simultaneouslyoptimize the SVM kernel function parameter and the penaltyparameter. The results have shown feasibility and effectivenessof the proposed approach.

I. INTRODUCTION

Bearings are frequently applied components in the vast majority of rotating machines.Their running quality influences the working performance of equipment. Statistically, 30% of rotational mechanical equipments malfunction is caused by the faults in bearings [1]. Therefore, many important researches had been done in the advanced field of bearing fault diagnosis [1],[2],[3],[4]. Using the vibration signal so frolling be a rings and components to monitor and diagnose their workings tate, is the common used method in the study of bearing fault diagnosis [1],[3]. Support Vector Machine (SVM) is a new machine learning method which was introduced by Vapnikon the foundation of statistical learning theory(SLT). However, since the middle of 1990s, the algorithms used for SVM started emerging with greater a vailability of computing power [5],[6]. The main difference between the known domain of artificial neural network (ANN) and SVM is in the principle of risk minimization (RM) [2],[3]. In case of SVM, structural risk minimization (SRM) principle is used to minimize an upper bound on the expected risk whereas in ANN, traditional empirical risk minimization (ERM) is used tominimize the error on the training data. The difference in RM leads to better generalization performance for SVM than ANN. According to the literature, SVM has been successfully applied to many applications, such as pattern identification, regression analysis, function approximating, etc [7],[8],[9],[10]. The results give the evidence that the technique is not only quite satisfying from a theoretical point of view, but also can lead to high performance in practical applications.

Finding out good features is an important phase in distin-guishing the different mechanical failure; As an interesting example, Wavelet Packet analysis has been utilized for impulse mechanical failure classification [1].Parameter optimization is the key to perform SVM. At present, the widely used methods of parameter optimization for SVM are network search method, K-order cross-validation method, Leave-one-out method, etc. These algorithms have the disad-vantage of huge amount of computation, and the calculated parameters are not always the best. In recent years, a series of intelligent bionic algorithms are proposed based on the bio-logical behavior study in the natural, such as genetic algorithm (GA) and particle swarm optimization, (PSO) [11],[12],[13]. PSO was proposed by Kennedy and Eberhart [14],[15]. And it is inspired by the social behavior of bird flocking, fish schooling and swarm theory, etc. The theoretical framework of PSO is very simple, and PSO possesses the properties of easy implementation and fast convergence [13],[16]. In this paper, the higher dimension time series data is reduced and the state eigenvector is extracted by autoregressive modeling, then the eigenvector are being used as an input of a multi fault classifier which is composed of SVM. Because particle swarm optimization is powerful, easy to implement, and computationally efficient [15], this study introduces PSO as an optimization technique to simultaneously optimize the SVM parameters.

II. FEATURE EXTRACTION

The fault diagnosis is essentially a problem of patternrecognition, of which, an important step is feature extraction.In this study, two types of feature extraction are applied:Wavelet packet transform and Autoregressive modeling.

A. Wavelet packet algorithm

The step of feature extraction based on three layer waveletpacket is given as follows :Firstly,The vibration signal x(t) was decomposed by a motherwavelet, the signal features in eight frequency bands fromlow to high were extracted in the third layer.

Salah ChenikherLaboratory of Electrical Engineering, LABGET

University of Tebessa, Algeria

Tawfik ThelaidjiaLaboratory of Electrical Engineering, LABGET

University of Tebessa, Algeriaphone: +213 37 49 95 43 phone: +213 7 95 66 67 13

[email protected] [email protected]

55 © 2014 UTA

Secondly,The signal in each frequency band is extractedand the wavelet packet decomposition coefficient wasreconstructed. D30 presents the reconstructed signal of d30 ,D31 presents the reconstructed signal of d31, and so on. Thecomposed signal is defined as :

D = D30+D31+D32+D33+D34+D35+D36+D37 (1)

Finally, The signal energy of each frequency band is calculatedas :

E3j =n∑k=1

|djk|2 (2)

Normalized, Let T =∑7j=0E3j .

Eigenvector E was constructed based on each frequency bandenergy:

E =

[E30

T,E31

T,E32

T,E33

T,E34

T,E35

T,E36

T,E37

T

]. (3)

B. Least-Squares method for AR parameter estimation

In this section, we derive a method of AR estimator, witchbased on a least-squares (LS) minimization criterion using thetime-domain relation A(z)y(t) = e(t) [17],[18]. Let x(n) bean AR process of order p. Then x(n) satisfies:

e(n) = x(n) +

p∑k=1

αkx(n− k) = x(n) + x(n) (4)

We interpret x(n) as a linear prediction of x(n). from the nprevious samples x(n−1), ....., x(n−p), and we interpret e(n)as the corresponding prediction error.The vector α = [α1, ..., αp]

′ that minimizes the predictionerror variance ρ = E

|e(n)|2

is the AR coefficient vector,we

have:

ρ = E|e(n)|2

= E

|x(n)− x(n)|2

= rxx(0) + αHr + rHα+ αHRα (5)

where α, R, r are defined by:

α = (α1, α2, ..., αp)T , (6)

r = (rxx(1), rxx(2), .., rxx(p))T , (7)

R =

rxx(0) rxx(−1) . . rxx(−p+ 1)rxx(1) rxx(0) . . rxx(−p+ 2).

......

......

...rxx(p− 1) rxx(p− 2) . . rxx(0)

(8)

The vector α that minimizes (5) is given by:

α = −R−1r (9)

with corresponding minimum prediction error

ρ = rxx(0)− rHR−1r (10)

The least-squares AR estimation method is based on a finite-sample approximate solution of the above minimization prob-lem. Given a finite set of measurementsx(n)Nn=1 we ap-proximate the minimization of ρ = E

|e(n)|2

by the finite-

sample cost function

f(α) =

n1∑n=n0

|e(n)|2 =

n1∑n=n0

∣∣∣∣∣x(n) +p∑k=1

akx(n− k)

∣∣∣∣∣2

(11)

f(α) = ‖h+Xα‖2 (12)

Such that:

h =

x(n0)

x(n0 + 1)...

x(n1)

;

X =

x(n0 − 1) ... x(n0 − p)x(n0) ... x(n0 + 1− p)

......

...x(n1 − 1) ... x(n1 − p)

;

where we assume x(n) = 0 for n < 1 and n > N The vectorα that minimizes f(α) is given by

α = − (X ∗X)−1

(X ∗ h) (13)

where, as seen from (12) the definitions of X and h dependon the choice of (n0, n1), when n0 = p+ 1 and n1 = N thischoice is often named the covariance method.

III. SUPPORT VECTOR MACHINE

The support vector machine (SVM) is a supervised learning method that generates input-output mapping functions from a set of labeled training data. For classification, nonlinear kernel functions are often used to transform input data to a high-dimensional feature space in which the input data become more separable compared to the original input space [19]. Support vector machine (SVM) based on statical learning theory is proposed according to optimal hyperplane in the case of linear separable[1].If the hyperplane separate all samples correctly, it must satisfy the following condition[4]:

yk(〈w;x〉 − λ0) ≥ +1,∀k ∈ 1, ..., n (14)

In order to find the optimal hyperplane, we need to minimizethe following functionals[4].

ϕ(w) =1

2‖w‖2 (15)

Solution of the optimal problem is given by the saddles ofLagrange function as below.

L(w, λ0, α) =‖w‖2

2−

n∑k=1

αk [yk(〈w;x〉 − λ0)− 1] (16)

T. Thelaidjia and S. Chenikher Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 55-59

56 © 2014 UTA

where α = (α1...αn) is the Lagrange coefficient;αi ≥ 0,∀iThe original problem can be transferred to the dual problemas below.W (α) =

n∑k=1

αk −1

2

n∑k,k′=1

αkαk′ ykyk′ 〈xk;xk′ 〉 ,

(17)

subject to:n∑k=1

αkyk = 0 et αk ≥ 0.

If α∗ is the optimal solution, then

〈w∗;x〉 =n∑k=1

α∗kyk 〈xk;x〉 (18)

It means that the weight coefficients of the optimal hyperplaneis the linear combination of the training sample vector.According to the Kuhn− Tucker condition, the solution ofoptimal problem must satisfy

α∗k [yk(〈w∗;xk〉 − λ∗0)− 1] = 0. (19)

were λ∗0 is given by

λ∗0 =1

Nsv

Nsv∑s=1

(ys − xTs w∗), s = 1 : Nsv (20)

After solving the above problem, we can get the optimalclassification function as below.

D(x) = sgn

[∑k∈S

α∗kykxTi x− λ∗0

](21)

The nonseparable problem can be solved by soft-margin SVM [6],[20],[21].If we used the inner κ(xk, x) substitute for the inner of the optimal hyperplane, the original feature space is mapped to new feature space[22]. And the optimal function can be formulated as below:

W (α) =

n∑k=1

αk −1

2

n∑k,k′=1

αkαk′ ykyk′κ(xk, xk′ )

subject to:∑nk=1 αkyk = 0 et 0 ≤ αk ≤ c The corresponding decision

function is written as below

D(x) = sgn

[∑k∈S

α∗kykκ(xTi x)− λ∗0

](22)

here, κ(xk, x) is called kernel function.Usually, the kernel function can be expressed as below[6],[7],[23].Polynomial:

κ(x1, x2) = (1 + 〈x1;x2〉)q (23)

where parameter q is the degree of the polynomial.Radial basis function (RBF):

κ(x1, x2) = exp(−‖x1 − x2‖2 /2σ2) (24)

where parameter σ2 is the variance of the Gaussian functionSigmoid:

κ(x1, x2) = tanh(α0 〈x1;x2〉+ β0) (25)

where α0 and β0 are the parameters of kernel function. The classification performance of SVM are affected by three techniques, i.e., the selecting of the kernel, the choosing of the kernel parameters, and the choosing of the regularization parameter c [4].Most of cases in practical are multi-classed, such as in the rolling bearing classifying, it can be sorted into normal, outer race fault and inner race fault, etc. So, we have to design an approach to expend the application of SVM to a multi-classifying field because the SVM can deal with only two classes. The different combination principles constitute different classifying algorithm [7], [24], [25]. We employ the one-against-the-rest method to compose a multi-fault classifier. Since the SVM generalization performance heavily depends on the right setting of c and σ, these two parameters need to be set properly by the user. According to the experience from numerical experiments [26],[27], c and σ exhibit a (strong) interaction. As a consequence, they should be optimized simultaneously, rather than separately.

IV. THE PARTICLE SWARM OPTIMIZATION

The particle swarm optimization (PSO) consists of a swarm of particles flying t hrough t he s earch s pace. E ach p article is treated as a point in a D-dimensional space. The i-th particle is represented Zi = (Zi1, Zi2, ..., Zid, ..., ZiD). The best previous position of any particle is recorded and represented as Pi = (Pi1, Pi2, ..., Pid, ..., PiD). The index of the best particle among all the particles in the population is represented by the symbol G. The rate of the position change (velocity) for the i-th particle is represented as Vi = (Vi1, Vi2, ..., Vid, ..., ViD). The updated velocity and position of the i-th particle at the k-th iteration are [11]:

V kid = ω.V k−1id + c1.r1.(Pid − Zk−1id ) + c2.r2.(PGd − Zk−1id )(26)

Zkid = Zk−1id + V kid (27)

Where c1 and c2 are constants knows as the cognitive andsocial acceleration coefficients, respectively, ω is the inertiaweight, r1 and r2 are random numbers between 0 and 1.The first part of (26) represents the previous velocity, whichprovides the necessary momentum for particles to fly acrossthe search space. The second part is the ”cognition” part,which represents the private thinking of the particle itself..The third part is known as the ”social” component, whichrepresents the collaboration among the particles. In addition,the implementation of PSO also requires placing a limit onthe particle velocity, and the limit, i.e. the maximum allowedvelocity Vmax, determines the searching granularity of space.The inertia weight ω plays the role of balancing the globalsearch and local search, and it can be a positive constant oreven a positive linear or nonlinear function of time.


57 © 2014 UTA

V. EXPERIMENTAL PROCEDURES FOR TRAINING ANDTESTING

The database is composed from a set of vibration signalswhich are divided into five different classes C1 to C5, includ-ing a normal bearing and four faults of roller bearing (outerrace completely broken fault bearings, broken cage with oneloose element fault bearings, damaged cage with four looseelements fault bearings and badly warned ball-bearings).To guarantee valid results for making predictions regardingnew data, the data set was further randomly partitioned intotraining sets and independent test sets via a k-fold crossvalidation. Each of the k subsets acted as an independentholdout test set for the model trained with the remaining k−1subsets. The advantages of cross validation are that all of thetest sets were independent and the reliability of the resultscould be improved. The data set is divided into k subsets forcross validation. This study used k = 10, meaning that allof the data will be divided into ten parts, each of which willtake turns being the test data set. The other nine data partsserve as the training data set for adjusting the model predictionparameters.We split the data into ten groups using stratified 10-fold crossvalidation. Each group contains training and test sets. Thetraining set is used to build the SVM model. The test set isused to evaluate the model’s classification accuracy.

VI. RESULTS AND DISCUSSION

The classification performance of SVM are affected by two techniques, the choosing of the kernel parameters, and the choosing of the regularization parameter c [4].The proposed approaches for SVM parameter optimization (PSO), is illustrated in figure 1.The swarm size is set to 20 particles. The searching ranges for c and σ are as follows: c ∈ [0, 100], and σ ∈ [0, 100]. Preliminary experiments also let this study set the personal and social learning factors (c1, c2) = (1.3, 1.3) that achieves better classification accuracy.The inertia weight is set to the following equation:

ω(k) = ωmax −(ωmax − ωmin)

kmax.k (28)

where ωmax is the initial weight, ωmin is the final weight,kmax is the maximum number of iterations or generation, andk is the current iteration number.

The predefined maximum iteration is 10. When the maxi-mum iteration is reached, the accuracy of test set is calculatedby the predicted output of the trained SVM classifier.In order to select the optimal values of the parameters p (orderof autoregressive modeling), for bearing fault classification,a series of experiments had been carried out by varying thevalues of this parameter. The important variation range ofparameter p is given as follows:• p from 15 to 25,

The classification results in validation and in test obtained fordifferent values of p are shown in figure 2.

Optimization of c , σ

?

Start PSO-SVM

?Define:

Initial range of c , σ.Quantity of particles .c1, c2, ω, Vmax.

Maximum iterations .

?Initialize:

Position of each particle.Velocity of each particle.

?Perform SVM on each particle

in population and computethe prediction accuracy basedon 10-fold cross validation.

?

Renew Pi and G

N

Y

?

@@@

@@@

Max. Interation?

?

Optimal c , σ for SVM

6

Update:Velocity of particles Eq.(26).Position of particles Eq.(27).

Fig. 1. PSO-SVM procedure

Fig. 2. SVM classification results of bearing fault using different values of p


58 © 2014 UTA

SVM+ Kernel c σ Validation TestFunction Rate% Rate%

AR Gaussian 8.1317 0.5789 98.00 96.00ModelingWavelet Polynomial 27.1447 17.2956 91.33 62.00Packet

TABLE ICLASSIFICATION RESULT OF BEARING FAULT IN VALIDATION AND TEST

Table I gives the classification result for the bearing faultclassification problem based on :• the proposed method (SVM based on autoregressive

modeling feature extraction).• SVM based on wavelet packet feature extraction, where

Discrete wavelet packet transform was used to decom-pose the time signals into eight packets at level 3 viaDaubechies-8.

from many experiments we obtain the following results:• when the order of autoregressive modeling (p) increase, It

was noted a general trend of increasing of The Validationand test rate.

• the classification accuracy is poor either in validation ortest when we combine SVM with wavelet packet.

• As the results in Table I, the best classification result ofbearing fault in the validation set 98% and in the test set96% is obtained by using the proposed method of SVM-PSO based on autoregressive modeling feature extraction.

These results clearly show the high percentage of correctclassification reached for the validation set and test set, whichclearly shows the good generalization capacity of SVM-PSObased on autoregressive modeling for fault diagnosis of rollerbearing.

VII. CONCLUSION

For the Bearing faults diagnosis, input feature subset se-lection and the SVM parameters setting are crucial problems.This paper presents a new method AR-SVM-PSO for bearingfault classification, AR modeling is utilized for feature ex-traction. After feature extraction from bearing fault vibrationsignal, a particle swarm optimization is employed to simulta-neously optimize the SVM kernel function parameter and thepenalty parameter.

PSO optimization requires only simple mathematical opera-tors. This algorithm is simple to implement and effective, andis inexpensive in terms of memory and time required.

The proposed method can overcome the inefficiency forselecting reasonable parameters according to the experience inthe traditional fault diagnosis. Compared with other methods,AR-SVM-PSO is simpler and easier to realize. The combinedAR modeling and SVM-PSO based technique is tested forbearing faults and provides satisfactory results.

REFERENCES

[1] M. Li and P. Zhao, “The application of wavelet packet and svmin rolling bearing fault diagnosis,” IEEE International Conference onMechatronics and Automation, August 2008.

[2] P. K. Kankar, C. S. Satish and S. P. Harsha, “Fault diagnosis of ballbearings using machine learning methods,” ELSEVIER, 2011.

[3] L. Shuang and L. Meng, “Bearing fault diagnosis based on pca andsvm,” IEEE International Conference on Mechatronics and Automation,August 2007.

[4] G. Xian, “Mechanical failure classification for spherical roller bearingof hydraulic injection moulding machine using dwt-svm,” ELSEVIER,August 2010.

[5] V. N. Vapnik, The Nature of Statistical Learning Theory. Springer-Verlag, 1995.

[6] V. N. Vapnik, Statistical Learning Theory. Springer, 1998.[7] C. Cortes and V. N. Vapnik, Support-vector networks. Machine

Learning, 1995.[8] C. J. Burges, “A tutorial on support vector machines for pattern recog-

nition,” Data Mining and Knowledge Discovery, vol. 2, pp. 121–167,1998.

[9] S. R. Gunn, “Support vector machines for classification and regression,”pp. 1–28, 1998.

[10] J. Christopher and C. Burges, “A tutorial on support vector machinesfor pattern recognition,” Kluwer Academic Publishers, Bell Lab, LucentTechnologies, Boston, pp. 1–43, 1998.

[11] R. Yuan and B. Guangchen, “Determination of optimal svm parametersby using ga/pso,” Journal of computer, vol. 5, no. 8, August 2010.

[12] H. Cheng-Lung and D. Jian-Fan, “A distributed pso-svm hybrid systemwith feature selection and parameter optimization,” ELSEVIER, 2008.

[13] X. Yun-Jie and X. Shu-Dong, “A new and effective method of bearingfault diagnosis using wavelet packet transform combined with supportvector machine,” Journal of computers, vol. 6, no. 11, November 2011.

[14] J. Kennedy and R. C. Eberhart, “Particle swarm optimization,” Pro-ceedings of International Conference on Neural Networks, IEEE, pp.1942–1948, 1995.

[15] J Kennedy, R C Eberhart and Y.Shi, “Swarm intelligence,” MorganKaufmann Publishers Inc, 2001.

[16] J. LI, “A combination of pso and svm for road icing forecast,” Journalop computers, vol. 5, no. 9, p. 2010, September.

[17] P. Stoica and R. Moses, Spectral analysis of signal. Pearson Education,2005.

[18] A. Gattal, S. Chenikher, K. M. Pekpe and J. P. Cassar, “Bearing faultsdiagnosis using artificial network with dimensionality reduction by pca,”1st International Conference on Electrical Engineering CIGET09, 2009.

[19] L. Wang, Support Vector Machines: Theory and Applications. Springer-Verlag Berlin Heidelberg, 2005.

[20] P. Clarkson and P. J. Moreno, “On the use of support vector machinesfor phonetic classification,” IEEE Proceedings of the internationalconference acoustics, speech and signal processes, vol. 2, pp. 585–588,March 1999.

[21] P. Ding, Z. Chen, Y. Liu and B. Xu, “Asymmetrical support vectormachines and applications in speech processing,” IEEE Proceedingsof the international conference acoustics, speech and signal processes,vol. 1, pp. 73–76, May 2002.

[22] C. Wang and Y. Song, “Support vector machine for mechanical faultsdiagnosis,” IEEE International Conference on Measuring Technologyand Mechatronics Automation, 2010.

[23] B. Scholkopf, A. Smola, R. C. Williamson and P. L. Bartlett, “Newsupport vector algorithms,” Neural Computation, vol. 12, pp. 1207–1245, 2000.

[24] M. XiaoXiao and H. XiYue, “2ptmc classification algorithm based onsupport vector machines and its application to fault diagnosis,” Controland Decision, vol. 18(3), pp. 272–276, 2003.

[25] J. Weston and C. Watkins, “Multi-class support vector machines,”European Symposium on Artificial Neural Networks, pp. 219–224, April1999.

[26] X. F. Yuan and Y. N. Wang, “Parameter selection of svm for functionapproximation based on chaos optimization,” Journal of Systems Engi-neering and Electronics, vol. 19, pp. 191–197, 2008.

[27] B. Ustun, W. J. Melssen , M. Oudenhuijzen and L. M. C. Buydens,“Determination of optimal support vector regression parameters bygenetic algorithms and simplex optimization,” Analytica Chimica Acta,vol. 54, pp. 292–305, 2005.


59 © 2014 UTA

Evaluation of antioxidative and antibacterial potentials of Crataegus monogyna Jacq. from

Mahouna mountain (Algeria) Samah Djeddi and Hanifa Boutaleb

Laboratory of of Ecobiology of Marine and Littoral Environment, Department of Biology, Faculty of Science, University of Badji Mokhtar BP 12, Annaba 23000, Algeria

Abstract-Crataegus species (Rosaceae), known as “Howthorn” have found special medicinal use for the treatment of mild heart diseases. The present work was focused on the study of the polar extract of Crataegus monogyna Jacq. collected from Mahouna mountain, Guelma city (North East of Algeria). The antioxidant activity of the MeOH extract was investigated in vitro using the DPPH method. The results obtained showed that the tested extract of the plant can act as a high radical scavenger reaching 95.88%, this is due to the presence of a high percentage of polyphenol and reflected by the color change (from purple to yellow). The extract was also tested in vitro for its antimicrobial potential against nine strains of pathogenic bacteria, using the disk diffusion method. The extract tested responded positively to almost all microbial strains tested even those resistant to some antibiotics.

Key Words: antibacterial activity, antioxidant activity, Crataegus monogyna, inhibition zone, phytochemistry.

I. INTRODUCTION

Crataegus species (Hawthorn) from the Rosaceae family, is found in northern temperate regions such as East Asia, Europe, and Eastern North America. The two most common species used are Crataegus laevigata (syn Crataegus oxyacantha) and Crataegus monogyna. Hawthorn was first mentioned as a drug in the Tang-Ben-Cao (659A.D.), which is the world’s earliest officially published pharmacopoeia [1].

Crataegus have found special medicinal use for the treatment of mild heart diseases. The extracts obtained from several parts of the plant including fruits are rich in proanthocyanidins and flavonoids [2] and exert their beneficial effects on cardiovascular function which are based on the results obtained from in vitro and in vivo laboratory investigations. Many studies showed that some extracts increased the force of myocardial contraction [3-5], enhanced coronary flow [6], improved oxygen utilization by cardiomyocytes [4] and reduced the ocurrence of reperfusion-induced cardiac arrhythmias [4, 7]. Other studies demonstrated that extracts could protect the myocardium from injury in animal models of

coronary ischemia and reperfusion [8, 9]. The plant is also used as therapeutic agent for cancer, diabetes and sexual weakness in Arab traditional medicine, and is considered to be generally safe and well-tolerated [10], while its different parts are used in Turkish traditional medicine for various diseases such as cough, flu, asthma, stomach ache, rheumatic pain, nephritis and hemorrhoids [11].

In Algeria C. monogyna is distributed all over the country except highlands; it's represented with many species [12].

During recent years, herbal medicines have become increasingly popular in some parts of the world, while in other parts it has always been an essential element of healthcare. In Algeria, traditional medicine takes an important part to cure diseases. The north-east part of the country involves an extremely rich and varied flora, characterized by its originality on the numerous endemic plants and its wide use in folk medicine. These characteristics make the flora study of this region of great scientific interest in the field of traditional use [13].

Based on above findings, the present study was undertaken to assess whether polar extract of C. monogyna growing in Mahouna mountain located in Guelma city (36°21'0" N and 7°24'0" E) posses a significative antioxidant and antibacterial activities.

II. MATERIALS AND METHODSII.1. Plant material

C. monogyna was collected in March 2012 from natural population in Mahouna mountain. The plant was collected and authenticated by Mrs. Amina Beldjazia (Department of Biology and vegetable Ecology, Faculty of Sciences, Ferhat Abbas University Setif, Algeria). A voucher specimen has been kept in the Biology Department, University of Badji Mokhtar Annaba.

60 © 2014 UTA

II.2. Plant extraction

The freshly cut plant was air dried at room temperature and powdered. 250g of the plant were macerated in 500 ml of MeOH for 24 hours, the extractions were repeated three times. After filtration the extract obtained was concentrated under reduced pressure. II.3. Radical scavenging activity

From the crude extract obtained, different concentrations were prepared in methanol: 100, 250, 500, 750 and 1000mg/l. The antioxidant activity of C. monogyna methanol extract was carried out using free radical-scavenging activity using DPPH (2,2-diphenyl-1-picrylhydrazyl).

This activity was estimated by using a modified DPPH-method (2,2-diphenyl-2'-picrylhydrazyl) [14]. According to this method, 2ml of a methanol solution of DPPH of concentration 24μg/ml were added to 100μl methanol solution of extracts of the various concentrations and let stay in the dark for 30 minutes. After this time the absorbance was measured at 516 nm in a Schimadzu 160-UV spectrophotometer. Shorter times have also been reported by some authors, such as 5 min [14] or 10 min [15], but in our experiments the time of 30 min proved to be the optimum (time needed for stable signals). The radical-scavenging activity was calculated using the following equation:

Scavenging activity (%) = [(A0 - A1)/A0] × 100 A0 is the absorbance of the control (sample without extracts). A1 is the absorbance of samples with extracts. II.4. Antibacterial activity

The agar disc diffusion method was used for the evaluation of the antibacterial activity of the extract tested [16]. The following clinical bacterial strains isolated directly from patients from the “Centre Hospitalo-Universitaire Ibn Sina,” Annaba (Algeria) were used : Staphylococcus aureus (Gram-positive) and Acinitobacter baumannii, Citrobacter sp., Enterobacter cloacae, Enterobacter intermedius, E. coli1 , E. coli2 , Klebsiella pneumonioa and Serratia marcescens (Gram-negative).

All bacteria species were cultured overnight at 37 ºC in Mueller-Hinton medium (Bio-Rad). Suspensions of the tested micro-organisms (0.1 ml of 107-108 cells/ml) were spread over the surface of Petri plates. The inocula were stored at +4ºC for further use. Filter paper discs (Whatman n° 1; 6.0 mm in diameter) were impregnated with 20μl of the sample and placed on the inoculated agar

plates. Cefazoline (30µg) was used in order to control the sensitivity of the test organisms. The plates containing the bacteria were incubated for 24 h at 37 ºC. The resulting inhibition zones diameters (IZD) have been measured in millimetres [17]. All experiments were performed in triplicate.

III. RESULTS AND DISCUSSION III.1. In vitro antioxidant effects of C. monogyna

In the present study, the methanolic extract of the

indigenous C. monogyna from Algeria was estimated for its radical scavenging potential. From the results mentioned on table 1, it was noticed that the polar extract of the studied plant showed a strong radical scavenging effect reaching (95,88 %) at 1000mg/l, this concentration was followed by 750mg/l and 500mg/l where high antioxydant activity was noticed (94,86 % and 94,07%, respectively). The lowest scavenging activity was noticed at concentration 100mg/l (84,23%). However, the scavenging effect of the well known synthetic antioxydant BHT (butylated hydroxyltoluene) was only 76,13% at 1000 mg/l. Table 1: DPPH radical scavenging activity (%) from C. monogyna methanolic extract and BHT

Concentrations BHT* MeOH extract*

100mg/l 16,70 ± 5,28 84,23 ± 4,669 250mg/l 32,46 ± 7,69 90,22 ± 1,910 500mg/l 62,83 ± 5,51 94,07 ± 0,384 750mg/l 64,73 ± 0,68 94,86 ± 0,600

1000mg/l 76,13 ± 3,68 95,88 ± 0,098 *Values represent mean ± standard deviation of three replicates

A previous study, where 52 indigenous Crataegus

species of Turkey were compared in terms of their antioxidant capacities, as a general result, C. monogyna samples have exhibited markedly high antioxidant activity in comparison with all samples tested [11].

Bahorun and co-workers [18] have shown that in the leaf extracts of C. monogyna, flavonoids account for most of the antioxidant activity observed, whereas proanthocyanidins and catechins do the same activity in flowers. Moreover, the species collected from Bolu district have shown significantly high activity, this district is surrounded by several forests and lakes which provide oxygen-rich and clean air to this district. Additionally, Bolu is known for its cold weather conditions-enriched with drought during winter months, which is also a demonstrated factor in raising the antioxidant capacities values of Crataegus leaves [19]. Other results showed that the ethanolic extract of C. monogyna had stronger antioxydant activity in comparaison with the aqueous on [20].

S. Djeddi and H. Boutaleb Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 60-63

61 © 2014 UTA

III.2. In vitro antimicrobial effects of C. monogyna

The antibacterial activity test was carried out on methanol crude extract and was evaluated using disk diffusion method by measuring inhibition zone diameters [16]. Cefazoline was included as positive control (Table1 and Figure 1). The methanol extract of C. monogyna showed a strong antimicrobial activity with the highest dilution (1000mg/l) where it was very active against A.

baumannii (20.00 mm) as well as K. pneumonioa, E. intermedius and E. cloacae (19.66, 18.66 and 18.00 mm, respectively).

While it was noticed a high activity even with the lowest dilution (100mg/l) against A. baumannii (17.00mm) and a moderate effect against E. coli1 (10.66mm) and E. coli2 (10.33mm). It should be mentioned that there are no background antibacterial study on C. monogyna.

Table 2: Antimicrobial activity of C. monogyna expressed as the diameter of the inhibition zone in mm.

*Values represent mean ± standard deviation of three replicates

The sensitivity to the different strains was classified by

the diameter of the inhibition zone as follows [21]: -: diameter less than 8 mm; not sensitive +: sensitive; diameter 9–14 mm ++: very sensitive; diameter 15–19mm +++: extremely sensitive for diameter larger than

20mm.

Fig. 1: Sensitivity of Staphylococus aureus Against C.

monogyna methanol extract dilution

IV. CONCLUSION

The potential isolation and use of extracts from plants is still very productive playground for the development of new drugs to improve health care in some medical fields. It is essential to emphasize that extensive in vitro and in vivo tests must be conducted to assure the selection of active and nontoxic antimicrobial natural compounds. It can be concluded that therapeutic claims on C. monogyna

used as traditional plant have been supported by the results which show positive activity against free radicals. Further isolation of bioactive constituents in the extract would help to ascertain its potency and safety to provide leads and candidates of antioxidants for dietary cosmetic and pharmaceutical uses.

In the other hand the problem of microbial resistance is growing and the out look for the use of antimicrobial drugs in the future is still uncertain. Therefore, actions must be taken to reduce this problem by developing research on natural extracts.

Acknowledgment Authors warmly thanks Mrs. Amina Beldjazia (Ferhat Abbas University Setif, Algeria) for the identification of the plant, they also want to thank the financial support provided by the Algerian MESRS (Ministère de l’Enseignement Supérieur et de la Recherche Scientifique) and DGRSDT (Direction générale de la recherche scientifique et développement technologique), project PNR n°237/ ANDRS/2011.

REFERENCES [1] M. Yao, H. E. Ritchie and P. D. Brown-Woodman,

“A reproductive screening test of hawthorn”, J. Ethnopharmacol, 118, 127–132, 2008.

[2] T. Bahorun, B. Gressier, F. Trotin, “Oxygen species scavenging activity of phenolic extracts from hawthorn fresh plant organs and pharmaceutical preparations” Arzneimittel-Forschung, 46, 11, 1086–1089,1996.

Microorganisms IZD (mm)*

100mg/l 250mg/l 500mg/l 750mg/l 1000mg/l Cefazoline

S. aureus 14.33±1,74 16.33±1,32 16.66±0,22 17.33±4,15 18.66±2,26 00 A. baumannii 17,00±0,10 17,00±2,45 18,00±0,51 18,33±1,52 20,00±0,45 00 Citrobacter sp. 12,66±1,12 16,33±2,68 16,33±0,24 17,00±2,74 17,33±1,21 15.10±3,04 E. cloacae 16.00±1.45 16.33±0.98 17.00±4.21 18.00±0,15 18.00±2,24 13.21±0,19 E. intermedius 12.00±0.25 14.66±5.65 16.33±2.84 17.33±3,48 18.66±1,47 00 E. coli1 10.66±2.45 11.66±1.92 12.33±0,35 16.66±0,45 17.00±0,56 12,10±2,23 E. coli2 10.33±1.85 10.33±0.56 11.33±1,15 13.00±2,01 16.66±1,66 17,23±0,11 K. pneumonioa 12.00±5.02 16.00±1.61 17.33±1,32 18.00±2,31 19.66±1,02 nt S. marcescens 11.33±1,11 11.66±1,42 12.66±0,45 13.66±3,89 18.33±1,57 15.02±2,54

250mg/l

100mg/l 750mg/l

1000mg/l 500mg/l


62 © 2014 UTA

[3] S. Popping, H. Rose, I. Ionescu, Y. Fischer, H. Kammermeier, “Effect of a hawthorn extract on contraction and energy turnover of isolated rat cardiomyocytes”, Arzneimittel for schung, 45, 1157–1161, 1995.

[4] A. Muller, W. Linke, Y. Zhao and W. Klaus, “Crataegus extrac prolongs action potential duration in guinea-pig papillary muscle. Phytomedicine”, Inter. J. Phytotherapy & Phytopharmacology, 3, 257–261, 1996.

[5] R. H. Schwinger, M. Pietsch, K. Frank, K. Brixius, “Crataegus special extract WS 1442 increases force of contraction in human myocardium cAMP-independently” J. Cardiovascular Pharmacology 35, 700–707, 2000.

[6] M. Schussler, J. Holzl and U. Fricke, “Myocardial effects of flavonoids from Crataegus species” Arzneimittel for schung Drug Research, 45, 842–845, 1995.

[7] A. Garjani, H. Nazemiyeh, N. Maleki and H. Valizadeh, “Effects of extracts from flowering tops of Crataegus meyeri A. Pojark, on ischaemic arrhythmias in anaesthetized rats”, Phytotherapy Research, 14, 428–431, 2000.

[8] T. Krzeminski and S. S. Chatterjee, “Ischemia and early reperfusion induced arrhythmias: beneficial effects of an extract of Crataegus oxyacantha” L. Pharmaceutical and Pharmacological Letters, 3, 45–48, 1993.

[9] M. Veveris, E. Koch and S. S. Chatterjee, “Crataegus special extract WS(R) 1442 improves cardiac function and reduces infarct size in a rat model of prolonged coronary ischemia and reperfusion”, Life Sciences, 74, 1945–1955, 2004.

[10] P. Ljubuncic, H. Azaizeh, I. Portnaya, “Antioxidant activity and cytotoxicity of eight plants used in traditional Arab medicine”, J. Ethnopharmacol, 99, 1, 43–47, 2005.

[11] M. Özyürek, M. Bener, K. Güçlü, A. Dönmez, S. Süzgeç-Selçuk, S. Pırıldar, A. H. Meriçli and R. Apak, “Evaluation of Antioxidant Activity of Crataegus Species Collected from Different Regions of Turkey”, Rec. Nat. Prod., 6, 3, 263-277, 2012.

[12] P. Quézel and S. Santa, “Nouvelle flore de l’Algérie et des régions désertiques méridionales”, tome 1, Paris, CNRS, 565, 1962.

[13] A. Zellagui, N. Gherraf, M. Kaabache, S. Rhouati, “Phytochemical and biological survey from two endemic Species: genista microcephala coss Et dur. and filago Pomelli Batt. et Trab” Plant Sciences Feed, 1, 11, 190 – 193, 2011.

[14] J. Lebeau, C. Furman, J.L. Bernier, P. Duriez, E. Teissier, N. Cotelle, “Antioxidant properties of di-tert-butylhydroxylated flavonoids”, Free Radical Biol. Med., 29, 9, 900-912, 2000.

[15] K. Schwarz, G. Bertelsen, L. R. Nissen, P. T. Gardner, M.I. Heinonen, A. Hopia, “Investigation of plant extracts for the protection of processed foods against lipid oxidation. Comparison of antioxidant assays based on radical scavenging, lipid oxidation and analysis of the principal antioxidant compounds”, Eur. Food Res. Techn., 212, 319–328, 2001.

[16] NCCLS (National Committee for Clinical Laboratory Standards), “Performance standards for antimicrobial disk susceptibility test”, 6th Ed. Approved Standard M2-A6, Wayne, PA, 1997.

[17] L. Jirovetz, G. Buchbauer, A.S. Stoyanova, E.V.Georgiev, S. T. Damianova, “Composition, quality control and antimicrobial activity of essential oil of long time stored dill (Anethum graveolens L.) seed from Bulgaria” J. Agricultural Food Chemistry, 51, 3854-3857, 2003.

[18] T. Bahorun, F. Trotin, J. Pommery, J. Vasseur and M. Pinkas, “Antioxidant activities of Crataegus monogyna extracts”, Planta Medica, 60, 323-328, 1994.

[19] A. Kirakosyan, E. Seymour, P. B. Kaufman, S. Warber, S. Bolling and S. C. Chang, “Antioxidant capacity of polyphenolic extracts from leaves of Crataegus laevigata and Crataegus monogyna (Hawthorn) subjected to drought and cold stress”, J. Agric. Food Chem., 51, 3973-3976, 2003.

[20] J. Bernatonienė, R. Masteikova1, D. Majienė, A. Savickas, E. Kėvelaitis, R. Bernatonienė, K. Dvoráčková, G. Civinskienė, R. Lekas, K. Vitkevičius and R. Pečiūra, “Free radical-scavenging activities of Crataegus monogyna extracts Medicina (Kaunas)”, 44, 9, 706-712, 2008.

[21] A. G. Ponce, R. Fritz, C. E. Del Valle, S. I. Roura, “Antimicrobial activity of essential oils on the native microflora of organic Swiss chard”, Lebensmittel-Wissenschaft und-Technology, 36, 679-684, 2003.


63 © 2014 UTA

Screening of antagonistic activity of indigenous bacteria against two fusarium speciesS. Mezaache-Aichour1 , N. Sayah1, N. Haichour1, A. Guechi1, M. M. Zerroug1

1Laboratoire de Microbiologie Appliquée, Faculté des Sciences de la Nature et de la Vie, Université Ferhat ABBAS Sétif “1”, 19000, Sétif, ALGERIE.

E-mail: [email protected]; [email protected]

Abstract-The genus Fusarium is one of the mostimportant fungi that include many pathogenic species, causing a wide range of plant diseases. The genus Fusarium is an ubiquitous soil saprophyte and has been isolated from debris and roots, stems and seeds of a wide variety of plants. Since, resistant plant varieties are not available for several soil-borne pathogens and chemical control is often insufficiently effective in soil. The enhancement of disease suppressive properties of soils will limit disease development, thus, being of great importance for sustainable agriculture as well as organic farming systems. The aim of this research is to test the inhibitory effect of some indigenous bacterial strains isolated from two potato fields, in Sétif (east of Algeria) against two fusaria strains, using the confrontation test. The results showed that among 50 bacterial strains only 11 showed an important antifungal activity against the tested phytopathogenic fungi, Fusarium oxysporum f. sp. albedinis and Fusarium solani var. coeruleum. This activity varies within the fields. The percent inhibition rate was from 0 to 92.30%, especially against Fusarium oxysporum f. sp. bi with a rate of 92%. The strain 8 from the first field inhibited Fusarium oxysporum f. sp. albedinis with 53,84 % and had no effect against Fusarium solani var. coeruleum, while the strain17 from the second field inhibited Fusarium solani var. coeruleum with 85% and had a very low effect against Fusarium oxysporum f. sp. albedinis with 1,25%.

Keywords - Fusarium, indigenous bacteria, antagonism,biocontrol.

I. INTRODUCTION

D. Fusarium are the most common and widespread among fungi, the genus is an ubiquitous soil saprophyte which can be isolated from debris, roots, stems and seeds of a wide variety of plants (Leslie and Summerell, 2006). Fusarium includes many pathogenic species, causing a wide range of plant diseases, of many crops such as wheat, corn, rice, bananas, potatoes, cotton and linen. They are also responsible for the deterioration of stored products and can cause disease for humans, animals and insects. The large distribution of Fusarium species may be attributed to the ability of these fungi to grow on various substrates and their efficient mechanisms for

dispersal [1]. Natural suppressive soils are probably the best examples in which indigenous microflora protect effectively plants against soil-borne pathogens such as Fusarium species. Initially, the soil suppressiveness became apparent because the incidence or severity of the disease is lower than what was expected in a receptive environment, or by comparison with infested soils. Suppressive soils have been described for many telluric pathogens such as Gaeumannomyces graminis var. tritici, Fusarium oxysporum. Cook et al. [2] postulated that many plant species have developed defense strategy against soil-borne pathogens involving selective stimulation and support rhizospheric antagonistic microorganisms populations.

The aim of this work is to evaluate the antagonistic activity of indigenous rhizobacteria against two economically important phytopathogenic fungi.

II. MATERIAL AND METHODS

II.1. Material

Strains of Fusarium solani var. coeruleum (Institut Pasteur Paris, France), Fusarium oxysporum f. sp. albedinis (INRA Algiers, ALGERIA) were used for antagonism tests.

II.2. Isolation and Selection of antagonistic rhizobacteria

Soil rhizosphere samples were collected, by random sampling, from potato fields located in Sétif (Algeria), known for their high agricultural yields, as described by Mezaache-Aichour et al. [8]. The bacterial isolates inhibiting hyphal growth were purified by a series of successive passages of an isolated colony on nutrient agar 2 to 5 times [4]. A series of analyzes was performed to identify purified bacterial strains [8], dual culture with soil fungi, namely Fusarium oxysporum f. sp. abedinis, and Fusarium solani var. coeruleum allowed to highlight the antagonistic nature of these bacteria [7].

64 © 2014 UTA

III. RESULTS

III.1. Antagonistic activity of rhizobacteria

The direct tests performed in vitro by dual culture between bacterial isolates and fungal strains studied (Fusarium solani var. coeruleum, Fusarium oxysporum f. sp. albedinis) show an inhibitory action of these microorganisms. After purification of the 50 bacterial isolates, only 11 were retained according to antagonism criteria cited bellow. The inhibition ranges from 0 to 92.30 %, according to the isolate and the pathogen considered. Some isolates are highly antagonistic against the studied fungi as 202nd; while the antagonistic activity of the isolates 172nd and 242nd was lower, some other isolates are antagonist against only one fungus (Table I).

Table I. Fungi inhibition rates.

Zone 1: spermosphere ; Zone 2: rhizosphere ; 1st and 2nd: first and second soil samples; Fsc: Fusarium solani var. coeruleum ; Foa : Fusarium oxysporum f. sp. albedinis ; Nd: not determined.

IV. DISCUSSION Exploring biodiversity in microorganisms from soil

from potato fields highlighted two categories of cultivable microorganisms, predominantly bacteria (81.96 %). Bacteria represented 80.44 % of microorganisms in the soil, while the estimation of bacteria in the rhizosphere revealed 85%. Similar results were reported by Kebe et al. [6]. 11 strains selected in vitro for their antagonistic activity against fungal strains showed an important effect. Among these strains, three isolates showed an interesting activity against the phytopathogenic isolates studied, an antagonistic effect in vitro against a special formae of F. oxysporum (albedinis) a deuteromycete one (Foa) , and an ascomycete (Fsc = Selenosporium coeruleum). The

major effects were obtained against the Foa which presents itself as the most sensitive. Léon et al. [7], had reported that within 80 isolates with antifungal activity greater than 40 %, and selected from 150 microorganisms isolated, six showed antagonistic activity against the phytopathogenic fungi studied (Ascomycetes, Deuteromycetes and Oomycetes). Antagonist isolates belong to Gram-positive and Gram-negative bacteria. Among the strains of Gram-positive bacteria tested, 4.16% inhibit the growth of Fusarium solani var. coeruleum, and 6.25% inhibit growth of Fusarium oxysporum f. sp. albedinis. Gram-negative bacteria are most effective, 16.66% inhibit the growth of Fusarium solani var. coeruleum, and 14.58% inhibit growth of Fusarium oxysporum f. sp. albedinis. Of the 11 inhibitory strains, that colonize the rhizosphere, two strains inhibit only one of the two plant pathogenic fungi tested. This result is in accordance with Nion and Toyota [9] and Kamilova et al. [5] results. Nion and Toyota [9] found that from 270 isolated strains, five isolates of Burkholderia proved antagonistic activity against Ralstonia solanacearum. Kamilova et al. [5], reported that from 16 isolates exhibiting the characteristic of fluorescent Pseudomonads colonizing the tomato rhizosphere, of which only one isolate inhibit effectively four of the five tested fungi. The rhizospheric bacterial isolates origin does not seem playing, in vitro, a role in their antagonistic power. Thus, isolates from potato rhizosphere showed a diverse antagonism activity against fungi affecting other plants and from other habitats such as Foa. Défago and Haas [3] reported that antagonistic compounds produced by fluorescent Pseudomonas spp. show reduced selectivity with respect to fungi. Our results suggest the possibility of a broad spectrum effect on all fungi belonging to the twice fungal phyla studied.

V. CONCLUSION

The studied isolates belonging to different bacterial groups have developed a various antagonistic effect against phytopathogenic fungi in the laboratory, which could presage a good ability to spread through a biological control program. The practical significance of this type of studies acquires its real importance when considering the need to replace chemical control procedures for the treatment of soil and/or plant diseases.

Inhibition rate % Zones Strains

Fsc Foa

21st 37.5 53.48

51st 34.72 76.74

91st 6.94 6.97

81st 0 53.84

1

1 2nd 6.25 37.5

162nd 41.25 42.5

17 2nd 85 1.25

18 2nd 46.25 Nd

202nd 82.5 92.30

222nd 52.5 43.75

2

24 2nd 32.5 32.5

S. Mezaache-Aichour et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 64-66

65 © 2014 UTA

VI. REFERENCES

[1] L. W. Burgess “General ecology of the Fusaria”, In: “Fusarium: diseases, biology, and taxonomy”. P. E. Nelson, T. A. Toussoun, and R. J. Cook (ed.). Pennsylvania State University Press, University Park., pp. 225-235, 1981. [2] R. J. Cook, L. S. Thomashow, D. M. Weller, D. Fujimoto, M. Mazzola, G. Bangera and D. Kim, “Molecular mechanisms of defense by rhizobacteria against root disease”, Proc. Natl. Acad. Sci., 92, pp 4197-4201, 1995. [3] D. Haas and G. Défago, “Biological control of soil-borne pathogens by Fluorescent Pseudomonads” Nat. Rev. Microbiol., pp 1-13, 2005. [4] K. C. A. Jalal, M. D. Zahangir Alam, A. Suleyman Muyibi , and P. Jamal “Isolation and Purification of Bacterial Strains from Treatment Plants for Effective and Efficient Bioconversion of Domestic Wastewater Sludge” American Jour. Environ. Scien. 2 (1), pp 1–5, 2006. [5] F. L. Kamilova, S. Validov, T. Azarova, I. Mulders and B. Lugtenberg, “Enrichment for enhanced competitive plant root tip colonizers selects for a new class of biocontrol bacteria”, Environ. Microbiol., 7(11), pp 1809-1817, 2005. [6] I. B. Kebe, J. Mpika, K. F.N’guessan, P. K. Hebbar, G. S. Samuels, and S. Ake, “Isolement et identification de microorganismes indigènes de cacaoyères en Côte d’Ivoire et mise en évidence de leurs effets antagonistes vis-à-vis de Phytophthora palmivora, agent de la pourriture brune des cabosses” Sci. Nat. 6 (1):, pp 71–82, 2009. [7] M. Léon, P. M. Yaryura, M. Montecchia, A. I. Hernàndez, O. S. Correa, N. L. Pucheu, N. L. Kerber, and A. F. Garcia, “ Antifungal Activity of Selected Indigenous Pseudomonas and Bacillus from the Soybean Rhizosphere”, Inter. J. Microbiol., pp 1–9, 2009. [8] S. Mezaache-Aichour, N. Haichour, N. Sayeh, A. Guechi and M. M. Zerroug, “Isolation and Selection of Indigenous Bacterial Strains with Suppression Properties from the Rhizospheres of Potato and Wheat”, Ann. Rev. Res. Biol. 3(4), pp 405-415, 2013. [9] Y.A. J Nion, and K. Toyota, “Suppression of bacteria wilt and Fusarium wilt by a Burkholderia nodosa strain isolated from Kalimantan soils, Indonesia”. Microbes Environ. 23 (2), pp 134–141, 2008.

S. Mezaache-Aichour et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 64-66

66 © 2014 UTA

A Survey of the Possible Role of German Cockroaches as a Source for Bacterial Pathogens

Taha MENASRIA1*

, Samir TINE1, Souad EL-HAMZA1, Djaouida MAHCENE1, Fatima MOUSSA

1,

Leyla BENAMMAR2, Mohamed Nacer MEKAHLIA1

1Department of Natural and Life Sciences, Faculty of Exact Sciences and Natural and Life Sciences, University of Tebessa, Tebessa 12002, Algeria.

2Department of Natural and Life Sciences, Faculty of Sciences, University of Batna, Batna 05000, Algeria.

* Address correspondence to Taha MENASRIA, Department of Natural and Life Sciences, Faculty of Exact Sciences andNatural and Life Sciences, University of Tebessa, Tebessa 12002. Algeria. E-mail: [email protected].

Abstract This s tudy w as c arried o ut to isolate and identify bacterial f lora f rom German c ockroaches (Blattella germanica) collected from dwellings and health facilities in Tebessa city. A total of 187 bacteria were isolated from 46 c ockroach s pecimens. T he m ost c ommon an d abundant bac teria sp ecies belonged t o Enterobacteria group (54.54%), w hile p athogens l ike Staphylococcus aureus as well as Pseudomonas aeruginosa were isolated with a r ate of 12.83 % a nd 4 .81 %, respectively. The results indicated that B. germanica is a possible reservoir and carriers of bacterial pathogens v ia their bodies and thus may spread multiple drug resistance species.

Key Words: Blattella g ermanica, Pseudomonas aeuruginosa, Staphylococcus au reus, antibiotic resistance.

1. Introduction

Of the 4,000 known species of cockroaches, only a dozen can be considered as pests, living in or around human structures, especially where food is stored, served and prepared [1, 2, 3]. At night, these insects become largely active but, during the daytime, they are hidden in isolated locations [4].

Scientists suggested that cockroaches are potential vectors of pathogenic organisms in public facilities buildings and hospital environment. [3, 5, 6]. Since cockroaches occur together with the bacteria they harbour in different human habitats, such as restaurants, offices, homes even hospitals and markets; it may cause the spread of diseases and thus play the role of potential vector of pathogenic agents.

Indeed, numerous published papers recognized the association of cockroaches with several infection diseases, spread of pathogenic and drug resistant

microbes such as Staphylococcus a ureus, P seudomonas aeruginosa, Escherichia coli, Klebsiella spp, Salmonella typhi and even some fungi and viruses [7, 8, 9].

In the recent years, the incidence and severity of Staphylococcus aureus and Pseudomonas ae ruginosa have increased rapidly, and current trends indicate that it is likely and even become a greater worldwide problem [10, 11], not only because of their pathogenic potential, but also due to their antibacterial resistance increasing [12].

Notwithstanding the importance of cockroaches’ studies of in preventive medicine, there is no information on the status of the microorganism’s abundance and bacteria species harboured by the cockroaches in different habitats, dwellings, and even hospitals in Algeria. The objective of the present study is to investigate the bacterial flora isolated from alimentary tract and external surface of the German cockroach captured in Tebessa, north-eastern Algeria. As well as the determination of the antibiotic susceptibility of the two considered pathogen species Staphylococcus aur eus and Pseu domonas aeruginosa.

2. Materials and Methods

2.1 Capture and cockroaches’ collection In aseptic conditions, a set of 46 cockroaches

(Dictyoptera: Blattellidae) were trapped manually from hospitals and dwellings in Tebessa (Northeast Algeria). Each individual cockroach was conserved in sterile bottle and transported to the laboratory for microbiological analysis.

2.2 Bacterial isolation and identification Each of trapped cockroaches was frozen for 10

minutes. Then, a swab and/or bacterial suspension from the outer surface of cockroaches was performed to

67 © 2014 UTA

analyse external bacterial flora [13]. In order to evaluate the internal microbial flora, the digestive organs of each B. ge rmanica specimen were separately dissected and a homogenous suspension was prepared in nutrient broth. Aliquots (0.01 mL) of both prepared samples were separately cultured in (i) MacConkey agar (Fluka

Analytical), (ii) manitol salt agar (Difco) and (iii) cetrimide agar (Merk) overnight at 37 °C and analyzed for the total viable bacterial load. The colonies were picked and surface-streaked several times until purification [14].

Pure isolates were identified on basis of their microscopic, morphological and biochemical characters using standard methods i.e. API System (bioMérieux, France). 2.3 Susceptibility testing

Antimicrobial disk susceptibility tests were performed on Mueller-Hinton agar as recommended by the Antibiogram Committee of the French Society for Microbiology [15]. The sensitivity was recorded in each investigation as from Sensitive (S) to Resistant (R) respectively. The susceptibility was determined for ten antibiotics: [Amoxicillin + Clavulanic Acid, Cefuroxim Clindamycin, Erythromycin, Fusidic Acid, Pristinamycin, Oxacillin, Rifampicin, Spiramycin, Vancomycin]. 2.4 Data analyses

The distribution of bacterial densities over the different observations was approached using relative abundance “RA”. The variation of bacteria species richness harboured by cockroaches was tested using multiple ANOVA. The test was considered statistically significant (*) and highly significant (**) when the probability levels were P < 0.05 and P < 0.01, respectively. 3. Results and Discussion

In this study, cockroach specimens were identified as Blattella g ermanica. Totally, forty-six cockroaches were collected in Tebessa city. Only seven insects were captured from dwellings and 39 specimens from hospitals, in each of two groups, separate experiments were established and examined for the presence of external and internal bacterial flora. Three different types of flora were screened of each insect: Enterobacteria, Staphylococci and Pseudomonads group (Table 1).

Among the 187 bacteria isolated from collected cockroaches, 109 (RA=54.54%) belonged to the group of Gram-negative Enterobacteria, and only forty-two strains of Pseudomonas spp. were identified, presenting the second group of the major isolated bacteria. Indeed, German cockroach’s females exhibited the most

important reservoir of Pseudomonads with 23 strains versus males with thirteen strains only. Therefore, Pseudomonas aeruginasa presented a frequency of 4.81% of the total isolated bacteria presenting one of the minor groups of the total isolated germs (Table 1). Table 1. Relative abundance of bacterial isolates harboured by Blattella g ermanica (N = 46) collected in Tebessa.

Bacteria Abundance

Frequency

Enterobacteria 109 54.54% Pseudomonas sp. 33 17.64% Staphylococcus aureus 24 12.83% Non-pathogenic Staphylococcus

9 4.81%

Pseudomonas aeruginosa 9 4.81% Overall 187 100%

Numerous studies have investigated the relationship between the environment and germs carried by cockroaches, highlight potential risk of human contamination in residential areas and hospitals [13].

The present results indicated a bacterial contamination of all German cockroaches collected from the surveyed location (dwellings and hospitals), while looking the bacterial diversity of these insect by isolating and identifying their bacterial flora. Far, a great number of germs were isolated from cockroaches captured in dwellings, hospitals or other local [13, 16]. Indeed, Bouamama and colleagues suggested that cockroaches harbored several pathogenic bacteria including Staphylococci and Pseudomonas [7].

Table 2. ANOVA testing the effect of factors “habitat”, “sex”, their interactions on bacterial richness harboured by German cockroach. Test statistics (F and P values) are Type III. (NS: not significant, *: significant, **: highly significant).

Source Df SS MS F P Habitat 1 45.38 45.38 10.48 0.002 ** Sex 1 37.50 37.50 8.66 0.004 ** Habitat vs. Sex 1 16.67 16.67 3.85 0.053 NS

No significant difference in bacterial species richness of cockroach individuals was observed compared to the habitat and sex. However, the MANOVA test revealed that the habitat and sex showed highly significant effect (P = 0.002, P = 0.004, respectively) on variation of bacteria species richness isolated from cockroaches.

Overall, the relative abundance of Staphylococci presented similarities between the two sexes of Blattella germanica. The predominant Staphylococci group bacteria was S. aur eus with a frequency of 12.83%, followed by non-pathogen Staphylococci with only nine species presented a relative abundance of 4.81% (Table

T. Menasria et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 67-70

68 © 2014 UTA

1), distributed on both sides of the surface and/or the alimentary tract. Female cockroaches exhibited the most important reservoir of Staphylococcus aureus with RA = 36.11% and RA = 36.6% in both part of the body, versus males with 18.1% in alimentary tract and 35% on external surface (Figure).

0

20

40

60

80

AT ES AT ES

Male () (n=17) Female () (n=29)

Bac

terial

fre

quency

%

Pseudomonas sp. Pseudomonas aeruginosaStaphylococcus sp. Staphylococcus aureus

Figure. Bacterial frequency of major bacterial groups

isolated from Blatella germanica (AT: Alimentary tract, ES: External surface).

The susceptibility of the two selected pathogens (Staphylococcus aureus and Pseudomonas aeurginosa) is summarized in Table 3. The results indicated that all the isolated strains of Pseudomonas aeruginosa were resistant to most tested antibiotics including: oxacillin, ampicillin, cefuroxime, pristinamycin, fusidic acid, erythromycin, vancomymcine and spiramycin. While, only two antibiotics (rifampicin and clindamycin) showed anti-Pseudomonas activity with significant differences.

Table 3. Antibiotic susceptibility patterns of Staphylococcus aureus and Pseudomonas aeruginosa isolates form Blatella germanica. [AMC: Amoxicillin + Clavulanic Acid, CXM : Cefuroxim, CN: Clindamycin, E: Erythromycin, FA: Fusidic Acid, OX: Oxacillin, PT: Pristinamycin, RA: Rifampicin, SP: Spiramycin,VA: Vancomymcin]. ND: Not determined. Res: Resistant; Sen: Sensetive.

Antibiotic susceptibility % P. aeruginosa (N=8) S. aureus (N=7)

Antibiotic

Res Sen Res Sen OX 100 0 62.5 37.5 AMC 100 0 75 25 CXM 100 0 ND ND RA 0 100 50 50 PT 100 0 25 75 FA 100 0 87.5 12.5 CN 0 100 12.5 87.5 VA 100 0 37.5 62.5 E 100 0 75 25 SP 100 0 37.5 62.5

In the present investigation, all Staphylococcus aureus

isolates were found to be highly resistant to oxacillin (62.5%), ampicillin (75%), 87.5% for fusidic acid and 75% for erythromycin. However, susceptibility patterns of S. aureus isolates in our findings showed that the majority of the resistances were exhibited toward to the commonly used antimicrobial.

The majority of Pseudomonas aeruginosa and Staphylococcus aureus tested in the present investigation, showed resistance. Further, research indicated that multi-drug resistance can be displayed [17, 18, 19, 20].

Thus, the findings clearly showed that all collected cockroaches in homes even hospitals harbored a large diversity of microorganisms. This high prevalence of the microorganisms and resistant ones harbored in the body of the insects portends public health risks, transmission of community-acquired and nosocomial infections. Therefore, further studies are clearly necessary to be carried out in order to investigate relevant control methods against cockroaches.

Acknowledgements We gratefully acknowledge the staff of Microbiology

Laboratory at the Department of Natural and Life Sciences (University of Tebessa) for all facilities provided during carrying out this study. References

[1] G. W. Bennett, “Cockroaches and Disease”, In: J. L. Capinera, (Ed.) Encyclopedia of Entomology. Springer, pp. 948–952, 2008.

[2] J. Dubus, M. Guerra, A. Bodiou, “Cockroach allergy and asthma”, Allergy, 56, 351–352, 2001.

[3] I. Pavlik, J. O. Falkinham, “The Occurrence of Pathogenic and Potentially Pathogenic Mycobacteria in Animals and the Role of the Environment in the Spread of Infection”, In: J. Kazda, I. Pavlik, J. O. Falkinham, K. Hruska, (eds.) The Ecology of Mycobacteria: Impact on Animal’s and Human’s Health. Springer, pp. 199–281. 2009.

[4] D. G. Cochran, “Cockroaches: their biology, distribution and control, World Health Organization, Document No: WHO/CDS/CPC/WHOPES/99.3. Geneva, Switzerland, 1999.

[5] M. F. Cotton, E. Wasserman, C. H. Pieper, D. C. Theron, D. Van-Tubbergh, G. Campbell, F. C. Fang, J. Barnes, “Invasive disease due to extended spectrum beta lactamase-producing Klebsiella pneumoniae in a neonatal unit: the possible role of cockroaches”, J. Hosp. Infect, 44, 13–17, 2000.


69 © 2014 UTA

[6] J. Tachebe, W. Erku, L. T. Gerbe-Michae, M. Ashenafi, “Cockroach associated food-borne bacteria from hospital and restaurant in Addis Ababa, Ethopia: Distribution and antibiograms”, J. Rural. Trop. Public Health, 5, 34–41, 2006.

[7] L. Bouamama, M. Lebbadi, A. Aarab, “Bacteriological analysis of Periplaneta americana L. (Dictyoptera; Blattidae) and Musca domestica L. (Diptera; Muscidae)” in ten districts of Tangier, Morocco, Afr. J. Biotechnol, 6, 2038–2042, 2007.

[8] R. Dajoz, “ Dictionnaire d’entomologie : Anatomie. Systématique “, Biologie, Lavoisier, Tec & Doc. Paris, pp. 38–39. 2010.

[9] M. A. Prado, F. C. Pimenta, M. Hayashid, P. R. Souza, M. S.Pereira, E. Gir, “ Enterobacteria isolated from cockroaches (Periplaneta americana) captured in a Brazilian Hospital “ Pan. Ame. J. Public Health, 11, 93–98, 2002.

[10] D-S. Lee, S-H. Eom, S-Y. Jeong, H. J. Shin, J-Y. Je, E-W. Lee, Y-H. Chung, Y-M. Kim, C-K.Kang, M-S. Lee, “Anti-methicillin-resistant Staphylococcus au reus (MRSA) substance from the marine bacterium Pseudomonas sp. UJ-6“ Env. Toxicol. Pharmacol, 35, 171-217, 2013.

[11] F. Nanvazadeh, A. D. Khosravi, M. R. Zolfaghari, N. Parhizgari, “Genotyping of Pseudomonas aeruginosa strains isolated from burn patients“ by RAPD-PCR.Burns. .doi.org/10.1016 /j.burns.2013.03. 008.

[12] J. R.Fitzgerald, “Evolution of Staphylococcus aureus during human colonization and Infection” Inf, Genet Evolution,.doi.org/10.1016/j.meegid 2013.04.020.

[13] X. Fu, L. Ye, F. Ge, “Habitat influences on diversity of bacteria found on German cockroach in Beijing” J. Environ. Sci., 21, 249–254, 2009.

[14] J. P. Guiraud, “ Microbiologie alimentaire “, Dunod, Paris. 2003

[15] Comité de l’Antibiogramme de la Société Française de Microbiologie, Int. J. Antimicrob. Agents, 21, 364–391, 2003.

[16] B. Kutrup, Cockroach Infestation in Some Hospitals in Trabzon, Turkey. Turk. J. Zool., 27, 73-77, 2003.

[17] H. Fathpour, G. Emtiazi, E. Ghasemi, “Cockroaches as Reservoirs and Vectors of Drug Resistant Salmonella” spp. Iran. Biomedical J., 7, 35–38, 2003.

[18] H.-H. Pai, W.-C. Chen, C.-F. Peng, “Cockroaches as Potential Vectors of Nosocomial Infections” Infect. Control. Hosp. Epidemiol, 25, 979–984, 2004.

[19] A. Salehzadeha, P. Tavacol, H. Mahjub, “Bacterial, fungal and parasitic contamination of cockroaches in public hospitals of Hamadan”, Iran. J. Vector. Dis., 44, 105–110, 2007.

[20] G. R. Oliva, C. Díaz, O. Fuentes González, M. D. Martínez, C. Fernández, R. Cordovi, P. M. Lago, N. Herrera, “Blatella germanica as a possible cockroach vector of micro-organisms in a hospital”, J. Hosp. Infect., 74, 93–95, 2010.


70 © 2014 UTA

Abstract— The behavior of fiber reinforced polymer (FRP)-confined concrete in circular columns has been extensively studied, but much less is known about concrete in FRP-confined square columns in which the concrete is non-uniformly confined and the effectiveness of confinement is much reduced. An experimental study has been carried out on square reinforced concrete (RC) columns strengthened with carbon fiber-reinforced polymer (CFRP) sheets. A total of 48 specimens were loaded to failure in axial compression and investigated in both axial and transverse directions. Slenderness of the columns, number of wrap layers and concrete strength were the test parameters. Compressive stress, axial and hoop strains have been recorded to evaluate the stress-strain relationship, ultimate strength, stiffness, and ductility of the specimens.

Key words: Square concrete column, confinement, CFRP composite, compressive strength, ultimate strain.

I. INTRODUCTION

n increasing number of reinforced concrete structures have reached the end of their service life, either due to

deterioration of the concrete and reinforcements caused by environmental factors, or due to an increase in applied loads. These deteriorated structures may be structurally deficient or functionally obsolete, and most are now in serious need of extensive rehabilitation. Carbon fiber reinforced plastics sheets or plates are well suited to this application because of their high strength-to-weight ratio, good fatigue properties, and excellent resistance to corrosion. Their application in civil engineering structures has been growing rapidly in recent years, and is becoming an effective and promising solution for strengthening deteriorated concrete members. Because CFRPs are quickly and easily applied, their use minimizes labor costs

and can lead to significant savings in the overall costs of a project.

During the last decade, the use of FRP composites has been successfully promoted for external confinement of reinforced concrete (RC) columns all over the world. Several studies on the performance of FRP wrapped columns have been conducted, using both experimental and analytical approaches (Saadatmanesh et al., 1994; Nanni and Bradford, 1995; Karbhari and Gao, 1997; Berthet et al., 2005; Benzaid et al. 2009; Benzaid et al. 2010). Such strengthening technique has proved to be very effective in enhancing their ductility and axial load capacity. However, most of the available studies on the behavior of FRP confined concrete columns have concentrated on circular shaped columns with normal strength. The data available for columns of square or rectangular cross sections have increased over recent years but are still limited (Chaallal et al., 2003; Al-Salloum 2007, Rochette et al. 2000; Benzaid et al. 2008). Also the validation of these results and their applicability to large-scale RC columns is of great practical interes. This field remains in its infancy stages and more research investigation is needed on this subject to study the effect of slenderness (Pan et al., 2007) and that of concrete strength.

This study deals with a series of tests on square plain concrete (PC) and reinforced concrete (RC) columns strengthened with CFRP sheets. A total of 48 concrete specimens were tested under axial compression. The data recorded included the compressive loads, axial strains, and radial strains. The parameters considered are the number of composite layers (1 and 3), the compressive strength of the unconfined concrete (25MPa and 60MPa) and the columns’ slenderness ratio L/a (2; 4 and 7.4). To comply with existing RC members in practice, where reduced cover is often present, the corners for all prismatic specimens were almost kept sharp for CFRP application.

Mechanical Behavior of Square Concrete Columns Wrapped with CFRP Composite

Riad Benzaid 1,2, Nasr eddine Chikh 2, Habib Mesbah 3.

1L.G.G. Laboratory, Jijel University. BP. 98 Ouled Issa, Jijel- 18000, Algeria. 2 L.M.D.C. laboratory, Dept of Civil Engineering, Mentouri University – Constantine,

Route Ain El Bey, Constantine 25000, Algeria. 3 L.G.C.G.M. laboratory, Department of Civil Engineering, INSA de Rennes, France.

20, Av. des Buttes de Coësmes – CS 70 839 – 35708 – Rennes Cedex7, France.

E-mail: [email protected]; 2 habib-abdelhak.mesbah@ univ-rennes1.fr; 3Chikh_ne@ yahoo.fr

A

71 © 2014 UTA

II. EXPERIMENTAL PROGRAM

A. Materials

Two kind of concrete mix have been realised to investigate the influence of concrete strength as indicated in Table 1. The two categories represent normal strength concrete (NSC) and high strength concrete (HSC).

The carbon-fiber sheets used were the SikaWrap-230C product, a unidirectional wrap. The manufacturer’s guaranteed tensile strength for this CFRP is 4300 MPa, with a tensile modulus of 238 GPa, an ultimate elongation of 1.8 % and a fiber thickness of 0.13mm. The Sikadur-330 epoxy resin was used to bond the carbon fabrics over the square columns.

This research work was carried out in the laboratory of the civil engineering department (I.U.T-University of Rennes 1- France). Eight series of experiments were performed to investigate the behavior of PC and RC square columns confined by CFRP composite. Table II summarizes the specimens involved in the experimental program.

TABLE I. CONCRETE MIXTURE PROPORTIONS

Concrete mixture No. I II Concrete strength f’co, MPa 25.93 61,81 Cement, kg/m3 280(1) 450(2) Water, kg/m3 180.00 170.00 Crushed gravel, kg/m3 Ø 4/6 122.90 115.60 Ø 6/12 258.20 242.80 Ø 12/20 769.50 728.50 Sand Ø 0/4, kg/m3 729.10 685.60 Sika Viscorete-Tempo12(3) , ml - 1550.00 W/C 0.64 0.38

(1) Portland cement: CPA CEM II R 32.5 MPa. (2) Portland cement: CPA CEM I R 52.5 MPa.

For all RC specimens the diameter of longitudinal and transverse reinforcing steel bars were respectively 12 mm and 8 mm. The longitudinal steel ratio was constant for all specimens and equal to 2.25%.The yield strength of the longitudinal and transversal reinforcement was 500 MPa and 235 MPa; respectively.

The specimen notations are as follows. The specimen notations are as follows. The first two letters refer to the type of concrete: PC for plain concrete and RC for reinforced concrete, followed by the concrete mixture: I for normal strength (24.77MPa) and II for high strength (59.53MPa). The next letter indicates the slenderness ratio: x for L/a=2, y for L/a=4 and z for L/a=7.14. The last number specifies the number of layers.

B. Specimen Preparation

After concrete columns were fully cured, FRP wrapping procedure was performed according to the procedure specified by the manufacturer. The CFRP jackets were applied to the specimens by manual wet lay-up process. The concrete specimens were cleaned and completely dried before the resin was applied. The epoxy resin was directly applied onto the substrate.

TABLE II. DETAILS OF TEST SPECIMENS

Specimen designation

Concrete mixture

Nominal dimensions (diameter x

height) [mm]

Number of layers

Unconfined concrete

strength [MPa]

Number of specimens

PCI. x0 140x140x28 - 2

PCI. x1 140x140x28 1 2

PCI. x3 140x140x28 3 2

RCI. x0 140x140x28 - 2

RCI. x1 140x140x28 1 2

RCI. x3 140x140x28 3 2

RCI. y0 140x140x56 - 2

RCI. y1 140x140x56 1 2

RCI. y3 140x140x56 3 2

RCI. z0 140x140x10 - 2

RCI. z1 140x140x10 1 2

RCI. z3

I

140x140x10 3

24.77

2

PCII. x0 140x140x28 - 2

PCII. x1 140x140x28 1 2

PCII. x3 140x140x28 3 2

RCII. x0 140x140x28 - 2

RCII. x1 140x140x28 1 2

RCII. x3 140x140x28 3 2

RCII. y0 140x140x56 - 2

RCII. y1 140x140x56 1 2

RCII. y3 140x140x56 3 2

RCII. z0 140x140x10 - 2

RCII. z1 140x140x10 1 2

RCII. z3

II

140x140x10 3

59.53

2

The fabric was carefully placed into the resin with gloved hands and smooth out any irregularities or air pockets using a plastic laminating roller. The roller was continuously used until the resin was reflected on the surface of the fabric, an indication of fully wetting. A second layer of resin was applied to allow the impregnation of the CFRP. The following layer is applied in the same way. Finally, a layer of resin was applied to complete the operation.

Each layer was wrapped around the column with an overlap of ¼ of the perimeter to avoid sliding or deboning of fibers during tests. The wrapped specimens were left at room temperature for 1 week before testing.

Fig. 1. Specimens after curing and wrapping

R. Benzaid et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 71-75

72 © 2014 UTA

C. Test Procedures

Specimens were loaded under a monotonic uni-axial compression load up to failure. The load was applied at a rate corresponding to 0.24 MPa/s and was recorded with an automatic data acquisition system. Axial and lateral strains were measured using appreciable extensometer. The instrumentation included one lateral linear variable differential transducer (LVDT) placed in the form of a square frame at the mid-height of the specimens. Measurement devices also included three vertical LVDTs to measure the average axial strains. Prior to testing, all CFRP-wrapped columns were capped with sulfur mortar at both ends.The test setup for the various specimens is shown in Fig. 2.

Fig. 2. Test setup

III. TEST RESULTS AND DISCUSSION Compression behavior of the CFRP wrapped specimens

was mostly similar in each series in terms of stress-strain curves and failure modes of the specimens. All confined cncrete columns failed by fracture of the composite wrap at one of the corners, because of the high stress concentration at these locations, Fig. 3. The collapse occured in a sudden and explosive way, though some popping noises were heard during various stages of loading and were attributed to microcracking of the concrete. The strain values observed for the jacket tensile failure were quite lower that the FRP failure strain, as many authors have already published.

For short specimens (L/a =2), the fiber rupture starts mainly in their central zone, then propagates towards both ends. Regarding slender specimens, the collapse was mostly concentrated in their end regions, indicating that the greater the slender ratio, the smaller the area of CFRP ruptured.

For these columns at ultimate load, when confinement action was no longer provided due to FRP fracture, the internal steel started buckling and the crushed concrete fell down between the fractured FRP. Hence, this indicates that the concrete core is significantly damaged (but yet confined) even before reaching ultimate load.

For all confined specimens, delamination was not observed at the overlap location of the jacket, which confirmed the adequate stress transfer over the splice.

The average experimental results are reported in Table III, with the increase in terms of compressive strength (f’cc/fco) and ductility (εcc/εco), intended as ultimate axial displacement.

Fig. 3. Failure of CFRP confined specimens

Representative stress-strain curves for each series of tested

CFRP-wrapped specimens are reported in Fig. 4 (a-c) for NSC and in Fig. 5 (a-c) for HSC. These figures give the axial stress versus the axial and lateral strains for specimens with zero, 1 and 3 layers of CFRP wrap considering various slenderness ratio L/a (2, 4 and 7.14).

A. Stress-Strain Response

For NSC, all CFRP strengthened specimens showed a typical bilinear trend with a transition zone. Three zones can be observed for the stress-strain curves of the CFRP-confined cylinders. The first zone is essentially a linear response governed by the stiffness of the unconfined concrete, which indicates that no confinement is activated in the CFRP wraps since the lateral strains in the concrete are very small. The unconfined concrete specimens show a sudden drop in stiffness and strength after reaching the maximum load point. In the second zone, a nonlinear transition occurs as the concrete expands, thus producing larger lateral strains. The CFRP wrap reacts accordingly and a confining action is created on the concrete core. During this stage a loss of stiffness occurs due to the rapidly growing network of cracks in the concrete. Finally, in the third zone, the concrete is fully cracked and the CFRP confinement is activated to provide additional load carrying capacity by keeping the concrete core intact. The stress-strain curve here increases linearly up to failure. However, no distinct post behaviour is observed for specimens with higher slenderness ratio. On overall, both ultimate compressive strength and ultimate strain are variably enhanced depending on the number of layers and the slenderness ratio.

As for the previous case of NSC, the first slope of the curve, regarding specimens with HSC, is also not substantially altered by the presence of CFRP. In this initial elastic zone, the confined and the unconfined specimens behave in the same manner, irrespective of the number of layers. The strengthening effect of the CFRP layers begins only after the concrete has reached the peak strength of the unconfined concrete: transversal strains in the concrete activate the CFRP jacket. The increase of load would produce large lateral


73 © 2014 UTA

expansions, and consequently a higher confining pressure, provided that the number of composite layers is quite sufficient. With low levels of confinement (one CFRP layer), the second part of the bilinear curve shifts from strain hardening to a flat plateau with a drastically reduced ductility.

No distinct post behaviour is observed as the slenderness ratio increases and little improvement is achieved in both strength and ductility.

B. Effect of CFRP Strengthening Ratio

In all cases the increase of the numbers of sheets generated an increase of compressive strength as well as axial deformation capacity. The level of increase is important for NSC specimens. Considering the cases of 1 and 3 CFRP layers, from results displayed in Table III and Fig. 4, it can be evaluated that the increase in the bearing capacity varies roughly from 17% to 58% as compared to the relative unconfined specimens, while the ultimate vertical deformations increase on average from 80% to 280%. From these findings, it is possible to assert that the increase in the number of CFRP sheets has a significant influence even though the increase in terms of strength is not as important as that of axial deformations which increase almost proportionally to the FRP strengthening ratio.

The effect of the number of CFRP layers on HSC specimens is relatively moderate compared to previously. In this situation, the confinement pressure is activated at higher load (around 80% of the ultimate value). Consequently, the enhancement in the loading carrying capacity is reduced and varies roughly from 8% to 29%, whereas the ultimate axial deformations undergo a significant reduction displaying an increase on average from 21 to 125%, as illustrated in Table III and in Fig. 5.

It should be emphasized that the presence of quite sharp corners in all tested CFRP jacketed columns produced a cutting effect on confining sheets and hence affected the rate of enhancement in their load carrying and deformation capacities.

C. Effect of slenderness ratio

The comparison of results recorded for a slenderness ratio varying from 2 to7.14 shows for NSC wrapped RC specimens a moderate decrease in the load carrying capacity and an important reduction in the axial deformation.

However, in the case of HSC jacketed specimens, the strength was almost not affected whereas the ductility undergoes a moderate decrease. This may be explained by the late activation of the confinement pressure which occurred at higher load (around 80% of the ultimate value).

On overall, the efficiency of the confinement provided by composite wraps was greatly affected by the premature damage of the CFRP fabric at the sharp column corner.

TABLE III. DETAILS OF TEST SPECIMENS

Specimendesignation

f’co

[MPa]

f’cc [MPa]

f’cc/fco εcc

[‰] εcc/ εco

εh,rup

[‰] εh,rup/ εho

PCI. x0 24.7 24.77 1.00 2.20 1.00 3.88 1.00

PCI. x1 27.66 1.11 6.58 2.99 24.33 6.27

PCI. x3 32.03 1.29 5.89 2.67 17.01 4.38

RCI. x0 33,59

33.59 1.00 7.61 1.00 17.00 1.00

RCI. x1 39.52 1.17 15.06 1.97 20.58 1.21

RCI. x3 49.12 1.46 15.66 2.05 25.50 1.50

RCI. y0 30.49

30.49 1.00 1.67 1.00 9.78 1.00

RCI. y1 36.73 1.20 3.02 1.80 8.69 0.88

RCI. y3 41.85 1.37 5.61 3.35 7.77 0.79

RCI. z0 24.69

24.69 1.00 0.96 1.00 - -

RCI. z1 33.92 1.37 2.05 2.13 - -

RCI. z3 39.17 1.58 3.64 3.79 - -

PCII. x0 59.5 59.53 1.00 3.66 1.00 3.31 1.00

PCII. x1 61.30 1.02 2.46 0.67 5.43 1.64

PCII. x3 70.35 1.18 3.09 0.84 13.39 4.04

RCII. 0

63.79

63.79 1.00 3.05 1.00 10.30 1.00

RCII. 1

74.84 1.17 3.87 1.26 16.36 1.58

RCII. 3

79.59 1.24 5.29 1.73 7.96 0.77

RCII. 0

63.62

63.62 1.00 2.05 1.00 0.35 1.00

RCII. 1

80.78 1.26 2.82 1.37 0.76 2.17

RCII. 3

82.44 1.29 2.79 1.36 0.76 2.17

RCII. z0 69.98

69.98 1.00 2.08 1.00 0.49 1.00

RCII. z1 75.77 1.08 2.53 1.21 0.82 1.67

RCII. z3 81.51 1.16 2.70 1.29 1.36 2.77


74 © 2014 UTA

0

10

20

30

40

50

60

70

80

90

100

-30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30


Co

ntr

ain

te a

xial

e (M

Pa)

Non confiné

1couche

3 couches

0

10

20

30

40

50

60

70

80

90

100

-30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30


Co

ntr

ain

te a

xial

e (M

Pa)

Non confiné

1couche

3 couches

0

10

20

30

40

50

60

70

80

90

100

-30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30

Déformatio radiale(‰) Déformation axiale(‰)

Co

ntr

ain

te a

xial

e (M

Pa)

Non confiné

1couche

3 couches

Fig. 4. Stress strain curves of NSC CFRP confined specimens

IV. CONCLUSIONS

In this paper an experimental program has been presented whose aim is to study the axial compression behaviour or reinforced concrete columns of a square cross-section confined externally with CFRP. The following conclusions can be drawn from the study:

• The failure of all CFRP wrapped specimens occurred in a sudden and explosive way preceded by typical creeping

sounds. For short specimens (L/a =2), the fiber rupture starts mainly in their central zone, then propagates towards both ends. Regarding slender specimens, the collapse was mostly concentrated in their upper or lower regions, indicating that the greater the slender ratio, the smaller the area of CFRP ruptured;

• On overall, CFRP strengthened specimens showed a typical bilinear trend with a transition zone. The first zone is essentially a linear response governed by the stiffness of the unconfined concrete. No distinct post behaviour is observed as the slenderness ratio increases and little improvement is achieved in both strength and ductility.

• Increasing the amount of CFRP sheets produce an increase in the compressive strength of the confined column but with a rate lower compared to that of the deformation capacity which almost proportional to the CFRP strengthening ratio;

• The CFRP confinement on low-strength concrete specimens produced higher results in terms of strength and strains than for high-strength concrete similar specimens. Therefore, the effect of CFRP confinement on the bearing and deformation capacities decreases with increasing concrete strength;

• The effect of increasing the slenderness ratio results in a decrease of the strengthening effect on strength and ductility. The rate of decrease is more important for NSC specimens.

REFERENCES [1] Y. A. Al-Salloum, “Influence of Edge Sharpness on the Strength of

Square Concrete Columns Confined With FRP Composite Laminates”, Composite Part B, 38, 640–650, 2007.

[2] J. Berthet, E. Ferrier, P. Hamelin, “Compressive behavior of concrete externally confined by composite jackets”, Part A: experimental study. Constr. Build. Mater., 19 (3), 223-232, 2005.

[3] R. Benzaid, N-E. Chikh, H. Mesbah, “Behaviour of square concrete columns confined with GFRP composite wrap”, J. Civ. Eng. Manag., 14 (2), 115-120, 2008.

[4] R. Benzaid, N-E. Chikh, H. Mesbah, “Study of the compressive behavior of short concrete columns confined by fiber reinforced composite”, Arab. J. Sci. Eng., Section B, 34 (1B), 15-26, 2009.

[5] R. Benzaid, H. Mesbah, N-E. Chikh, “FRP-Confined Concrete Cylinders: Axial Compression Experiments and Strength Model” J. Reinforced Plastics and Composites, 29 (16), 2469-2488, 2010.

[6] O. Chaallal, M. Hassen, M. Shahawy, “Confinement model for axially loaded short rectangular columns strengthened with FRP polymer wrapping” J. ACI Struct., 100 (2), 215-221, 2003.

[7] V. M. Karbhari, Y. Gao, “Composite jacketed concrete under uniaxial

compression-verification of simple design equations”, J. Mater. Civ. Eng., 9 (4), 185-193, 1997.

[8] A. Nanni, N. M. Bradford, “FRP jacketed concrete under uniaxial compression”, Constr. Build. Mater., 9 (2), 115-124, 1995.

[9] J. L. Pan, T. Xu, Z. J. Hu, “Experimental investigation of load carrying capacity of the slender reinforced concrete columns wrapped with FRP”, Construct. Build. Mater., 21, 1991–1996, 2007.

[10] P. Rochette, P. Labossière, “Axial Testing of Rectangular Column Models Confined with Composites”, ASCE J. Compos. Construct., 4 (3), 129–136, 2000.

[11] H. Saadatmanesh, M. R. Ehsani, M. W. Li, “Strength and ductility of concrete columns externally reinforced with composites straps” J. ACI Struct., 91(4), 434-447, 1994.


75 © 2014 UTA

Abstract— A method for predicting the chloride ingress into concrete structures, which is the first and main phase of chloride induced reinforcement corrosion, is developed. The heat transfer, moisture transport and chloride diffusion, that contribute to the rate and amount of transported chlorides into concrete are modeled. In modeling the chloride transport, a modified version of Fick’s second law is used, in which processes of diffusion and convection due to water movement are taken into account. The proposed finite element model and its associated program are capable of handling pertinent material nonlinearities and variable boundary conditions that simulate real exposure situations. The numerical performance of the model was examined through few examples to investigate the effect of the different model parameters and mechanisms considered on the process of chloride penetration into concrete, and eventually, on the service life of concrete.

Index Terms— chloride penetration, concrete structure, modeling, cold regions, transport processes

I. INTRODUCTION

he corrosion of reinforcing steel in concrete, mainly caused by chloride penetration, is one of the major deterioration mechanisms that cause severe damage of

reinforced concrete structures in cold or marine regions. When chloride concentration at re-bar depth reaches a certain value, called chloride threshold concentration, the iron oxide film protecting the steel is destroyed. Depassivation of steel bars leads to the formation of rust in the interface zone between the embedded steel and concrete. The large volume expansion associated with rust formation causes cracking in concrete cover, lost of bond between the steel and concrete, and eventually the reduction of cross section area of the steel, which reduces the load carrying capacity of the reinforced concrete structural member, or even causes collapse of the structure. The sources of chlorides in concrete are either internal from components of concrete mixture (aggregates and water) or external mainly from seawater or soluble deicing salts. The chloride ions penetrate into saturated concrete either driven by concentration gradient (diffusion) or by electrical potential gradient (electromigration). The latter process is relevant only when a strong electric potential is imposed on the concrete member. For non-saturated concrete, moisture

movement driven by the pressure gradient contributes to the chloride penetration into concrete, because when moisture moves chloride ions are carried by the moisture. This process is the fastest and is active when the concrete member is partially saturated.

There are two different approaches in modeling the chloride penetration into concrete. One is based on the conventional diffusion model, i.e., Fick’s first and second laws [1-4]. The other is based on the basic laws of electrochemistry, particularly the Nernst-Plank equation [5-6]. The major difference in the two approaches is that Fick’s laws consider single ion (i.e. chloride ion) diffusion; while Nernst-Plank equation includes the coupling effects of multispecies diffusion in concrete (Na, K, etc, in addition to chloride). A common feature of the two approaches is that the parameters in the governing differential equations are all effective parameters at large scale, which means they are averaged parameters over a representative volume. In recent years, a new approach was developed that treats the problem at multiscale levels [7-10]. In this approach, diffusion theory based on Fick’s laws was used to formulate the problem at the macroscopic level, and the composite material theory was used to incorporate the effects of aggregate and cement paste pore structure on the material parameters at the mesoscopic level; and furthermore, the diffusion mechanisms of chloride ions were taken into account at the microscopic level.

For saturated concrete, Xi and Bazant [7] modeled the chloride diffusion using the new multiscale approach, and they used the finite differences method to solve the governing differential equations. The material models developed were more detailed and took into account the effect of temperature, concrete age, concrete composition (e.g. aggregate content), and type of cement. Furthermore, the material parameters in the model were considered to be dependent on the chloride concentration, rendering the system of high nonlinearity. Hansen and Saouma [4] modeled the chloride diffusion problem using the finite elements method and took into account the effect of temperature and time on the process.

For non-saturated concrete, Saetta et al. [1] developed a model that took into account the effect of humidity and temperature. Thus, it was able to simulate the chloride penetration in non-saturated concrete structures exposed to marine environments or in cold regions. Martin-Pérez [11] and Isgor [12-13] followed the same methodology as Saetta et al. and coupled the chloride penetration to the heat

Modeling Of Chloride Penetration In Concrete Structures In Cold Regions

H’mida Hamidane1, Ayman Ababneh2

1 Department of Civil Engineering, University of Tebessa, Tebessa, Algeria ([email protected]) 2 Department of Civil Engineering, Jordan University of Sciences and Technology, Irbid, Jordan

T

76 © 2014 UTA

and moisture transfer, and further, considered the propagation stage of corrosion. Ababnah et al. [8] extended the work of Xi and Bazant [7] on saturated concrete to unsaturated concrete. Later on, Suwito et al. [10] further extended the model by characterizing the chloride and moisture diffusions as two fully coupled processes and solving the coupled partial differential equations by an advanced parallel computing technique.

The focus of this paper is the effect of low temperature on chloride penetration in concrete. There is a pressing need to study the chloride penetration process in concrete in cold regions and understand the effect of cold weather on the chloride-induced corrosion problem of concrete infrastructure. The framework of the multiscale modeling of chloride penetration [7, 8, 10] is used and further extended to cover the coupling effects among chloride penetration, moisture diffusion, and heat transfer. It is important to emphasize that the focus of the present study is the diffusion processes in the low temperature range. Therefore, pore pressure built-up and vaporization of pore water under high temperatures are not considered in the model. Moreover, the coupling effects among chloride, moisture, and temperature is considered to be partial coupling: the effects of moisture and temperature on the chloride penetration are taken into account in the model, the effect of chloride penetration on moisture diffusion are considered, but the effects of chloride penetration and moisture movement on temperature are not modeled. In another word, the chloride and moisture diffusions are fully coupled, and the heat transfer is considered to be an independent process.

In the following sections, the governing equations will be described first, followed by a detailed introduction on material parameters involved in the equations. The governing partial differential equations will be solved by a two-dimensional time dependent non-linear finite element technique. The numerical performance of the model will be analyzed through a sensitivity analysis, and the predicted results by the present model will be compared to available test data.

II. MODEL DESCRIPTION

A. Review Stage

The transport of chlorides, moisture and heat can be described by the following coupled system of differential

equations:

CC wt f Div D grad C CCl f fC t tf

(1)

Cw w H fDiv D grad HH

t H t t

(2)

TC Div D grad TT

t

(3)

in which Ct and Cf are total and free chloride concentration in concrete, respectively; w is moisture content and H is pore relative humidity in concrete; T is temperature; μ is a coupling parameters that reflect the effect of moisture transport on the chloride penetration; θ is the coupling parameter for the effect of chloride penetration on moisture diffusion; is the chloride diffusion coefficient of concrete expressed in [cm2/day]; is the chloride binding capacity of concrete; is the humidity diffusion coefficient of concrete expressed in [cm2/day]; is the concrete moisture capacity; is the thermal conductivity of concrete [W/m.˚C]; is the specific heat capacity of concrete in [J/kg.˚C]; and is the concrete density in [kg/m3].

In the above system of equations, there are three state variables: the free chloride concentration Cf, pore relative humidity H, and temperature T. The first equation describes chloride penetration process in concrete. The first term on the right hand side is for the driving force of free chloride gradient, the second term for the coupling effect of moisture diffusion on chloride penetration. This equation includes the effects of all three state variables on the chloride penetration in concrete. The second equation characterizes the moisture diffusion process in concrete. The first term on the right hand side is for the driving force of moisture gradient, and the second term for the coupling effect of chloride penetration on moisture diffusion. This equation includes the effects of two state variables, Cf and H on the chloride penetration in concrete. The third equation represents heat transfer in concrete. The coupling effects of chloride penetration and moisture diffusion on heat transfer are neglected. The solutions of the three equations provide comprehensive information of free chloride concentration, pore relative humidity, and temperature in the concrete at any location and at any time.

III. MATERIAL PARAMETERS

There are totally eight material parameters in Eqs. (1), (2), and (3): three for the chloride penetration, three for the moisture diffusion, and two for the heat transfer. The material models for the eight parameters are very important because the accuracy of the prediction model depends on the accuracy of material models. The difficulty in developing reliable material models is due partly to the heterogeneous features of concrete materials and partly to continuous development of microstructure of cement paste with time. The heterogeneity of material components of concrete exhibits in different sizes and shapes, ranging from cement particles at micrometer level to coarse aggregate at millimeter level, and from high angularity crushed stone to smoothed surface river gravel. All components in concrete contribute to transport properties of the material. The overall effect of the components will be characterized by composite models at different scale levels. At the mesoscale level, the concrete is considered as a composite with coarse aggregates as inclusions and cement mortar as matrix. At the micrometer level, the mortar is considered as a composite with fine sand as inclusions and cement paste as matrix. At the nanometer level, the cement paste is considered as a composite with cement particles and crystals as inclusions and C-S-H gel as

77 © 2014 UTA

H. Hamidane and A. Ababneh Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 76-82

matrix. As one can see that a composite model for effective transport properties of a composite material can be used at all three levels, but the components at different scale level will be different.

IV. FINITE ELEMENT FORMULATION

The partial differential equations governing the three transport processes, Eqs. (1), (2), and (3) could be written in a unified format as the following:

h=Div[Dgrad( ) ]

t t t

(4)

where is the main field variable. The correlations among the variables in Eq. (4) and Eqs. (1) to (3) is shown in Table I. Depending on whether δ = 0 or both ξ = 0 and δ = 0, Eq. (4) can be decreased to the same form of Eq. (2) and Eq. (3).

Applying Galerkin weighted residuals method [15, 16] to the Eq. (4):

h-Div[Dgrad( ) ] d 0

t t t

(5)

TABLE I

CORRESPONDENCE BETWEEN THE SIMPLIFIED AND EXPLICIT EXPRESSIONS OF

THE PDE’S GOVERNING THE PROBLEM.

Using Green’s Theorem and discretizing the domain into finite elements, we have:

e

eT

eeT

e

eT

eee

dNDNndDNt

dNN ......

0....

t

dNNt

hdNNN

e

eT

e

eTe

ee

(6)

where denotes the domain boundary and n the outward

normal to the domain surface. N is a row vector containing

the element shape functions associated with each node, and

e is a vector containing the unknown nodal values.

Superscript (e) refers to element values.

If we note by B N , and knowing that the flux

is given by nDJ . (7)

The Eq. (6) can be written as:

A K F (8)

where

TA N . N d

(9)

dNNNhdNNdBBDK TTT .... (10)

dNdNNF TT .. (11)

In order to obtain a numerical solution, Eq. (11) is integrated in time by means of a finite difference approximation [17]:

ttttKtA

ttttt FFtKtA 11 (12)

To solve the system of nonlinear equations resulting from the time discretization of Eq. (12), we used the Modified Newton Raphson method (MNR) [18].

V. NUMERICAL RESULTS AND DISCUSSIONS

To examine the performance of the prediction model and the effect of different model parameters on the chloride penetration in concrete in cold regions, a 10 cm depth concrete slab similar to

the slab shown in figure (1) was used. The finite element mesh was only subject to flux type boundary conditions along the upper side, and the other sides were considered to be sealed (no flux is allowed). The material parameters, for all analyzed cases, are given in the table II.

A. Comparison with Test Data

The concrete slab shown in figure (1) was exposed to a 5% chloride solution and 100% relative humidity on the

Pro

blem

D h

Chl

orid

e P

enet

rati

on

fC

f

t

C

C

Cl

ClD

t

w

0 H 0

ClEnvC HH

EnvH

Moi

stur

e D

iffu

sion

H

H

w

H

HD

0

0

fC EnvH H

Hea

t T

rans

fer

T

C

T

TD 0 0

0 0 EnvT T

78 © 2014 UTA


upper side. The temperature in the surrounding environment was 25˚C. The slab was initially free of chlorides, at 50% relative humidity and 15˚C temperature. The results obtained by the present model, for two types of concrete with water to cement ratios of 0.4 and 0.6, were compared to experimental results from the 90-days ponding test [19]. The concrete slab used for the test is similar to that shown in figure (1) and exposed to similar conditions. Figure (2) shows that the predicted chloride concentration profiles by the present model agree very well with the test data for the partially saturated concrete.

Fig. 1. Concrete slab used in the numerical analysis.

TABLE II

MATERIAL PARAMETERS USED IN THE ANALYSES.

Specific heat, 1170 [J/kg.˚C]

Thermal conductivity, 3.6 [W/m.˚C] Convection heat transfer coefficient,

0.07 [W/m2.˚C]

Surface moisture transfer coefficient,

2.43×10-7 [m/s]

Surface chloride transfer coefficient,

1 [m/s]

Cement type I Curing time 28 days

Fig. 2. Comparison of the numerical results with test data.

B. Effect of Concrete Mix Parameters

To show the sensitivity of the numerical model to concrete mix parameters, i.e., the water to cement ratio and the volume fraction of aggregate, the concrete slab was exposed to the same initial and boundary conditions as described above, and the mix designs for the concrete were varied. Figure (3) shows the profiles of free chloride concentration after one year of exposure for concretes made of 65% aggregate volume fraction and three different water to cement ratios, 0.45, 0.55 and 0.65. The general trend is that the higher the water to cement ratio, the higher the diffusivity of concrete, and thus

the higher the free chloride concentration at a fixed depth. This is because as the water to cement ratio increases, the porosity of the concrete increases, and thus, chloride penetrates in a faster rate. Figure (3) shows exactly the general trend in terms of the effect of water to cement ratio.

The effect of the aggregate volume fraction on the chloride penetration into concrete is shown in figure (4). The water to cement ratio was fixed as 0.55 and different volume fractions of aggregate, 55%, 65% and 75% were used in the model. One can see from figure (4) that the free chloride concentration decreases with increasing the volume fraction of aggregate. This is due to the lower diffusivity of aggregate which reduces the effective diffusivity of concrete.

C. Effect of Temperature

The effect of temperature on chloride penetration was studied by using two examples. In the first example, the seasonal temperature variation was simulated sinusoidally over a period of one year with maximum and minimum annual temperatures of 25˚C and -15˚C; and in the other example, the environmental temperature was kept a constant at 25˚C during the entire year. The external relative humidity and chloride concentration were kept constant at 100% and 5%, respectively. The slab is initially free of chlorides, at 50% relative humidity and 15˚C temperature. The resulting free chloride profiles after a period of exposure of one year are shown in figure (5).

Fig. 3. Effect of water to cement ratio on the chloride penetration into concrete.

Fig. 4. Effect of the aggregate volume fraction on the chloride penetration into concrete.

One can see that at a certain depth, the chloride concentration under the constant temperature is higher than that under the variable temperature at the same depth. This is because the constant temperature used in the second example (25˚C) is the highest temperature in the variable temperature range used in the first example (from 25˚C to -15˚C). Therefore, the temperature effect in the case of variable

79 © 2014 UTA


temperature is either equal or lower than the temperature effect of the case of constant temperature.

D. Effect of Moisture

To study the effect of moisture on the chloride penetration into concrete, three different examples were used. The first one is to simulate drying effect on chloride penetration, the environmental relative humidity was kept constant at the level of 75% and the concrete slab was initially at 100% internal relative humidity, thus simulating a drying condition. The second one is to simulate wetting effect on chloride penetration, the environmental relative humidity was kept constant at the level of 100% and the concrete slab was initially at 50% internal relative humidity, thus simulating a wetting condition. In the third one, the environmental relative humidity was allowed to vary sinusoidally over a period of one year with maximum and minimum annual relative humidities of 100% and 50%, and the concrete slab was initially at 50% internal relative humidity. Boundary temperature and chloride concentration were kept constant at

25 °C and 5%, respectively. The initial internal chloride

concentration is 0%. The resulting chloride profiles after periods of exposure of one month, six months and one year are shown in figures (6), (7), and (8), respectively.

Fig. 5 Effect of temperature on the chloride penetration into concrete.

From the figures, one can see that the moisture diffusion affects greatly the chloride penetration. During the wetting process, the moisture gradient and the chloride concentration gradient are in the same direction. Thus, the diffusion of the water containing chloride ions into concrete increases the rate of chloride penetration into concrete, and subsequently the free chloride concentration profile is the highest among the three examples. During the drying process, the moisture gradient and the chloride concentration gradient are in opposite directions, and the penetration of chloride ions is slowed down. So, the free chloride concentration profile is the lowest among the three examples. In the case where environmental relative humidity varied, the direction of moisture gradient varied and the direction of chloride concentration gradient remained the same, and as a result the free chloride concentration profile is between those for the drying and wetting processes. In general, when a concrete is subject to wetting- drying cycles, the resulting chloride profile depends on the frequency of the cycles (period) and the rate of wetting or drying during the cycles.

Fig. 6 Effect of moisture on the chloride penetration into concrete (one month).

Fig. 7 Effect of moisture on the chloride penetration into concrete (six months).

Fig 8 Effect of moisture on the chloride penetration into concrete (one year).

E. Chloride Concentration in Concrete Structures under Service Conditions

Chloride penetrations into concrete structures under various service conditions were studied by three examples. In the first example, the chloride concentration was a constant at 5%, thus, simulating a concrete slab that is submerged in a chloride solution, sea water for example. In the second example, which simulates a concrete slab exposed to de-icing salts that are applied only in the winter, the chloride concentration was varied according to a step function. In the first two examples, the atmospheric temperature and relative humidity were kept constant at 25 °C and 100%. In the third example, to simulate field conditions, the chloride concentration was allowed to vary according to a step function over a period of one year where CEnv is 5% during winter and zero elsewhere, the relative humidity was allowed to vary sinusoidally over a period of one year with maximum and

80 © 2014 UTA


minimum annual relative humidities of 100% and 50%, and the temperature was allowed to vary sinusoidally over a period of one year with maximum and minimum annual temperatures of 25˚C and -15˚C.

The free chloride profiles resulting from the first two examples after different periods of exposure are shown in figures (9) and (10). When the slab is exposed to chlorides, the free chloride concentration in the region close to surface is the highest, but during periods when chlorides no longer exist in the environment, the free chloride concentration in the surface decreases as chlorides that were in the surface continue to diffuse further inside the concrete as the concentration gradient still exists.

Fig. 9 Free chloride concentration profile for constant chloride concentration in the environment.

Fig. 10 Free chloride concentration profile for variable chloride concentration

in the environment.

Fig. 11 Variation of free chloride concentration for constant chloride concentration in the environment.

M Fig.12 Variation of free chloride concentration for variable chloride

concentration in the environment.

Figures (11) and (12) shows the variation of free chloride concentration with time at different depths after 25 years of exposure, for the first two examples. The fluctuations in the free chloride concentration, as shown in figure (12), are due to changes of the boundary condition of chloride concentration, whether deicing salts are applied or not in the surrounding environment. At a fixed depth in concrete, when there is a supply of chloride ions, the free chloride concentration increases with time, and when there is no chloride ion on the boundary, the free chloride concentration

Fig.13 Free chloride concentration profile for variable chloride concentration in the environment after 25 years of exposure.

decrease with time.

Figure (13) shows a comparison between the free chloride profiles resulting from the second and third examples after 25 years of exposure. At the shallow zone close to the surface, the free chloride concentration resulting from the third example (all environmental conditions, chloride, humidity, and temperature are variables) is higher than that resulting from the second example (only chloride concentration is variable). This is due to the coupling effects of relative humidity and temperature that accelerate the chloride penetration process.

VI. CONCLUSIONS

A general model for chloride penetration into concrete structures was proposed in this study. The mathematical formulation involved the coupling of heat transfer, moisture transport and chloride diffusion mechanisms. The multiscale modeling of material parameters involved in the coupled governing equations resulted in a

81 © 2014 UTA


more realistic characterization of the chloride ion ingress into concrete. The governing equations were solved numerically using the transient finite element method, in which the time integration was performed using a finite difference scheme and the systems of nonlinear equations were solved using the modified Newton Raphson method. Although the computational tool is able to deal with different exposure environments, special emphasis was on the cold environments where chlorides mainly come from the de-icing salts. From the analytical study performed, the model proved to be promising as a prediction tool for chloride profiles and time to depassivation estimates, since it showed sensitivity to almost all relevant parameters involved in the process. Comparison of the model predictions to available test data (90-day ponding test) showed its ability to simulate chloride profiles.

REFERENCES

[1] A. V. Saetta, R. Scotta, R. Vitaliani, "Analysis of chloride diffusion into partially saturated concrete", ACI Mater. J.,90 (5), 441-451, 1993.

[2] R. Frey, T. Balogh, G. L. Balázs, Kinetic method to analyze chloride diffusion in various concretes. Cement. Conc. Res.,24 (5), 863-873, 1994.

[3] S. L. Amey, D. A. Johnson, M. A. Miltenberger, H. Farzam, "Predicting the service life of concrete marine structures: An environmental methodology", ACI Struct. J., 95 (2), 205-214, 1998.

[4] E. J. Hansen, V. E. Saouma, "Numerical simulation of reinforced concrete deterioration-Part 1: Chloride diffusion", ACI Mater. J.,96 (2), 173-180, 1999.

[5] C. Andrade, "Calculation of chloride diffusion coefficients in concrete from ionic migration measurements", Cement. Conc. Res.,23, 724-742, 1993.

[6] S. Chatterji, "Transportation of ions through cement based materials, Part I. Fundamental equations and basic measurement techniques", Cement. Conc. Res., 24 (5), 907-912, 1994.

[7] Y. Xi, Z. P. Bazant, "Modeling chloride penetration in saturated concrete", J. Mater. Civ. Eng.,11 (1), 58-65, 1999.

[8] A. Ababnah, F. Benboudjema, Y. Xi, "Chloride penetration in nonsaturated concrete", J. Mater. Civ. Eng.,15 (2), 183-191, 2003.

[9] A. E. Nakhi, "Damage impact on chloride diffusion through concrete: Experimental, theoretical and numerical studies" PhD thesis, Boulder, CO, USA, University of Colorado, 2004.

[10] X-C. C. Suwito, Y. Xi, "Parallel finite element method for coupled chloride moisture diffusion in concrete", Int. J. Numer. Anal. Model, 3 (4), 481-503, 2006.

[11] B. Martin-Pérez, "Service life modeling of reinforced concrete highway structures exposed to chlorides", PhD thesis, Canada, University of Toronto, 1999.

[12] O. B. Isgor, "A durability model for chloride and carbonation induced steel corrosion in reinforced concrete members", PhD thesis, Ottawa, Canada, Carleton University, 2001.

[13] O. B. Isgor, A. G. Razaqpur, "Predicting the initiation and propagation of corrosion in reinforced concrete structures", 4th Structural Specialty Conference of the Canadian Society of Civil Engineering, Montréal, Québec, Canada, 2003.

[14] A. Ababneh, "The coupled effect of moisture diffusion, chloride penetration and freezing-thawing on concrete durability", PhD thesis, Boulder, CO, USA, University of Colorado, 2002

[15] D. L. Logan, "A First Course in the Finite Element Method", (4th ed), Toronto, Thomson, 2007

[16] O. C. Zienkiewicz, R. L. Taylor, "The Finite Element Method", 1: The Basis. 5th ed, Oxford, Butterworth-Heinemann, 2000.

[17] J. H. Argyris, L. E. Vaz, K. J. William, "Higher Order Methods for Transient Diffusion Analysis", Comput. Methods Appl. Mech. Eng. ,12, 243-278, 1977.

[18] O. C. Zienkiewicz, R. L. Taylor, "The Finite Element Method", 2: Solid Mechanics. 5th ed, Oxford, Butterworth-Heinemann, 2000.

[19] C. Andrade, D. A . Whiting, "Comparison of chloride ion diffusion coefficients derived from concentration gradients and non-steady state accelerated ionic migration", Mater. Struct., 29,476-484, 1996.

82 © 2014 UTA


Abstract—Stone-concrete buildings are the most common residential construction in Jordan. Jordanians use the locally available natural limestone as masonry units in the external facades of buildings. While external walls were built at the dawn of the last century using heavy stone masonry units and different kinds of cementing mortars, this practice changed over the years to combine smaller and thinner stone masonry units with plain concrete. Earthquake awareness in Jordan has increased over the past few years as worldwide earthquakes have caused huge losses in human lives as well as substantial economical losses. Most of these losses were directly correlated to heavy damage or collapse of buildings. Occurrence of earthquake damage depends upon strength, period and duration of seismic motions in addition to the dynamic properties of the structure. The Fundamental Period of vibration of a building is one of the most important factors that determine how the building can respond to ground motion. This study is concerned with the dynamic properties of residential stone-concrete buildings (old versus new) in Jordan. The fundemantal period of vibration will be evaluated experimentally using ambeint vibration measurements (Nakamura technique) and analytically using SAP2000N to analyze models of representative buildings. Ambient vibration measurements are widely used to determine the fundamental period of vibration for buildings using the Nakamura technique also known as H/V ratio. The H/V ratio is defined as the ratio of the intensity of the horizontal waves to the intensity of vertical waves. The H/V spectral ratio is plotted against the frequency, and a clear peak designates the dominant frequency of the building under consideration .Recommendation will be made regarding the fundemantal period of vibration equation in the Jordanian code for earthquake-resistant buildings. KEY WORDS— Residential Building, Code, Vibration, Seismic Design, Stone-concrete

I. INTRODUCTION ordan is af fected b y the seism ic acti vity of t he Dead Sea transform fault, which extends 1000 km from the Red Sea to Turkey. Current studies indicate the probable occurrence

of a major earthquake in the region every 100 years. Most of the bui ldings in J ordan were built w ithout earthquake

resistance, which in creases the incid ence of di sasters in t he event of a major earthquake. Stone-concrete buildings are the most com mon in t he Jordan ian residential co nstruction. Jordanians use the locally av ailable natu ral lim estone as masonry uni ts in t he ex ternal facades of bu ildings. While external walls were built at the dawn of the last century using heavy stone m asonry u nits and d ifferent k inds of c ementing mortars, this pr actice chang ed over the years to co mbine smaller and thinner stone masonry units with plain concrete.

Nowadays, residen tial buildings ty pically in clude plai n concrete as back-fill behi nd “ thin” stone masonry courses as shown in Figure 1a. Acti vated by publ ic awareness of the importance of thermal insulation, the construction practice has also been modified to include polystyrene boards or rock-wool within the ex ternal w alls. Figure 1b shows a ty pical cross section of exterior walls in new local residential buildings. As shown in the figure, cou rses of th in (around 50 mm) sto ne masonry un its ar e back-filled wi th a 170 mm plain concrete layer that i s cast agai nst an other l ayer of con crete bl ocks, 100mm in thick ness. Pol ystyrene bo ards are commonly used between the concrete blocks and plain concrete layer.

(a) Single layer (b) Double layer

Fig. 1. Typical cross section in stone concrete walls

Modeling and Measuring the Fundamental Period of Vibration for Low to Medium Rise

Residential Buildings in Jordan

Musa Resheidat1, Hanan Al Nimry2, Marwa Al Jamal3 1 ,2 Jordan University of Science & Technology, Jordan

Email: [email protected] [email protected] 3 Graduate CE Student, Jordan University of Science & Technology Jordan

Email: [email protected]

J

83 © 2014 UTA

Earthquake awareness in Jordan has increased over the past few years as worldwide earthquakes have caused huge losses in human lives as well as sub stantial economical losses. Most of th ese losses w ere d irectly co rrelated to heavy damage or collapse of buildings. Occu rrence o f e arthquake dam age depends upon strength, period and duration of seismic motions in ad dition t o th e dynam ic propertie s of the structure. The Fundamental Per iod of vib ration o f a bu ilding is on e of the most important f actors that d etermine how the b uilding can respond to ground motion. The Jordanian code for earthquake resistant buildings estimates values f or th e elastic p eriod of vibration ( in seconds) f or any b uilding b y the following formula:

Ta = k (h) 0.75 (1)

Where k depends on the material and type of the structure and h is the height of the building.

National experts in Jordan suggested a value of k = 1.2 5 for stone-concrete buil dings m erely b ased o n en gineering judgment without any experim ental or even an alytical proof . The structures targeted under investigation r epresent m ost commonly po pular co mmercial and/or housing buildings, made of reinforced concrete elements such as beams, columns and s labs in addition to masonry or concrete walls and block panels.

Fig. 2. Types of Low-rise Buildings

A. Objectives Motivated by these conside rations, an ext ensive research

stage was in itiated and wil l be carr ied out to d etermine t he natural frequencies and structure vibrati on m odes. Th e f irst problem, from t he analytical po int of view, to ob tain th e dynamic characteristics, is the postulation of a model without being too complicated that reflects accep tably th e m ain characteristics. Sec ondly the result s o f th e r esearch must be compared w ith the r esults of the m odels to ex amine its components. Thus, th e obj ectives o f th e r esearch co me out. They are:

(a) Determine experimentally, the dynamic characteristics of the considered structures.

(b) Establish the inf luence of the int eraction soil-structure on the dynamic characteristics

(c) Calibrate th e m athematical m odel u sed, to obtain t he structures’ dy namic char acteristics, determining the relative im portance of bo th, the structural and not structural elements on them.

The proposed study is co ncerned w ith the d ynamic properties of residential ston e-concrete buil dings ( old versus new ) in J ordan. The f undemantal perio d o f vibration will be evalu ated experimentally using ambeint vibration m easurements (Nakamura technique) an d analytically using SAP2000N to an alyze models o f representative buildings. Reco mmendation wi ll be m ade, if possible, regarding the f undemantal period of vibration equation in the J ordanian code for earthq uake-resistant buildings.

B. Literature. Review

Amanat and Hogue [1] inv estigated the natural p eriod o f vibration of regular RC frame buildings having infill walls by using 3D finite element models .Compared with the values for the period of vib ration co mputed usin g building codes, good agreement was found in case of models containing infill walls. On t he co ntrary, th e r esults co nfirm very large differences between period of vibration v alue of RC frame buildings without infill walls and those predicted by codes. Crowley and Pinho [2] stu died eleve n cases o f RC frames having infill panels f rom f ive d ifferent European count ries built between 1930 and 19 80 wi th di fferent num ber of stories varying between two to eight stories. The effect of frame height on the period of vibration was analyzed usin g a fiber – m odeling Finite E lements program for seism ic anal ysis of f rame structures. The study presented a period-height relationship for RC frame buildings with infill walls and compared it with the empirical equation of Euro Cod e 8 ( EC8).The an alytical formula values of period of vibrati on s hows a very large difference with values ob tained f rom EC8.The autho rs suggested revising the f ormula of EC8 based on the ir results. Demagh et al. [3] investigated the period of vibration for eight RC f rames and four ste el f rames with different number of stories. The authors derived a f ormula f or the period o f vibration of multi story buildings from the f ree vibration of a cantilever colu mn. The nu mber o f sto ries and the plan dimensions and elem ent stif fness are the major factors effecting the period of vi bration. N umerical anal ysis h ave been p erformed using S AP2000 N prog ram .The v alues of

M. Resheidat et al. Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 83-87

84 © 2014 UTA

period of vibration obtained from the auth ors formula show a good agreement wi th values o btained from t he numerical analysis and codes equation. A seven story reinforced concrete residential building was select ed and anal yzed by Helou and Touqan [4] using nine num erical m odels. The f undamental period of vibration in second s for the building was also evaluated by using IBC 2003 code equation. The result shows that the addi tion of m asonry wall increased the period of vibration; this is due to the increase in mass without increasing stiffness. Kose [5] ev aluated t he f undamental perio d of vibration of 189 RC f rame bui lding considering the effect of presence of i nfill w all. The studi ed bu ildings were selected based on th e par ameters af fecting the f undamental per iod of vibration such as the height of the building, type of the frame, and percentage of infill wal l. 3D finite elem ent m ethod was used to analyze s ome of the f rames. Iter ative modal analysis was used to determine the fundamental period of vibration for buildings co ntaining i nfill wal ls taking into account their nonlinear behavior. The r esults obt ained f rom th e num erical analysis were compared with the code equation. Fundamental period of vibration determined by code equation was less than the fundamental period of vibration calculated from computer modeling. I t was f ound that the h eight of the building is a major parameter affecting the fundamental period of vibration. Masi and Von a [6] investigated th e f undamental period of vibration of reinf orced concrete res idential buildings. The period of vibration w as determ ined usin g exp erimental tests (ambient vibration m easurement) an d n umerical anal ysis (eignvalue analysis) taking into account height of the building, presence of m asonry wal l elem ent stif fness. Per iod-height formula proposed was for some reinforced concrete buildings based on eig envalue anal ysis. The results show very close agreement with experimental measurement and codes formula. Masi et al. [7] investigated the period – height relationship by studying medium rise RC frame buildings. Parametric analysis used to analyze the selected f rame buil ding f rame b y usin g three cases, case one: total ly in filled f rame, case two: frame without in fill wall , case th ree: partially infilled f rame. Micro tremor measurements were also used to det ermine the period of vi bration f or RC f rames under consideration. The results show is a large d ifference between numerical an d experimental val ues. Nav arro et al . [8] verified empirical period formula with the dat a of eigh ty nine pr ivate bui ldings and eleven public building in Granada cit y wi th nu mber of stories r anging f rom three to sixteen obtai ned from t he ambient n oise m easurement. The m easured periods were in agreement with those values obtained from other local studies. It was concluded that the period of vibration increases with the number of stories. Ricci et al . [9] concl uded th at t he fundamental p eriod o f v ibration can be estimated using empirical formula and m odal analysis pr ovided from numerical model of the structure. Period of vibration is a key parameter to ass ess the seismic demand which dep ends o n mass and stiffness of the structure. Modal analysis was used to study reinforced con crete m oment resistant f rames with various morphology and infill ch aracteristics. T hey used t he data obtained from th e analysis to d evelop s implified formulas. Resul ts were c ompared with similar formulae obtained f rom th e literatu re based on experimental measurement of existing bu ildings. The comparison show ed

that there is a good agreement between the numerical analysis and experimental measurement based on ambient noise in case of uncracked infills. The cracked infills lead to overestimation of th e empirical period of vibration of in filled RC bui ldings. Verderame et al. [10] studied two types of Euro-Mediterranean RC bare f rame b uildings. The first t ype designed for gravity load only and the second type designed following the seismic design c riteria. Modal analys is w as used to analyze the structures. The result show that, in the lo ngitudinal directio n the period of vib ration o f the bu ilding desi gned wi th gravity load o nly was f ound to b e hig her t han t hat o f the building designed for seismic loads. In th e s hort directio n the p eriod becomes much larger than the period in the lon gitudinal direction due to dif ferent s tructural systems. The p eriod o f vibration evaluate using Euro code 8 for the studied buildings was found to be long er the period of vibration obtained from modal analysis.

C. Expected Outcome

The findings of this work would enable to suggest emperical equations f or the f undemantal period of vibration for t he residential stone-concrete buildings in Jordan to incorporate in the Jordanian cod e f or earthq uake-resistant buildings if possible. To better quan tify the f undemantal period of vibration for the dominant resident ial construction in J ordan, an experimental and anal ytical investigation is planned as follows: Residential building will b e classif ied based on t he

method of construction (old versus new) and number of stories.

Twenty four buildings will be chosen as representative of th e two m ethods of construct ion an d the various height catego ries (low -rise and m eduim-rise buildings) and verticle regularity.

Analytical models of the representative buildings,using SAP2000N, will be co nstructed an d anal yzed t o determine their fundamental period of vibration.

Ambeint vib ration m easurements based o n t he Nakamura tech nique will be used in coll oboration with the Natu ral Resours es A uthority (NRA) to measure the f undemantal p eriod of vi bration of t he representative buildings.

Analysis an d com parison o f th e experim ental an d analytical results will be performed.

Emperical equations for the fundemantal period of vibration for the res idential stone-con crete b uildings will be suggested to in corporate in t he Jordanian co de for earthquake-resistant buildings if possible.

The following expected outcomes may be outlined from this research:

1. Documentation of the m easurements of n atural fundamental periods o f th e chosen representative buildings

2. Obtaining sim ilar r esults by em ploying an alytical solutions.


85 © 2014 UTA

3. Comparison an d cali bration of r esults aim ing at introducing code f ormulae to assist and facilitate the design of such buildings.

II. METHODS AND TOOLS Dynamic characterizatio n of civ il engi neering structu res becomes increasing ly im portant f or dynamic response prediction, f inite elem ent modal u pdating, structural health monitoring, as well as passive and acti ve vibration control of the low to medium high -rise buildings. The structure can adequately be excit ed by non m easurable ambient, or natural, excitation suc h as m icro-seismic t remors. Ambient v ibration test has two m ajor adv antages co mpared t o f orced v ibration test for obtaining dynamic characteristics of civil engineering structures. On e is that no expensive an d h eavy excitation devices ar e r equired, and, theref ore ease and economic t o implementation. The o ther advan tage i s t hat al l (or p art) of measurement coordina tes can be used as r eferences. The closed spaced or even repeated modes can easily b e handled. Structural sy stem identification based on am bient response measurement has drawn great att ention i n civil engineering community i n recent year s f or extracting dynamic characteristics, such as n atural frequencies, damping ratios [11-13].

A. Potential of Microtremor Measurements

Microtremor methods have lo ng b een used ex tensively i n Japan, and much less in other parts of the world, due to some questions that were not satisfactorily answered. The Nakamura version of th e microtremor methods ha s proved to be one of the m ost c onvenient tech niques to estimate fundamental frequencies of soft deposits. It has been shown by several authors that there is no straigh tforward relati on bet ween t he H/V peak amplitude and site amplification. However, the peak amplitude of the H/V ratio is affected by the characteristics of the unco nsolidated soil s and the und erlying bedrock. The spectral f eatures and p olarization o f m icrotremors exh ibit a gross corr elation wit h the sit e geolo gical condi tion. This question is addressed in by use of one- and two-dimensional numerical modeling in combination with f ield measurements. Synthetics are used to derive relat ions between the amplitude of the H/V ratio. The theoretical results will then be compared with the results from field measurements. The ef fect of two-dimensional structures is to be analyzed and r elated to r esults obtained from a reference site technique.

III. DEVELOPMENT OF THE STRUCTURAL MODEL

A. Structural System. Structural frames are of ten f illed wi th m asonry walls

serving as partiti ons or as claddi ng. In th e structural design process, such f iller walls are considered to be inert “nonstructural” elements. The structure is assumed to carry the transverse loads b y the f rame elements resisting pr imarily in flexure. It is apparent f rom geometrical considerations t hat a reasonably tight fitted wall having finite stiffness will impede deformations com patible with f rame action . The f rame wit h filler wal l is considerably s tronger and stif fer than the frame alone. Ignoring the interaction between the frame and the filler

wall is tantamount to wasti ng a ver y important str uctural contribution. A lso th e cr itical r egions in t he f rame-wall composite may not be the same as those in t he f rame alo ne. And the designer may have a risk on brittle links of the frame-wall composite. There is a general agreem ent am ong researchers that inf illed f rames h ave g reater strengt h as compared to frames without infills. The presence of the infill will also increase the lateral stiffness considerably. Due to the change in stiffness and the m ass, the dynamic characteristics will also cha nge. U nderstanding t he beh avior of infilled frames and having a satisfactory method of analysis will help us to have m ore realistic and econ omical solut ions. In many occurrences of earthqu akes it was show n that , inf ills h ad an important ef fect on t he res istance an d stif fness of bu ildings and hence an effect of dynamic response of the structure.

The behavior of the in filled frame under seismic loading is very complex an d com plicated. Since t he beh avior is nonlinear and closely related to the link between the frame and the infill, it is very difficult to predict it by analytical methods unless these analytical m odels are s upported and revised b y using th e measurements of the exi sting buildings in t erms o f natural periods as w ell as the drift and/or the horizontal floor displacements. Due to the co mplex beh avior of s uch composite structures, this research is of great im portance t o determine the strength, stiffness and dynamic characteristics at each stage of loading. There are several an alytical methods to predict the beh avior, strength and s tiffness of infilled frames. Some of these m ethods ar e empirical or sem i-empirical, an d some are m ore r ational and use sop histicated mathematical models for geometry and materials. These analytical methods can be grouped into two categories; (a) Macroscopic approach, which try to predict the overall behavior and (b) Microscopic approach, modeling mechanical p roperties o f the materials to predict the behavior. Macroscopic models t ry to generate the force-deformation characteristics of the inf ill. They usually idealize the panel by an equ ivalent beam or s trut. Alt hough these methods require less com putational ef fort, they are usually valid only for the tests for which the d erivations are made. Changes in the topology of the panel du e to crack opening and closin g, and chang e in material properties i n macro structu re cannot be taken int o acco unt in simplified model implementation. This is one of the main disadvantages of macroscopic ap proaches. Microscop ic m ethods em ploy principals of mechanics of solids to model the f rame and the infill behav ior. Large computational ef forts are r equired t o obtain meaningful results. The finite element method is widely used for this purpose. Some of the finite element methods are based on the ory of elasticity and some a re more complicated and can take plasticity and str ain hardening into the accounts. New m ethods have also b een d eveloped to model t he nonlinear behavior at conn ections. However s uch m ethods have some difficulties as: Cyclic load behavior cannot be taken into account, since

the material models for such cases do not yield realistic results.

Boundary cond itions and conn ections may no t be modeled properly.

Friction between the frame and the infill may be modeled with reasonable accuracy.


86 © 2014 UTA

Elements are as sumed to be i sotropic, whereas they can be non-isotropic.

The microscopic models ha ve th ese disad vantages, bu t they have th e advan tage o f m odeling the real structu re an d using th ree di mensional m odeling. The new structural concepts ar e based on these types of m ethods, so t he microscopic methods improved by experimental findings will be used more extensively in the future. There are many factors or parameters to be considered: [13, 14] The type of stone walls, the thi ckness and the strength of

backing concrete, The size of the openings in these walls, The existence of block w alls so th ese stone walls an d

backed concrete form sandwich behavior, How these walls are to m odeled either by an equi valent

strut-tie models or by finite element models, Consideration of such walls as non structural elements or

structural elements, Considering 2 di mensional or 3 dimensional frame

analysis Attention should address geometric symmetry as well as

non symmetrical geometry The effect of soft stories The type of construction that will define the behavior of

walls

B. Earthquake records. Figure 3 shows some of the ear thquake records that may be

used to perform the dy namic analys is as a source of excitations. However, replica of these r ecords is li kely t o be implemented. [15, 16]

Northridge earthquake, January 17, 1994

San Fernando earthquake, February 9, 1971

Imperial Valley earthquake (El Centro Record), May 19, 1940

Fig. 3. Earthquake Records

IV. CONCLUSIONS The following conclusions may be drawn from this paper:

1. This paper introduces the action plan for this study

2. Due to time lim itation, measurements o f the n atural fundamental periods of the bu ildings hav e not b een recorded. Per haps d uring the conference t ime, so me records as wel l as some analytical results w ill be available.

3. It is expected the constan t k as g iven i n Equation 1 may have a value ranging from 0.75 -1.50 depending on type of the building under consideration.

4. This approach can easily be applied to medium high or even high rise buildings.

REFERENCES [1] K. Amanat, E. Hogue. "A Rationale for Determining the Natu ral Period

of RC Building Frame Having Inf ll", Engineering Structure 28, 495–502, 2006.

[2] H .Crowley, R. Pinho. "Simplified Equations For Estimating the Period of Vibr ation of Existing B uilding", First European Conference on Earthquake Engineering and Seismology, 3-8 September, 2006.

[3] K. D emagh, H. Chabil, H. Turkia, "Period o f Vibra tions o f F ramed Structures", Materials and Structures 39, 259–267, 2006.

[4] S. Helou, A . Touqan, " Dynamic B ehavior of R einforced Conc rete Structures with Masonry Walls", An-Najah Univ ersity, J. Res. N. Sc., Vol 22, 2008.

[5] M. Kose, "Param eters Affecting the Fundam ental Period of RC Buildings with Infill Walls", Engineering Structures, 31, 93-102, 2009.

[6] A. Masi, M. Vona. "Experimental and Numerical Evaluation of the Fundamental Period of Undamaged and Damaged RC Framed Buildings", Bulletin of Earthquake Engineering, 8, 643–656, 2010.

[7] A. Masi, " Seismic Vulnerability Assessment of Gravity Load Designed RC Frames", Bulletin of Earthquake Engineering, 1 , 371-395, 2003.

[8] M. Navarro, F. Vidal, M. erdear, T. Enomoto, F. Sanchez, I. Matsuda. "Expected Gr ound RC Building Struc tures Resonance Pheno mena in Granada City", 3th World Conference on Earthquake Engineering August 1-6, 3308, 2004.

[9] P. Ricci, G. M. Verderame, G. Manfredi. "Analy tical Inves tigation of Elastic Period of Inf illed R C MRF Buildings", Engineering Structures 33, 308–319, 2011.

[10] G.M. Verderame, I. Iervolino, G. Manfredi. " Elastic Period of Sub-standard Reinforced Concrete Moment Resisting Frame Buildings", Bull Earthquake Eng, 8, 955–972, 2010.

[11] F. Kind, D. Fäh, S. Steimen, F. Salami, and D . Gia rdini, "O n th e Potential of Microtremor Measurements", Proceedings of the 12th World Conference on Earthquake Engineering, Auckland, New Zealand 2000.

[12] Lingmi Zhang ,Yukio Tamura, "Ambient Vibration Testing & Modal Identification of an Of fice Building", Pr oceedings of the 20th IMAC, Los Angeles, USA, Feb. 2-5, 2002.

[13] Jamal Omary, "Behavior of Framed Structures Having Wall Panels" M. Sc. Thesis, Jordan University of Science and Technology, Irbid, Jordan, May 1990.

[14] F. Marjani and U. Ersoy, " Behavior of Brick Inf illed Reinforced Concrete Frames Under Reversed Cyclic Loading" ECAS2002 International Symposium on Structural and Earthquake Engineering, October 14, Middle East Technical University, Ankara, Turkey, 2002

[15] [http://www.iitk.ac.in/nicee/wcee/twelfth_conf_NewZealand/#a4 [16] http://www.clear.rice.edu/elec532/PROJECTS00/earthquake/earthquake

s.htm


87 © 2014 UTA

Abstract—Landslides made the ir appearance with the current of the decade 1990 and the beginning of y ears 2000 on muc h of places of the to wn of Constant ine and its area which extends to north to the M editerranean. These slips appeared after a rainy climate repeated over successive years. Acc ording to the results of analyses on s ites, lit hology, th e microtectonic one, water, the slope of t he r elief, are t he prin cipal cause s. The geolog y of northern of Constantine is cha racterized by the presence of clay and ma rl. Th e com plexity o f th e c lay s oils, t heir capac ity o f saturation, th e deterioration of t he rock s a nd their fast degradation by seepage waters in these zones are as many factors favorable to th e re lease and the acceleration of t he landslides. Geological research and meticulous geotechnics were carried out on th e site s accompanied by s eries b y measurements by inclinometer, and piezometers

Keywords—Landslide, north east Algeria, geotechn ical, c lay, inclinometer

I. INTRODUCTION

North east of Algeria knew l ately o f t he significant

losses caused by the landsl ides (see fig. 1).These losses are in general var ious types of constructions: Transportation roads, works of publ ic works, zones of new or old agglomerations. Movements o f grou nd ar e no t noticed th at i f th ey have one impacts on the company or the economy of the country .These landslides have t he sa me cha racteristics from the point o f view:

- form (opening, depth, rate of travel...); - unst able co nditions ( climate, geol ogy, topog raphy o f the

ground, microtectonic...); - d amage (fissured buil dings made uni nhabitable,

significant dep ression o f r oad section, works communication made impracticable (bridges, viaduct...).

II. GEOLOGIC AND GEOTECHNICAL ANALYSES

A. Geology The geology of the northern of constantine is

characterized by sedimentary formations made up primarily of clay and sandstone of Smendou, with appearance of alternation of clay at least schistous, black color, likings

yellowish end and marly limestone of the miopliocène [1].

Fig. 1. Situation of the zone of slip

They are overthrust folds created by the compression of with the sep aration of the plate Euro- African.Slip of t he sedimentary layers of basins was carr ied out on more 100km north in the s outh by the g one up vo lcanic ones r eleased by underground e nergy [2],[3]. On these bloc ks of sedimentary basins o ne meets layers o f s ame al ternations and the var ious alternations cau sed by the anticlines effects. O ne meets o f them the same types of clays and marls in the borders of the south of Europe [4].

B. Geotechnic On all the sites the depth of the slips is of an average of 25 with 35m this was determined by surveys, tests inclinometric and p iezometric . G eotechnics tests relat e to t he physicomechanical pr operties of the grounds; the latter gave appreciably similar results on all the area, characterized by a very s ignificant s welling of clays and marls at the p oint of internal loss of cohesion of the rock which caused a r upture by shearing. The complexity of absorption of water has a role essential in their behavior and consequently in he behavior of the ground on the sur face and in-dep th [5]. I n a gr ound containing of fine materials (clays and silts), the variations of the wa ter c ontent h ave e ffects o n i ts mechanical pr operties: absorption i nvolves a reduc tion in the shear strength, t he desiccation involves on the c ontrary an improvement of these properties on a small scale and on a large scale of the cracks and cracks more or less profound appear [6].

Behavior Geotechnical and Geologic of the Grounds Northeast Algerian Affected

by the LandslidesA.Saihia and M. Meksaouine

Faculty of Engineering Sciences University Badji Mokhtar Annaba Algeria [email protected], [email protected]

88 © 2014 UTA

While basing onesel f on the index of plasticity and the argillaceous fraction, one can have various behaviors of clays. According to Ske mpton, the value of th e A c (activ ity of clays), whi ch by d efinition is the fraction of the index o f plasticity (I P )(%) wit h t he co ntent o f c lay, ref lects the mineralogical character directly. Thus, we will have: Ac<0,75 i nactive clay, 0,75 <Ac<1,25 no rmal cla y, Ac>1,2 5 clay active) [7]. The results of the LTP Est (Laboratory of the Public works of the EAST) give values for the IP (index of plasticity) ranging between 3 5% and 40% and for th e WL (li mit of l iqidity) between 55% and 60% at the end of October 2005 [8]. Other tests car ried o ut at the en d o f Febr uary 2 006 gave f or like results between 35% and 45% for the IP and between 60% and 75% for th e WL. Of such v alues are cl ose to the l imiting values whi ch ar e abo ut 80% in clays and th e marls o f Constantine a ccording t o t he r epertory of th e Alg erian laboratories o f the gr ounds. ( the gy psum inflates i n the presence of water and becomes lubricating, which will explain the disorders observed on the surface) [9]. In addi tion, of th e shear tests to the t riaxial apparatus o n clays determined angles of repose natural ranging between 3° and 6°; th ey are values clo se t o the rupt ure for the in clined grounds. In the curve of Talbo t, cl ays reach porosity i n 0.84 [10]. I t is n oticed that the indices o f p lasticity are slightly homogeneous where they are definitely higher, which explains the te ndency t o the ground hea ving. The cla y o f mio-pliocene has coe fficients o f c ompressibility vary ing from 0,188 to 0,28 7 [11]. These values c orrespond for the majority of the surveys.

III. ARIOUS SITES OF SLIP This instability of ground is regional it affects 4 departments (Constantine, Mila, Jijel, Bejaia), this problem is shown only i f the condition s necess ary are met ,i t is emphasized by media and economic importance. * In Constantine more than 20% of the perimeter of city which counts more 5000ha * I n Mila department t he motorways of Sétif-Constantine (RN5) road section of Chelghoum Laid on 250m, The viaduct of Oued Dib, abutment rivets southern on the RN 27 in the district o f Grarem (Mila), New ag glomeration of 185 residences on the level of the place head. * In Jij el depar tment, road (RN27) Con stantine- Jijel wit h SEBARI section of 300m . * The town of Bejaia districts of Smina, Sidi Ahmed, Brise de Mer in Bejaia more 600ha.

A. Site of Mila 1 During the realization o f the viaduct o f Ou ed D ib of Mi la department, earthmovings surrounding the Southern abutment Constantine side, appeared, wi th falls invading the trunk road RN 27. It is the e mbankment of access of the southern abut ment; at summer set u p by co mpany (COG E-FAR. Italy) ; who was affected by a system of r uptures sign ificant an d generalized under form opening o f cra cks macroscopic on the in-d epth surface and found by the surveys in depth . The object of this study, is to show the relevance of the use combined of several methods o f investigation, su rface and in -depth h aving

contributed to determine in a precise way the facts of the case [12].

Fig. 2. Representation of the limit of the circle of slip

They are t he results of the investigations of the partner of geotechnical recognition u sing cored survey s, statements piezometric, inclinometric ( see fi g. 3 ) and to pographic, t he taken samples w ere t he sub ject o f ph ysical an d mechanical identification and were u sed for a study of s tability b y estimating the safety coefficient with proposal of solutions of confortement adapted t o the site and the problem arising (see fig. 2).

B. Site of Mila 2 The analysis of the landslide assigns the city 185 residences to Mila and its approximately near, of a population of more than 3000 inhabitants (see fig.4).The causes o f this slip are mainly by the e ffect o f water, the na ture o f the ground and with the slope of the ground. The pr eliminary tests sho wed that the ground is suited to construction, the constraint of the ground is about 1.2 bars and surmounted layers are those of marly clay. After one year of provisional acceptance of work of a phase of 40 residences, it was noted only the appearance of the macro-cracks on the l evel of the walls of a building and widening of the expansion joints (see fig.5) which are propagated towards the foundations. Progressively the phen omenon observed epugnant of other buildings.

C. Site of Mila 3 The layout of the section of motorway Sétif – Constantine), sectio n r oad of Chelghoum Laid on 250m, crosses ge ological formations t o d ominant ar gillaceous, of age mio-pliocene. Surface topography constitutes a slope with relatively weak slope where the relief has a moutonnée surface which co uld reflect pal éo-slips. Our own investigations, carried ou t on these movements o f mass, spread ou t in t ime, enabled us to estimate the width of this geological r isk. The recognitions of g round are distr ibuted over three periods and taking place in at the end of November 2005, at the beginning of January 2006 and end Mars 2007. We noted that these slips are of rational type, planar whose surface of separation is not very deep. T he faces of detachment o f t he various s lips can reach the uneven significant ones and the volume of materials moved can be estimated at several hundred or thousands cubic meters in certain cases.

A.Saihia and M. Meksaouine Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 88-91

89 © 2014 UTA

Fig.. 3. Inclinometric test result in the vicinity of the viaduct of O. Dib

Fig. 4. Slip of the roadway quoted Smina

Fig. 5. Widening of the expansion joint quoted 185 residences

This exp ressway is affected in three places where the roadway is seriously damaged. It proved that work of confortement (works, sheeting piles, gabionnage, etc.)

realized on these neuralgic sites did not have the effect and to hope for stabilization of these slips. In the current state it is difficult to bring a radical solution to this problem, however deepened geological investigations, and with the support of other adapted analyses, could contribute to a stabilization of these catastrophic slips.

D. Site of Jijel The study relates to a vast landslide located on principal road R.N. 27 between Constantine and J ijel, close to the vi llage of Sebari whose ruptu re occu rred wit hin a marly clay f ormation af fecting th e road on a great sectio n. Usin g piezometers one could raise uncertainty on t he localization of the slip surface, we examined various assumptions which took account of morphological features and the geometry of the soil horizons. It was conside red a sing le s urface who se p osition was located only at its two ends:

- a hig h zone bein g locat ed in the princip al nic he of wrenching at altitude 500.00 (NGA). - a lo w zo ne l imited b y t he ap pearance o f th e pad s at altitude 480.00. We also ex tended t he slip surf ace upwards beyond t he principal nic he t o visible cr acks of regres sion until at altitude (510.00) in order to determine with more precision the stabili ty of the whole of the slope. From the profiles thus de fined a nd fixed on th e lo ngitudinal profile b efore slip, it was implemented f our traditional m ethods of analysis of stability: Method of Bischop (circular rupture), method o f Fel lenius (circula r r upture), method o f Ju mbo (noncircular Rupture), method of disturbance (noncircular rupture).

E. Site of Bejaia In this city the slip touched more than 600 ha, they are the districts of Smina, Sidi Ahmed, Brise de Mer.

Fig. 6. Collapse of a construction quoted Smina

Constructions on loose grounds and inclined as it is in the case in the southern side of the city, can lead to disorders of great width. Instabilities were observed pri marily in the faded grounds. The slip of S mina completely destroyed part of the city and threat the stabi lity of the whole district ( see f ig. 6). With least s trong precipi tation, the sli p wor sens endangering human lives from where need for supervising the evolution of the slip. Upstream, one is in t he pres ence o f a rot ational standard slip which is spread out over a surface of more than

A.Saihia and M. Meksaouine Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 88-91

90 © 2014 UTA

600ha. A s the thalweg narr ows, the slip i s transf ormed int o muddy castings. This slip threatens a whole district now.

F. Site of Constantine Classified s ites as u nstable represent an ext ent of approximately 195 ha ( that is to say 3.9% o f the total surface of the ci ty) an d ex pose more t han 1 00 000 i nhabitants, t he equivalent wi th approximately of 15.38 % o f the t otal population of the city. The final delimitation of the unstable grounds by slips is not finalized yet [13], [14]. That means the question of the instability of the grounds with Constantine deser ves a very detailed attention. Beca use the town of Constantine, the metropolis of the East, is on the one hand, a p lace o f ver y s trong de mographic concentrati on. Observations on ground (see fig. 8) allowed to highlight which the dynamics of the grounds is more significant in wet periods than in periods sèches [15]. The presence and the action of the water o f t he rain s c onstitute o nly one of the cau ses of the release and the acceleration of the landslides. Constantine city and l ocated in t he North-eastern area Algerian, episodically knew sign ificant la ndslides ca using o f significant human damage and materials (see fig. 7).

Fig. 7. Slip of the barrier of security of a Constantine road

IV. CONCLUSION The w ork c ompleted i n north eastern of Constantine showed that th is zo ne is into p erpetual movement in wet season confirmed with m easurements of m onitoring (inclinometric piezometric géod ésiques).the microtectonic sup ported instability by a n etwork of active faults. The stu dies geotechnics of all the area appreciably gave the s ame results; the argil laceous grounds o f miopliocene are particularly of average with high plasticity. Some cases of subjects of slips were so lved by systems of supporting on the other hand for the slips of broad wide any solution is not pos sible, a zoning of prohibition to build was installed.

References

[1] P. -E. Coiffait, Un bassin post-nappe dans son cadre structural : l’exemple du bassin de Constantine. Thesis Sci ., University Henri Poincaré, Nancy I, France, 1992.

[2] J. -M. Vila, La chaîne a lpine d'A lgérie ori entale et des c onfins algé ro-tunisiens. Thesis Sci., University Pierre et Mar ie Curie, Paris VI, 1980.

[3] A. Benaissa, Proprié tés géote chniques de quelques foprmations géotechniques propices aux glissements de terrains dans l’agglomerations de Constantine. Rummel 6, pp111-120, 1998.

Fig. 8.Inclinometric test results town of Constantine

[4] S. Paulsen, E. Krauter, J. Hanisch, "Rapport d’expe rtise sur l es glissements de terrain de la ville de Constantine (Algérie)", Institut Fédéra l des Géo sciences et des ressources naturelles Hano vre, N° arch. 117989, 42 , R.F.A. 1999.

[5] P. Mouroux, P. Margon, J. -C. Pinté, "la construction économique sur sols gonflants", BRGM, PP 78-89, 1988.

[6] K. Terzaghi, B. R. Peck, "Soil mechanics in engineering practice" ( 2nd edition), 729, 1967.

[7] N. Mongereau, G. Sanglerat, L. Dav id et H. Millers, "Mouvements d e terrain en zone urbaine:exemple la ville de Lyon", Bulletin de l’AIGI, 31, pp 93-103,Paris, 1985.

[8] Résultats du LNT P Est 2003, 2004, 2005, 2006. (Laboratoire Nationale des travaux publics de l’EST), Constantine.

[9] Costet J. et Sanglerat G., Cours pratiques de mécanique des sols, T1.Dunod, Paris, pp 6- 45, 1981

[10] G. Sanglérat et J. Costet, "Cours pratique des mécaniques des sols", Plasticité et calcul du tassement , T 1. Dunod, p. 285, Paris 1975.

[11] [Philiponnat G., Fondation et ouvrage en terre , Ed .Eyrolles , pp.19-20 .1979

[12] M. P. Luong, Mécanique des sols,CR. Acad. Sc. Paris, pp. 305-307, pp. 313-315, 1978.

[13] Leroueil S ,Tavenas F., Propriétés fondamentales des sols compressibles dans le monde, Symposium international de mécanique des sols, Tiaret, Algérie, 1989 .

[14] A. Saihia, Mesures d es glissements de terrain dans la vi lle de Constantine par la géodésie et l’inclinométrie (Est Algérien) Thèse de doctorat, Université de Annaba, Algérie, 2007.

[15] A. Benaissa, Glissements de terrain, Université de Constantine, p. 102, 2003.

-70.0m

-60.0m

-50.0m

-40.0m

-30.0m

-20.0m

-10.0m

0.0

-5.0 0.0 5.0 10.0 15.0 20. 0 25.0 30.0 35.0 displacement ( cm.)

5/2003

7/2003

9/2003

11/2003

01/2004

03/2004

05/2004

07/2004

09/2004

11/2004

01/2005

Dates

A.Saihia and M. Meksaouine

Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 88-91

91 © 2014 UTA

Leader Election in mobile ad hoc networks using Omega failures detector

Leila Melit Department of Computer Science, Jijel University,

Algeria. [email protected]

Nadjib Badache Laboratory of Computer Systems, USTHB,

Algiers, Algeria. [email protected]

Abstract—Leader election is a fundamental control problem in both wired and wireless systems. The classical statement of the leader election problem is to eventually elect a unique leader from a fixed set of nodes. However, in the context of mobile ad hoc networks, complications may arise due to frequent and unpredictable topological changes. This paper presents a leader election algorithm based on the omega failures detector, along with proofs of correctness. This algorithm ensures that every connected component of the mobile ad hoc network will eventually select a unique leader, which is the node of the highest priority value from among the nodes in that component. The algorithm is well-suited for use in mobile ad hoc networks because it can tolerate arbitrary and concurrent topological changes induced by node mobility, network partitioning and merging components.

Keywords- leader election; mobile ad hoc networks; omega failures detector.

I. INTRODUCTION

A mobile ad hoc network or MANET (Mobile Ad hoc NETwork) is a collection of mobile nodes that can communicate via message passing over wireless links. Nodes that are in transmission range of each other can communicate directly otherwise, they communicate via message relay. Characteristics that distinguish ad hoc networks from existing distributed networks include concurrent and unpredictable topology changes due to arbitrary mobility pattern of nodes, dynamic wireless link formation and removal, network partitioning and disconnections, limited bandwidth and energy, and highly variable message delay. These characteristics signify mobile ad hoc network as a challenging domain for implementing distributed algorithms.

Leader election is a fundamental control problem in both wired and wireless systems. The classical statement of the leader election problem is to eventually elect a unique leader from a fixed set of nodes. Indeed, several algorithms have been proposed to solve this problem. However, in the context of mobile ad hoc networks, though, link changes are common and may cause the network to split into multiple connected components or cause some components to become separated from the leader. Additionally, two connected components, each with its own leader, may merge. It is important to realize that it is impossible to guarantee a unique leader at all times. For example, when a network partition occurs or when two

components merge, it will take some time for a new leader to be elected. Thus, the definition of the leader election problem has to be adapted to the mobile ad hoc environment. Thus, we define the leader election problem in MANETs as the problem of guaranteeing that “every connected component of the mobile ad hoc network will eventually have exactly one leader”.

Moreover, in many situations, it may be desirable to elect a leader with some system-related characteristic as mentioned in [1]. For example, in a mobile ad hoc network it might be desirable to elect the node with maximum remaining battery life or computation capabilities to other nodes, as the leader. Therefore, the elected leader should be the node which has the highest priority value from among all nodes within that connected component, where the priority value of a node is a performance–related characteristic. Thus, the requirements for leader election algorithm becomes: “Given a network of mobile nodes each with a priority value, after a finite number of topological changes, every connected component will eventually select a unique leader, which is the node of highest priority value from among the nodes in that component”.

The solution proposed in this paper to solve election problem in mobile ad hoc networks is different from known solutions in the literature. It is based on the Ω failures detector [2]. The concept of unreliable failure detector was introduced by Chandra and Toueg as a mechanism that provides (possibly incorrect) information about process failures [3]. Each process has access to a local failure detector module. The output of the failure detector module of Ω at a process p is a single process, say q. We say that p trusts q to be up at time t. Ω ensures that eventually all correct processes trusts the same process and that this process is correct. The contribution of this paper is to use the Ω failure detector to solve election problem in mobile ad hoc networks. Otherwise, we will implement omega in the context of mobile ad hoc network. To elect a unique leader, the algorithm requires mobile nodes to communicate only with their neighbors.

The rest of this paper is organized as follows: the next section focuses on some leader election protocols developed for mobile ad hoc networks; Section 3 describes the system model; the leader election algorithm for mobile ad hoc networks is depicted in Section 4; Section 5 presents proofs of correctness; and Section 6 concludes this paper.

92 © 2014 UTA

II. RELATED WORKS Leader election is an extensively studied problem for

networks with static topologies. But, few algorithms exist to implement leader election for mobile ad hoc networks. The algorithms presented in [4], which are classified as Non-Compulsory and Compulsory protocols, are unrealistic as they require nodes to meet and exchange information in order to elect a leader. The algorithms presented in [5], [6] and [7] are based on a routing algorithm called TORA [8] wherein nodes adjust a locally maintained variable, called the height, to point to the leader, in a decrementing manner over a Directed Acyclic Graph (DAG). The work in [7] improves on the idea of [6] by devising an algorithm that is self-stabilizing with relatively fast convergence. The algorithm can tolerate multiple concurrent topological changes. By introducing the time-interval-based computation concept, the algorithm ensures that a network partition can within a finite time converge to a legitimate state even if topological changes occur during the convergence time. Vasudevan, Kurose and Towsley presented in [1] another leader election algorithm based on the classical termination-detection algorithm for diffusing computations by Dijkstra and Scholten [9]. This algorithm always requires each and every node to maintain information about its dynamic neighbourhood. The leader sends periodic heartbeat messages to other nodes. The absence of these messages at a node for some predefined time out indicates a departure from the leader and triggers a diffusing computation at that node to elect a new leader. The diffusing computation consists of constructing a spanning tree with the node that started off the diffusing computation as the root. The root then informs all the reachable nodes about the identity of the elected leader. Different diffusing computations can be executed concurrently. A total order on these diffusions is defined to determine the diffusing computation of the highest priority. A node participates in only one diffusing computation at a time. It stops its participation in diffusing computations of lower priorities in favor of the highest one. The algorithm in [10] presents a consensus–based leader election algorithm. The algorithm elects a local extrema as leader. Moreover, this algorithm can be tuned to the global extrema of the network visiting all nodes instead of majority of those.

III. SYSTEM MODEL We consider a distributed system composed of a finite set

of n> 1 processes Π = p1... pn, that communicate only by sending and receiving messages over a wireless network. The network affected by topological changes is modeled as a dynamically changing, not necessarily connected, undirected graph. Node (process) mobility may result in arbitrary topology changes including network partitioning and merging. Furthermore, nodes can crash arbitrarily at any time and can recover from crash–failure again at any time.

Our model considers various types of links, all of which satisfy the following property:

• Integrity: Process q receives a message m from process p at most once, and only if p previously sent m to q.

• Fair lossy: if p sends an infinitely many messages to q and q is correct, then q receives infinitely many messages from P.

• Eventual timely link: The link from p to q is eventual timely if there is a time t and a bound d such that each message sent by p to q at any time t’ is received by q by time t’ +d.

A. Specification of Ω The concept of unreliable failure detector was introduced

by Chandra and Toueg as a mechanism that provides (possibly incorrect) information about process failures [3]. Each process has access to a local failure detector module. Each local module monitors a subset of the processes in the system, and maintains a list of those that it currently suspects to have crashed. Note that at any given time the failure detector modules at two different processes may have different lists of suspects.

One failure detector of particular interest is Ω [2]. At every process p, and at each time t, the output of the failure detector module of Ω at a process p is a single process, say q. We say that p trusts q to be up at time t. Ω ensures that eventually all correct processes trusts the same process and that this process is correct. Thus, a failure detector Ω can be seen as an algorithm for electing a leader: the process trusted by all correct processes is elected.

The contribution of this paper is to use the Ω failure detector to solve election problem in mobile ad hoc networks. So, each correct process p of our system must have a local variable leaderp that holds the identifier of a single process. Eventually, all correct processes of the same component should retain the identifier of a single correct process. Specifically, we need the following property:

For every component C, there is a correct process l in C and a time t after which for each process p in C, leaderp = l.

If at time t, leaderp contains the same process l for all alive processes p, then we say that l is leader at time t. Note that a process p never knows if leaderp is really the leader at time t, or not. A process only knows that eventually leader p is leader.

IV. LEADER ELECTION ALGORITHM In the proposed leader election algorithm, each process

starts the execution by electing itself and sends the identifier of its leader to all processes in the system. A competition will take place between the different leaders. The winner will be the process that has the highest priority value.

Each process p starts the execution by electing itself. During its execution, it checks if it is still the leader (leaderp = p). If so, it broadcasts a message (ALIVE, p, priorityp) to all every δ time. Each process p, receiving a message (ALIVE, q, pri) tests whether q is higher priority than its leader. If so, p updates the identifier and the priority value of its leader and then broadcasts (ALIVE, q, pri).

Figure 1 shows the proposed algorithm to solve the election problem in mobile ad hoc networks. Each process begins

L. Melit and N. Badache Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 92-95

93 © 2014 UTA

execution of the algorithm by election as a leader then starts its timer (a variable that is automatically incremented at each clock tick).

Figure 1. Leader election algorithm using Ω.

V. CORRECTNESS

We assume that T is the moment when the topology is static, ie, we suppose we have a finite number of topology changes.

We also assume that Pmax is the process that has the highest priority value in its component.

The following theorems establish the correctness of the leader election algorithm proposed above.

Lemma 1. There is a time after which Pmax permanently satisfies that leaderPmax = Pmax and sends a message (Alive, Pmax, prioritypmax) every δ time.

Proof:

Lemma 1 means that after time t> T, Pmax will not receive any message (Alive, i, priorityi) such as (priorityi > prioritypmax) or ([priorityPmax = priorityi] and [i < Pmax]). Therefore, it will not execute the lines (9-14) of the algorithm, and the property leaderPmax = Pmax is always satisfied at every time t> T. To see that it is satisfied, we distinguish the following cases:

Case 1: There was no process i in the Pmax’s component such as (priorityi > prioritypmax) or ([priorityPmax = priorityi] and [i < Pmax]). In this case, lines (9-14) of Task 1 had never been executed by the process Pmax. LeaderPmax = Pmax is always satisfied and will always be satisfied at every time t> T.

Case 2: There was at least one process i in the component such that (priorityi > prioritypmax) or ([priorityPmax = priorityi] and [i < Pmax]) which sent (Alive, i, priorityi) by executing line 8 of the algorithm. In this case, leaderPmax = k has been satisfied after execution of Line 11. Then, after a finite number of topology changes, the process i (and definitely all processes that are higher priority than Pmax) left the component or crashed. Therefore, Pmax becomes the process that has the highest priority value in its component. So, after a time t> T, line 14 will not be executed, the timer will not be restarted and it will finally expired (line 15) and leaderPmax = Pmax will be satisfied (line 16).

Finally, according to task 0, and as leaderPmax = Pmax, Pmax sends permanently (Alive, Pmax, prioritypmax) every δ time.

Lemma 2. There is a time after which every message (Alive, p, priorityp) with p ≠ Pmax disappears from the system.

Proof:

• The links are eventually timely: any sent message is delivered at most T + ∆.

• Note that initially, each process begins by being elected as the leader. i.e., leaderp = p (line 1). As long as it remains leader, it broadcasts (Alive, p, priorityp) (line 8). Also note that each process p receives a message (Alive, q, priorityq), makes leaderp = q if q is higher priority than p (lines 9-12).

At a time t> T, Pmax is the process that has the highest priority value in its component and it periodically sends the message (Alive, Pmax, prioritypmax) (Lemma 1). Each process p ≠ Pmax, upon receipt of (Alive, Pmax, prioritypmax), makes leaderp = Pmax (line 11) and sends (Alive, Pmax, prioritypmax) to all its neighbours as shown on line 13 of the algorithm.

After time t> T, timerp is sufficiently large that the process p can receive the message (Alive, Pmax, prioritypmax) before its expiration. The timer will not be expired as Pmax periodically sends the message (Alive, Pmax, prioritypmax), and p will never execute the line 18 of the algorithm (the message (Alive, p, priorityp) will never be sent).

Finally, every message (Alive, p, priorityp) with p ≠ Pmax was disappeared from the system.

Theorem. There is a time after which each process p in the same component have leaderPmax = Pmax, where Pmax is the process that has the highest priority value in that component.

Proof:

Lemma 1 shows that there is a moment after which the process Pmax maintains leaderPmax = Pmax (ie it is the leader of its component) and periodically sends the message (Alive, Pmax, prioritypmax) to notify the other processes in its component that it is the leader. Lemma 2 shows that no other message will circulate into the component. So the only message that is circulating is the message (Alive, Pmax, prioritypmax) and all processes of the component have Pmax as a leader.

So, the algorithm ensures that:

Code for each process p: Init: (1) leaderp ← p; (2) priorityleader ← priorityp; (3) timeoutp ← δ; (4) Set timerp to timeoutp; (5) Start tasks 0 and task1; Task 0: (6) loop forever (7) if [leaderp = p] and [have not sent (Alive, leaderp, priorityleader) within δ] then (8) broadcast (Alive, leaderp, priorityleader); Task 1: (9) upon reception of message (ALIVE, q, pr) do (10) if (priorityleader < pri) or [(priorityleader = pri) and

(q ≤ leader)] then (11) leaderp ← q; (12) priorityleader ← pri; (13) broadcast (Alive, q, pri); (14) reset timerp to timeoutp; (15) upon expiration of timer do (16) leaderp ← p; (17) priorityleader ← priorityp; (18) timeoutp ← timeoutp + 1;


94 © 2014 UTA

For each component C in ad hoc network, there is a correct process ℓ in C and a moment after which, for every process p in C, leader p = ℓ.

VI. CONCLUSION

In this paper, we have proposed a leader election algorithm that is well-suited for use in mobile ad hoc networks in the sense that it can tolerate intermittent failures, such as link failures, sudden crash or recovery of mobile nodes, network partitions, and merging of connected network components associated with ad hoc networks. We have also proved that our algorithm ensures that eventually each connected component of the topology graph has exactly one leader which has the highest priority value from among all nodes within that connected component. But this algorithm is not communication-efficient. We are currently working on how we can modify this algorithm to make it communication-efficient without any compromise with impulsive behaviors of ad hoc networks.

REFERENCES

[1] S. Vasudevan, J. Kurose, and D. Towsley, “Design and analysis of aleader election algorithm for mobile ad hoc networks”, Proceedings of the 12th IEEE International Conference on Network Protocols(ICNP’04), pp. 350-360, 2004.

[2] T. D. Chandra, V. Hadzilacos, and S. Toueg, “The weakest failuredetector for solving consensus”, Journal of the ACM, 43, pp. 685–722,1996.

[3] T. D. Chandra and S. Toueg, “Unreliable failure detectors for reliabledistributed systems”, Journal of ACM, 43, pp. 225–267, 1996.

[4] K-P. Hatzis, G-P. Pentaris, P-G. Spirakis, V-T. Tampakas, and R. B.Tan, “Fundamental control algorithms in mobile networks”, Proceedingof the 11th Annual ACM Symposium on Parallel Algorithms and Architectures, pp. 251–260, 1999.

[5] N. Malpani, J-L. Welch, and N. Vaidya, “Leader election algorithms formobile ad hoc networks”, Proceedings of the 4th InternationalWorkshop on Discrete Algorithms and Methods for Mobile Computingand Communications, pp. 96–103, 2000.

[6] A. Velayutham and S. Chaudhuri, “Analysis of a leader electionalgorithm for mobile ad hoc networks”, Technical report, Iowa StateUniversity, 2003.

[7] A. Derhab and N. Badache, “A self-stabilizing leader election algorithmin highly dynamic ad hoc mobile networks”, IEEE Transactions on Parallel and Distributed Systems, vol. 19, 7, pp. 926–939, 2008.

[8] V. D. Park and M. S. Corson, “A highly adaptive distributed routingalgorithm for mobile wireless networks”, Proceeding of the 16th IEEE INFOCOM, pp. 1405–1413, 1997.

[9] E.W. Dijkstra and C.S. Scholten, “Termination detection fordiffusing computations”, In Information Processing Letters, pp. 1-4, 1980.

[10] S.M. Masum, Amin Ahsan Ali, and Mohammad Touhid youl IslamBhuiyan, “Asynchronous leader election in mobile ad hoc networks”.Proceedings of the 20th International Conference on Advanced Information Networking and Applications, pp.827–831, 2006.


95 © 2014 UTA

Forecasting Crude Oil Price Based on Artificial Intelligent Model: A Smoothed Feedforward Neural Network

Manel HAMDI International Finance Group Tunisia, Faculty of

Management and Economic Sciences of Tunis, Tunisia, El Manar University, Tunis cedex, C.P. 2092, El Manar

Tunisia Email: [email protected]

Chaker ALOUI International Finance Group Tunisia, Faculty of

Management and Economic Sciences of Tunis, Tunisia, El Manar University, Tunis cedex, C.P. 2092, El Manar

Tunisia Email: [email protected]

Abstract— This paper develops a Feedforward Neural Network (FNN) model to predict the next day crude oil price. Based on historical observations of West Texas Intermediate (WTI) crude oil spot price (4 January 2000 to 1 December 2011) the FNN has been trained. Moreover, a smoothing procedure was applied to the oil price sequence in order to reduce the unforessen disturbances of oil market. The results prove that the smoothing procedure improve the forecasting accuracy therefore it’s not sufficient. Then, we re-smoothed the oil price sequence and repeated this phenomenon until the error stops minimized. The good prediction results giving reliability to re- smoothing the oil prices time series for forecasting tasks.

Keywords- forecasting, crude oil price, smoothed feedforward neural network

I. INTRODUCTION

As th e m ost im portant co mmodities arou nd t he g lobe, crud e oil p lay’s an increasingly: significant role in world economy. Therefore, it is o f significance to develop stronger techniques to predict oil price.

Formely, trad itional statistical and ec onometric methods are widely applied for forecasting cr ude oil price. Abramason and Finizza [1] used probabilistic models to pre dict oil price. For the sam e pupos e, Gulen [2] ap plied t he co -integration analysis. In an other research , Moran a [3] u tilized GAR CH model to short-term oil pri ce forecasting. Ye et al. [4] proposed a simple econometric model for forecasting crude oil spot price. I n next work, La nza et al . [5] developed a nother econometric tool as th e error correction m odels to predict o il price.

However, t hese t echniques failed t o offer go od prediction results du e to th e lin ear asp ect of crude oil market. To avercome th e d rawback, th e artificial in telligent (AI) models are predestinated to sol ve forecasting problem s. The most popular AI t ools to predict o il p rice m ovement is arti ficial neural network (ANN).

Moshiri a nd Foroutan [6] developed l inear an d n onlinear techniques to predict crude oil futures prices. M ore precisely, they compared ARMA and GARCH, to ANN. The empirical results sho wed th at ANN outperforms th e two other models. In a relate d research, Xie e t al. [7] com pared AR IMA and backpropagation n eural n etwork ( BPNN), to sup port vecto r machine (SVM). T he authors concluded that SVM fore caster is supe rior, however, the BPNN outperformed SVM and ARIMA for two sub-periods of all four sub-periods checked.

In light of these findings, an important number of studies have been focused on using ANN in oil price forecasting tasks.

Haidar et al. [8] proposed a three layer backpropagation FNN for short-term forecasting of th e crude oil spot price. In another work, Lack es et al . [9] ap plied a BF m ultilayer perceptron to predict price development of crude oil in sh ort-term, mid-term and long-t erm. Recently, se veral hy brid models have been introduced to p redict o il p rice ( [10]; [11]; [12]; [13]). In the presented study, we apply the FNN to pred ict the future crude oi l p rice on t he one hand a nd on t he ot her hand we introduce a smoothing algorithm in o rder to reduce the short term n oises effect wh ile maintaining the main an d long t erm characteristics of the dynamic oil market [13]. Corresponding to th e au thors, t he sm oothing pro cedure im proves the forecasting results.

The idea in this paper i s to re smoothing the sequence of oi l price already smoothed and repeat th is phenomenon until the error stops m inimized, the refore, reach t he best prediction results.

This paper is p receded as follows. Sectio n 2 presen ts th e experimental st udy m ore precisely; we des cribe data used t o predict fut ure oil price and the forecasting m odel applied. Then, we present the em pirical findings. Later, we conclude in section 3.

96 © 2014 UTA

II. EXPERIMENTAL STUDY

1. Data description

A sequence of WTI crude oil spot price for the period running from 4nd January 2000 to 1nd december 2011 is used to predict future o il price. Th e d aily data (29 90 data) was provided by US E nergy Information a dministration website. 80% of the data set (2392 observations) represent the training sample th at is u sed t o estimate th e parameters o f n etwork (synaptic wei ghts an d bi as) a nd t he rem ainder (598 observations) is the checking sample which is designed to test the predictive ability of network.

2. FNN model : presentation and structure As a universal app roximator us ually use d f or t ime seri es forecasting the FNN m odel is applied in this investiga tion ([14]; [15]). A standard FNN consists of an input layer, one or more hi dden l ayers an d a n output l ayer i nterconnected (see Fig. 1).

Fig.1. A standard FNN architecture

Nevertheless, there are no criteria for the choice of the optimal neural network architecture. A number of experiments were required to i dentify th e id eal n etwork con figuration due to absence of universal heuristics for an optimal architecture [9].

After se veral experiments, t he network c onfiguration ap plied in this application is as fo llow, the daily WTI spot price used as input variables of the model, one hidden layer with 9 nodes and a sing le neuron i n th e ou tput layer that rep resent th e predicted future price. We note that the horizon of prediction adopted in th is wo rk is one d ay ah ead, th e fact th at th e volatility of oil prices is nearly random and mainly related to the day to day events [16].

3. Simulation and results

All along this application, we used a FNN with sigmoid transfer fu nction in th e h idden layer and hyperbolic tang ent transfer function i n t he output l ayer. Also, t wo ev aluation criteria are employed as the mean squared error (MSE) and the root m ean squ ared e rror (R MSE) t o e valuate t he prediction performance. After t raining th e FNN, th e sim ulation results sh ow deficiencies i n te rms of prediction a s MSE= 3.389 and RMSE= 1.7433 . We note t hat th is finding d emonstrate a higher er ror th erefore a b ad pr ediction re sults. In order to improve the forecasting acc uracy, we a pplied the sm oothing procedure to oil price ti me series. As resu lts, th e MSE h as increased f rom 3.38 9 t o 1.5965 a nd R MSE from 1. 7433 t o 1.2635. This result confirmed the conclusion of [12]. Fig. 2 represents the actual and smoothed WTI crude oil prices and t he er ror between t hem i s sho wn i n Fig.3 w hile Fi g. 4 shows the ability of smoothing procedure to reduce the effect of significant oil market disturbances.

0 500 1000 1500 2000 2500 30000

50

100

150

Trading day

WT

I pr

ice

(US

$ pe

r ba

rrel

)

Smoothed oil price

Actual oil price

Fig.2. Actual and smoothed WTI crude oil prices

M. Hamdi and C. Aloui Journal of Advanced Sciences & Applied Engineering Vol. 01, N° 01 (2014) 96-100

97 © 2014 UTA

0 500 1000 1500 2000 2500 3000-15

-10

-5

0

5

10

Trading day

Error

Fig.3. The error between Actual and smoothed WTI crude oil

prices

Fig.4. Significant reduce of oil market disturbances effects

using smoothing algorithm

The new idea in this paper is to re-smooth the sequence of oil price already smoothed and repeat th is phenomenon until the error stops minimized. In Table 1, we summarize the results found.

MSE RMSE

FNN 3.389 1. 7433

SFNN* 1.5965 1. 2635

RE SFNN 1.0078 1.0038

RE-RE SFNN 1.6331 1. 2779

*Smoothing FNN Table 1. Comparison of evaluation criteria values between the

different models

Based on t hese resu lts, we g ive reliability to re-smoothing t he cru de oil seq uence al ready smoothed the fact that MSE r each 1 .0078 and RMSE= 1.0038. Nevertheless, smoothing t he oi l price re-smoothing p rovides hi gher e rror (MSE=1.6331) therefore bad prediction results. In this case we must stop sm oothing crude oil spot price sequence because it becomes unnecessary and the best fo recasting res ults was obtained with the Re SFNN (see Fig.5).

0 100 200 300 400 500 60060

70

80

90

100

110

120

Trading day

WTI price (U

S$ pe

r ba

rrel)

predicted WTI price

actual WTI price

(1)

0 100 200 300 400 500 60065

70

75

80

85

90

95

100

105

110

115

Trading day

WTI price

predicted WTI price

actual WTI price

(2)

0 100 200 300 400 500 60065

70

75

80

85

90

95

100

105

110

115Actual and predicted WTI prices

Trading day

WTI price

predicted WTI price

actual WTI price

(3)

Fig.5. Predicted and actual WTI crude oil spot price in testing

part: (1) FNN, (2) SFNN, (3) Re SFNN

III. CONCLUSION

This stud y pro poses a FNN model to pred ict th e n ext working day crude oil pri ce an d sh ows t hat t he sm oothing algorithm significantly reduces the effect of disturbances of oil market with keeping their dynamism. Based on the em pirical forecasting results, the smoothing procedure must be st opped when the error stopped minimized therefore obtained the best forecasting results. In light of this conclusion, the future work should be focused on looking for methods and algorithms that can reduce the energy market disturbances.


98 © 2014 UTA

IV. REFERENCES

[1] B. Abra mson, A. Finizza, ‘‘Probabilistic forecasts f rom probabilistic models: A case study in the oil market’’, International Journal of Forecasting 11, 63-72, 1995.

[2] G. S. Gulen, “ Efficiency in the Crude Oil Futures Markets”, Journal of Energy Finance and Development 3, 13–21, 1998.

[3] C. M orana, ‘ ‘A se miparametric appr oach to shor t-term o il pr ice forecasting’’, Energy Economics 23, 325-338, 2001.

[4] M. Ye, J. Zyren, J. Shore, “Forecasting Crude Oil Spot Price Using OECD Petroleum Inventory Levels”, International Advances in Economic Research 8, 324–334, 2002.

[5] A. Lanza , M . Mane ra, M. Giovannini, “Modelling and Forecasting Cointegrated Relationships Am ong Heavy Oil and Pr oduct Pr ices” Energy Economics 27, 831-848, 2005.

[6] S. Moshiri, F. Foroutan, ‘‘Forecasting nonlinear crude oil futures prices’’, Energy Journal 27, 81-95, 2005.

[7] W. Xie, L. Yu, S. Wang, ‘‘A ne w method for crude oil price forecasting based on sup port vector machines’’, Lectures Notes in Computer Science 3994, 441-451, 2006.

[8] I. Haidar, S . Kulkarni, H . P an, “ Forecasting model f or crude oil prices based on ar tificial neur al networ k”, in ISSNIP ’08: Proceedings of the international conference on Intelligent Sensors, Sensor Networks and Information Processing, 273-277, 2008.

[9] R. Lackes , C . Borger mann, M. Di rkmorfeld, ‘ ‘Forecasting the p rice development of crude oil with artificial neural networks’’, Lectures Notes in Computer Science 5518, 248-255, 2009.

[10] S. Wang, L. Yu, K. K. Lai, ‘‘Crude oil price f orecasting with T EI@I methodology’’, Journal of Systems Science and Complexity 18, 145- 166, 2005.

[11] L. Yu, K. K. Lai, S. Wang, K. He, “Oil price forecasting with an EMD-based m ultiscale neural network l earning paradig m”, Lectures Notes in Computer Science 4489, 925-932, 2007.

[12] J. Liu, Y. Bai, B. Li, “A new approach to f orecast crude oil price based on fuzzy neur al networ k”, in FSKD ’07: Proceedings of the Fourth International Conference on Fuzzy Systems and Knowledge Discovery, 273-277, 2007.

[13] A. Ghaffari, S. Zare, ‘‘A novel algorithm for prediction of crude oil price variation based on soft computing’’, Energy Economics 31, 531-536, 2009.

[14] K. Hornik, M. Stinchco mbe, H. White, ‘‘Multi layer feedforward s are universal approximators’’, Neural Networks 2, 359-366, 1989.

[15] Z . Tang, P. A. Fishwick, ‘‘Feedforward neural networks as models for time series forecasting’’, ORSA Journal on Computing 5, 374-385, 1993.

[16] F. Gori , D. Ludovisi, P. F. Ce rritelli, “ Forecast of oil pric e and consumption in th e shor t ter m under th ree scenar ios: Parabolic, linear and chaotic behavior”, Energy, vol. 32, pp. 1291-1296, 2007.


99 © 2014 UTA

Date post:	22-May-2018
Category:	Documents
Upload:	nguyenliem
View:	288 times
Download:	8 times

Journal of Advanced Sciences Applied Engineering - … · · 2016-07-03Journal of Advanced...

Documents