ARTICLE IN PRESS

Energy Policy 37 (2009) 1239–1248

Contents lists available at ScienceDirect

Energy Policy

0301-42

doi:10.1

� Corr

E-m

journal homepage: www.elsevier.com/locate/enpol

A fuzzy inference model for short-term load forecasting

Rustum Mamlook a,�, Omar Badran b, Emad Abdulhadi b

a Middle East University for Graduate Studies, Faculty of Information Technology, Amman 11942, Jordanb Al-Balqa Applied University, Faculty of Engineering Technology, Amman, Jordan

a r t i c l e i n f o

Article history:

Received 20 March 2008

Accepted 30 October 2008Available online 7 January 2009

Keywords:

Fuzzy sets

Power generation

Short-term load forecasting

15/$ - see front matter & 2008 Elsevier Ltd. A

016/j.enpol.2008.10.051

esponding author.

ail address: rstmamlk@hotmail.com (R. Maml

a b s t r a c t

This paper is concerned with the short-term load forecasting (STLF) in power system operations. It

provides load prediction for generation scheduling and unit commitment decisions, and therefore

precise load forecasting plays an important role in reducing the generation cost and the spinning

reserve capacity. Short-term electricity demand forecasting (i.e., the prediction of hourly loads

(demand)) is one of the most important tools by which an electric utility/company plans, dispatches the

loading of generating units in order to meet system demand. The accuracy of the dispatching system,

which is derived from the accuracy of the forecasting algorithm used, will determine the economics of

the operation of the power system. The inaccuracy or large error in the forecast simply means that load

matching is not optimized and consequently the generation and transmission systems are not being

operated in an efficient manner. In the present study, a proposed methodology has been introduced to

decrease the forecasted error and the processing time by using fuzzy logic controller on an hourly base.

Therefore, it predicts the effect of different conditional parameters (i.e., weather, time, historical data,

and random disturbances) on load forecasting in terms of fuzzy sets during the generation process.

These parameters are chosen with respect to their priority and importance. The forecasted values

obtained by fuzzy method were compared with the conventionally forecasted ones. The results showed

that the STLF of the fuzzy implementation have more accuracy and better outcomes.

& 2008 Elsevier Ltd. All rights reserved.

1. Introduction

Load forecasting is an important element for economicallyefficient operation and for effective control of power systems. Thepurpose of the short-term load forecasting (STLF) is to forecast inadvance the system load, that represented by the sum of all theconsumers’ load at the same time. Also, precise load forecasting isrequired to avoid high generation cost and the spinning reservecapacity. Under prediction of STLF leads to insufficient reservecapacity preparation and threaten the system’s stability, on theother hand, over prediction of STLF leads to the unnecessarilylarge reserve that leads to high cost preparation.

The following factors specify the pattern of the electricityconsumption variation: (a) weather; (b) time; (c) historical data;and (D) random disturbances. The nature of parameters that affectthe load forecasting includes many uncertainties. Fuzzy logic (FL)is characterized by generalizing classical two-valued logic forreasoning under nonlinear and uncertain conditions; it is there-fore the most appropriate method in describing the humanknowledge that contains vague concepts and huge amount ofdata. Load forecasting using fuzzy implementation is considered

ll rights reserved.

ook).

to be faster and more accurate than the conventional forecastingmethods which deal with rigid data and have long processingtime. Regression analysis is considered as a conventional wayin power demand prediction. In statistics, regression analysis isthe process used to estimate the parameter values of a function,in which the function predicts the value of a response variablein terms of the value of other variables (Tong, 2007); more-over, the method that based on ‘‘similarity’’ principle whichuse information of the day being similar to the weather condi-tion is considered as a conventional method (Pandian et al.,2006).

Conventional load-forecasting systems in Jordan are normallybased on statistical modeling techniques; they have limitedchances in predicting accurate loads for abnormal days whensome irregular events occur due to abnormal weather conditionsor sudden temperature changes. They also need precise values inthe forecasting process, while the fuzzy method uses interval ofvalues like ‘‘high or low temperature’’. Such changeable conditionslead to the application of fuzzy implementations to be usedinstead of statistical schemes in the forecasting systems. And eventhough in the US, and most developed countries, the standardload-forecasting system uses a neural network. Fuzzy logic wasused by Pandian et al. (2006) to achieve much closer results to theactual consumption than other conventional methods usesthe similarity principle. And Li-Chih and Pan (2007) found that

ARTICLE IN PRESS

Fig. 2. Monthly variation of electricity consumption in year 2006.

R. Mamlook et al. / Energy Policy 37 (2009) 1239–12481240

the adaptive network-based fuzzy inference gave better resultsthan artificial neural-network model.

Many attempts by researchers (Jingfei, 2006; Chow et al., 2005;Garcıa, 2006; Gabbi and Zanotti, 2003; Negnevitsky, 2005) havebeen made to improve load-forecasting process in many world-wide regions. Khan and Abraham (2003) used a hybrid of artificialneural network (ANN) and fuzzy logic to forecast the load in CzechRepublic. They found that hybrid fuzzy neural network and radialbasis function networks are the best candidates for the analysis ofthe load in Czech Republic.

Another study has been carried out for power systems in Indiaby Pandian et al. (2006). They used fuzzy approach for short-termload forecasting using the temperature and time parameters onlyas inputs for the system. They showed that with the aid of fuzzylogic, they can achieve much closer results to the actualconsumption than the conventional methods which use thesimilarity principle.

Mamlook (2006) used fuzzy set methodology for evaluatingenergy production alternatives to compare between differentpower systems. He showed that the fuzzy method has fasterlearning technique than the neuro-fuzzy method. Fuzzy logicmethodology has been utilized for wide range of evaluationapplications, i.e., evaluation and comparison between differentsolar systems (Mamlook et al., 2001), evaluation of factorsaffecting solar still production (Mamlook and Badran, 2007),evaluation of parameters that affect leakage in infrastructuresystems (Mamlook and Al-Jayyousi, 2003).

Li-Chih and Pan (2007) applied the adaptive network-basedfuzzy inference system model to forecast the regional electricityloads in Taiwan, and demonstrated the forecasting performance oftheir model. They found that the adaptive network-based fuzzyinference gave better results than artificial neural-network model.

An integrated evolving fuzzy neural network and simulatedannealing for load-forecasting method is presented by Liao andTsao (2004) to reduce the error of conventional load forecasting.Their load-forecasting scheme was tested using data obtainedfrom a sample study includes 1 year, 1 month, and 24 h timeperiods. They showed that their method gave good accuracy.

Another study by Kodogiannis and Anagnostakis (1999)discussed the development of improved neural-network-basedshort-term electric load-forecasting models for the power systemof the Greek island of Crete. The performance was evaluated

Fig. 1. Typical consumption curv

through a simulation study, using metered data provided by theGreek Public Power Corporation. Their results indicated that theload-forecasting models developed provided more accurate fore-casts than the conventional methods.

Maia and Gonc-alves (2008) proposed an approach for next daypeak load forecasting for electrical companies. A nonlinear modelfor the peak load is proposed taking into account the historicalload and the temperature, each of which was estimated using anon-line recursive algorithm.

Hobbs and Nelson (1992) used a bilevel programming in theelectric utility industry. The model is nonlinear and is used toanalyze various economic issues that affect electric utilityplanning. The electric utility at the upper level of the modelseeks to minimize costs or maximize benefits while controllingelectric rates and subsidizing energy conservation programs.Customers at the lower level attempt to maximize their netbenefit by consuming electricity and investing in conservation.

Gribik et al. (2007) suggested that the electricity marketmodels require energy prices for balancing, spot and short-termforward transactions. For the simplest version of the coreeconomic dispatch problem, the formulation produces a well-defined solution to the pricing problem in the usual intersectionof the supply marginal cost and the demand bids. This pricingsupports the equilibrium solution and satisfies a no arbitragecondition. In the more general economic unit commitment anddispatch models, there may be no corresponding uniform energyprice vector that supports the solution. This introduces a needboth define the appropriate energy prices and determine the

e pattern for a normal day.

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–1248 1241

associated uplift make—whole payments needed to support thesolution. Different approximations of the optimal value functionyield different price and uplift results.

The manual forecasting previously performed by the systemdispatchers has been replaced by fuzzy logic STLF modelincorporated with meteorological effects to improve load predic-tions of the power systems in Jordan.

The input fuzzy sets have been modeled to represent thetemperature, historical data for different days, the forecastedweather and the day duty ‘‘Holiday or normal workday’’, while thefuzzy output represents ‘‘forecasted load’’. The proposed fuzzySTLF methodology was applied to a case study of the NationalElectric Power Company (NEPCO) in Jordan that based on actualhourly load data for the last 7 years (January 2000–June 2007).Thus, efforts aimed at the implementation of accurate andeffective STFL are highly worthwhile.

2. Load attribute in Jordan

Good understanding of the system load characteristics helps indesigning a reasonable forecasting model to achieve accurateresults at variable conditions.

The system load represents the sum of all the consumers’ loadat the same time (Jingfei, 2006; Chow et al., 2005).

Fig. 3. Consumption variation versus temperature (2000–2007): (a) Temperature

at mid-day peak load. (b) Temperature at evening peak load.

A typical load curve pattern in Jordan for normal day of 24 h isshown in Fig. 1.

Fig. 2 shows that the highest consumption rates occurred inthe summer season, when the energy demand for cooling and airconditioning is high (Garcıa, 2006), while the high consumption inwinter is due to the usage of heating systems (NEPCO, 2006).

The load pattern curve undergoes different changes due tounstable conditional factors; these factors that should beconsidered during the load-forecasting process (Jingfei, 2006)are as follows:

�

16

Fig200

weather,

� time, � economy, and � random disturbance.

Each parameter will be explained to show its influence on theconsumption variation:

2.1. Weather

Weather factors include temperature, humidity, wind speed,cloud cover, light intensity, etc. The Jordanian National ElectricPower company (NEPCO, 2006) use limited weather inputs; theyuse temperature at the peak load only, while the fuzzy logicconsiders the temperature at each hour to track the accurate loadvariation to achieve better output.

Other factors like humidity and wind speed should be includedin the forecasting process to achieve high accuracy of prediction(Khan and Abraham, 2003; Gabbi and Zanotti, 2003).

The pattern of the load consumption in Jordan that varies dueto the change in temperature is illustrated in Fig. 3a and b. Fig. 3aand b shows the behavior of the power consumption variationduring morning and evening peak load at low, normal, and hightemperatures. As can be inferred from Fig. 3a and b, the higherMegawatt (MW) consumption takes place at low and hightemperatures, while it starts to decrease when the temperaturegets near to the normal condition. In other words, the quantity ofenergy required for heating or cooling is strongly dependent onthe weather condition; below normal temperatures in wintercreate higher demand for heating; above normal temperatures insummer create higher demand for energy to meet air conditioningneeds (Khan and Abraham, 2003; Gabbi and Zanotti, 2003).

2.2. Time

Time plays an important role in influencing the load con-sumption, which includes the time during the day or at night asshown in Fig. 4. The consumption will consider whether it is anormal workday or a holiday, and the season. It can be seen fromFig. 4 that the consumption rate increased during the mid-day

3% 2%15%

%

29%

35% HouseholdIndustryCommercialWater pumpingStreet lightsOthers

. 4. Sectorial distribution of electricity consumption for the year 2006 (NEPCO,

6).

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–12481242

which corresponds to the working hours in the industrial, private,and governmental sectors; therefore in the evening, the con-sumption rate will be increasing, specially in the night time due tothe high usage of lights and household utilities which represent3–35% of the total sectorial consumption (Fig. 4).

2.3. Economy

The economic situation also affects the utilization of electri-city. The degree of industrialization is an explicit examplecategorized under the ‘‘economy’’ factor. Since the industrialsector consumption in Jordan represents 29% of the totalelectricity consumption as shown in Fig. 5, it is important toconsider the production/consumption layout of the main factoriesin Jordan, during the year, to improve the load forecasting and toget more accurate results.

2.4. Random disturbances

Random disturbances such as wars and conflicts will lead tothe increase in the number of electricity consumers due toimmigrations and the use of frequent emergency services. In spiteof the prevailing circumstances, Jordan’s stability and securityforms a good environment for investment, consequently, the rateof growth of investments during 2002–2006 has a steep growthrate as shown in Fig. 5 (Jordan Investment Board, 2006). Such

240

200

160

120

80

40

0

-40Foreign

-35.9%

4.7% 14.3%

80.9%

2003 - 20042002 - 2003

[%]

Fig. 5. Percentage growth of local and foreign in

Fig. 6. Fuzzy input/ou

steep growth rate has increased the electricity consumptionin Jordan.

3. Fuzzy sets proposed methodology

Over the past two decades, there has been a tremendousgrowth in the use of fuzzy logic controllers in power systemsapplications. Recent series of tutorials in the IEE Power Engineer-ing focused entirely on the applications of fuzzy logic in powersystems can be an evidence of the importance of the fuzzy modelsin the energy field (Cirstea et al., 2002).

Since precise load forecasting remains a great challenge, themain objective of the fuzzy methodology in the current study is todevelop a practical model that can achieve an accurate forecastingresult. Fuzzy logic has the ability to absorb the human experience,contain a large amount of data and infer the desired actionsrepresented with an accurate forecasting.

The advantages of using fuzzy logic in load-forecastingapplications include the following:

�

yea

ves

tput

Fuzzy method uses fuzzy sets that enabled us to condenselarge amount of data into smaller set of variable rule.

� Fuzzy logic controllers are based on heuristics and therefore

able to incorporate human intuition and experience (Cirsteaet al., 2002).

Local

r

47.2%

188.5% 205.1%

108.2%

2005 - 20062004- 2005

tment (Jordan Investment Board, 2006).

combination.

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–1248 1243

The fuzzy inputs used are listed as following; they showed goodresponse in tracking the actual load consumption:

�

Last day consumption (MWh). � Last week consumption (MWh). � Last day temperature (1C). � Forecasted temperature (1C). � Weather. �

Fig. 10. Membership function for weather input.

Day (binary sets determine whether it is a normal work day orholiday).

The classical four steps for fuzzy implementation are shownbelow:

3.1. Determining the linguistic variables and the fuzzy sets

The fuzzy input/output combination is shown in Fig. 6. Eachinput and output is divided into number of fuzzy sets as shownin Figs. 7–11.

Fig. 8. Membership functions for th

Fig. 9. Membership functions for tempe

Fig. 7. Membership functions for the forecasted load in Megawatt (MW).

The output is divided into seven fuzzy sets; it shows the degreeof consumption for the forecasted load with respect to the inputssituation, the fuzzy output sets are as follows:

�

e h

ratu

Very very low consumption (VVL).

� Very low consumption (VL). � Low consumption (L).

istorical data input in MW.

re inputs in degree Celsius (1C).

Fig. 11. Membership function for ‘‘Day’’ input.

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–12481244

�

Normal consumption (N). � High consumption (H). � Very high consumption (VH). � Very very high consumption (VVH).

The output consumption intervals were divided into Gaussiansets as shown in Fig. 7.

The values of the output consumption shown in Fig. 7 arewithin the range from 900 to 2000 MW which represents theminimum and the maximum consumption limit, respectively.

The intervals for the ‘‘Last day consumption’’ and the ‘‘Lastweek consumption’’ inputs have been divided into seven member-ship functions as follows:

�

Very very low consumption (VVL). � Very low consumption (VL). � Low consumption (L). � Normal consumption (N). � High consumption (H). � Very high consumption (VH). � Very very high consumption (VVH).

The membership functions for the historical data input areshown in Fig. 8 as follows:

‘‘Last day consumption’’ and ‘‘Last week consumption’’ inputsare included in the fuzzy implementation to determine thebehavior of the consumption during the last day and the lastweek, respectively.

The intervals for the (last day temperature) and (forecastedtemperature) have been divided into three membership functions(Fig. 9):

�

Low temperature (L). � Normal temperature (N). � High temperature (H).

Fig. 12. Forecasting proce

The average temperature values in Fig. 9 are occurring between0 and 40 1C which represent the low and the high averagetemperatures, respectively.

The weather interval inputs have been divided into fourmembership functions (snow, rain, cloud, and clear) and shownin Fig. 10.

‘‘The weather’’ inputs are used in the fuzzy implementation todetermine which weather condition is dominant (i.e., snow, rain,cloud, or clear) as shown in Fig. 10.

(Day) is an input with binary boundaries which determines ifthe day is a normal workday or holiday; it is illustrated in Fig. 11.

The ‘‘Day’’ input is used to determine the day duty (holiday orworkday) within 0–1 range.

3.2. Constructing fuzzy rules

In the current study, 37 fuzzy rules have been used to predictthe load under the effect of different inputs shown in Fig. 6. Therules processed are shown in Fig. 12.

3.3. Performing fuzzy inference into the system

The fuzzy inference is the process of formulating the mappingfrom a given input to output using fuzzy logic. The mappingthen provides a basis from which decisions can be made orpatterns discerned. The process of fuzzy inference involvesmembership functions, fuzzy logic operators, and if-then rules(The MathWorks, 1984–2007).

This procedure is used to compute the mapping from the inputvalues to the output values, and it consists of three sub-processes,fuzzification, aggregation, and defuzzification (Fig. 13).

For example, rule 6 shown in Fig. 13 shows that the ‘‘Last dayconsumption’’ was 1650 MW and day equal to 1 is a ‘‘Workday’’,the degree of activation for the output set at rule six carries the

ss using fuzzy logic.

ARTICLE IN PRESS

Fig. 14. Fuzzy rules.

Fig. 13. Fuzzy implementation sequence.

R. Mamlook et al. / Energy Policy 37 (2009) 1239–1248 1245

same minimum input set ‘‘Last day consumption’’ degree becausethe operator is ‘‘and’’. The activated sets due to fuzzification sub-process will be aggregated in the next step to form the combinedshape shown in Fig. 13, after that, it will be defuzzified to get acrisp number (forecasted load ¼ 1680 MW); the rules in theMATLAB toolbox are shown in Fig. 14.

If the (day) is holiday, then the (last day consumption) will beignored and ‘‘Last week consumption’’ value should be consideredas a last holiday.

4. Case study

The data for this case study have been taken from theNational Electric Power Company in Jordan. NEPCO’s capa-city portfolio includes fuel oil-fired steam generating units, gas-fired combustion turbines, diesel-fired combustion turbines,and diesel engines. NEPCO used to provide all the bulk powerto the national grid, except for that supplied by interconnectedindustrial companies (about 100 MW). NEPCO also operates

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–12481246

the bulk power network in Jordan; it operates the mainsubstations to provide services to the various distributionnetworks, which have an aggregate capacity of about 2101 MW(NEPCO, 2006).

The rules in the fuzzy logic were made based on the historicalsystem load data sheet of the years (2000–2007). The fuzzy setswere tuned to get the right response based on the variation of

Table 1Comparison between conventional and fuzzy forecasted load for 24 h 23/5/2007.

Time Conventional

load (MW)

Fuzzy forecasted

load (MW)

Actual

load (MW)

Conventional

error (%)a

Fuzzy

error (%)

23/5/2007 Wednesday

1 1178 1180 1176 0.2 0.3

2 1130 1160 1129 0.1 2.7

3 1083 1110 1095 1.1 1.4

4 1070 1120 1098 2.6 2

5 1105 1120 1093 1.1 2.5

6 1085 1100 1080 0.5 1.9

7 1152 1200 1195 3.6 0.4

8 1281 1340 1327 3.5 1

9 1424 1480 1509 5.6 1.9

10 1505 1530 1567 4.0 2.4

11 1566 1600 1614 3.0 0.9

12 1635 1620 1640 0.3 1.2

13 1585 1590 1610 1.6 1.2

14 1588 1610 1600 0.8 0.6

15 1586 1590 1591 0.3 0.1

16 1580 1570 1547 2.1 1.5

17 1485 1510 1528 2.8 1.2

18 1483 1470 1482 0.1 0.8

19 1451 1450 1418 2.3 2.3

20 1670 1670 1700 1.8 1.8

21 1604 1620 1633 1.8 0.8

22 1490 1510 1515 1.7 0.3

23 1326 1410 1417 6.4 0.5

24 1310 1270 1293 1.3 1.8

Average error 2 1.3

a MAPE ¼ |((conventional or fuzzy load)–(actual load))/(actual load)|.

Table 2Comparison between conventional and fuzzy forecasted load for 24 h in 15/1/2007.

Time Conventional

load (MW)

Fuzzy forecasted

load (MW)

Actual

load (MW)

Conventional

error (%)

Fuzzy

error (%)

15/1/2007 Monday

1 1220 1190 1149 6.2 3.6

2 1120 1130 1082 3.5 4.4

3 1075 1090 1060 1.4 2.8

4 1050 1050 1046 0.4 0.4

5 1121 1120 1128 0.6 0.7

6 1196 1190 1164 2.7 2.2

7 1345 1290 1192 12.8 8.2

8 1367 1390 1373 0.4 1.2

9 1502 1490 1481 1.4 0.6

10 1575 1590 1581 0.4 0.6

11 1589 1580 1601 0.7 1.3

12 1670 1620 1610 3.7 0.6

13 1655 1580 1590 4.1 0.6

14 1686 1590 1575 7.0 1

15 1652 1570 1547 6.8 1.5

16 1699 1620 1501 13.2 7.9

17 1806 1710 1620 11.5 5.6

18 1860 1790 1800 3.3 0.6

19 1798 1740 1723 4.4 1

20 1761 1700 1675 5.1 1.5

21 1770 1680 1667 6.2 0.8

22 1636 1580 1563 4.7 1.1

23 1542 1490 1448 6.5 2.9

24 1333 1310 1298 2.7 0.9

Average error 4.6 2.2

different parameters, the following Tables 1–4, compare betweenthe conventional forecasting and the fuzzy forecasting. The meanabsolute percentage error (MAPE) is used to measure the error forboth methods.

The error caused by the fuzzy method is due to the variednature of the load consumption and due to the limited input datathat have been discussed earlier. Whenever new input data are

Table 3Comparison between conventional and fuzzy forecasted load for 24 h in 18/1/2007.

Time Conventional

load (MW)

Fuzzy forecasted

load (MW)

Actual

load (MW)

Conventional

error (%)

Fuzzy

error (%)

18/1/2007 Thursday

1 1101 1170 1191 7.6 1.8

2 1033 1130 1108 6.8 2

3 966 1070 1040 7.1 2.9

4 964 1050 1053 8.5 0.3

5 945 1090 1065 11.3 2.3

6 1025 1160 1164 11.9 0.3

7 993 1260 1209 17.9 4.2

8 1103 1370 1335 17.4 2.6

9 1272 1520 1513 15.9 0.5

10 1411 1550 1589 11.2 2.5

11 1527 1590 1352 12.9 17.6

12 1565 1600 1512 3.5 5.8

13 1580 1570 1492 5.9 5.2

14 1544 1540 1483 4.1 3.8

15 1555 1540 1445 7.6 6.6

16 1630 1530 1460 11.6 4.8

17 1720 1630 1550 11.0 5.2

18 1770 1760 1730 2.3 1.7

19 1660 1680 1660 0.0 1.2

20 1620 1660 1620 0.0 2.5

21 1575 1620 1610 2.2 0.6

22 1510 1550 1545 2.3 0.3

23 1422 1430 1415 0.5 1.15

24 1342 1300 1318 1.8 1.4

Average error 7.6 3.2

Table 4Comparison between conventional and fuzzy forecasted load for 24 h in 29/6/2007.

Time Conventional

load (MW)

Fuzzy forecasted

load (MW)

Actual

load (MW)

Conventional

error (%)

Fuzzy

error (%)

29/6/2006 Thursday

1 1250 1270 1285 3 1.2

2 1160 1190 1210 4 1.7

3 1120 1140 1170 4 2.6

4 1110 1130 1130 2 0

5 1080 1130 1145 6 1.3

6 1035 1060 1080 4 1.9

7 1110 1120 1140 3 1.8

8 1277 1330 1360 6 2.2

9 1447 1460 1485 3 1.7

10 1514 1560 1548 2 0.8

11 1569 1620 1612 3 0.5

12 1600 1620 1640 2 1.2

13 1595 1610 1631 2 1.3

14 1583 1620 1600 1 1.3

15 1548 1590 1605 4 0.9

16 1518 1560 1565 3 0.3

17 1470 1510 1486 1 1.6

18 1425 1470 1420 0.3 3.5

19 1374 1400 1380 0.4 1.4

20 1480 1520 1535 4 1

21 1570 1600 1615 3 0.9

22 1470 1520 1520 3 0

23 1445 1470 1475 2 0.3

24 1334 1370 1370 3 0

Average error 2.8 1.2

ARTICLE IN PRESS

Fig. 15. Comparison between conventional forecasting load and fuzzy forecasting load in (a) 23/5/2007, (b) 15/1/2008, (c) 18/1/2007, and (d) 29/6/2006.

R. Mamlook et al. / Energy Policy 37 (2009) 1239–1248 1247

ARTICLE IN PRESS

R. Mamlook et al. / Energy Policy 37 (2009) 1239–12481248

existed, the mission for the fuzzy method will be facilitated topredict the load variation in more accurate way.

During the power plant visits, we noticed that the conventionalmethod of forecasting uses limited data inputs for the load-forecasting process; so that it is highly recommended to considerother parameters like the temperature level at each hour,humidity, wind speed, temperature, and changes in productionor consumption rate. So that fuzzy set implementation incorpo-rate the weather database is found to be highly capable to trackthe load variation in more accurate way. Hence, the conventionalforecasting previously performed by the system dispatchers hasbeen replaced by the fuzzy STLF software as a modern energymanagement system.

Fig. 15a–d shows small deviation (o5%) between the actualload and the fuzzy set evaluation that is due to the limitedness ofthe dispatchers data collected by NEPCO and due to the suddenvariations occur in the load consumption rate under the effect ofdifferent climatical conditions. But as an overall view, the fuzzyforecasting shows similar trend to the conventional method withmore accurate predictions, because it takes into account the realvariations of weather and rate of consumptions fluctuations. Also,Fig. 15a–d of the fuzzy method forecasted accurately the rate ofthe power consumption during morning and evening peak load.

5. Conclusions

Load forecasting is the most important factor for estimatingthe power needed for consumers; it provides load prediction forgeneration scheduling and unit commitment decisions, andtherefore precise load forecasting plays an important role inreducing the generation cost.

Fuzzy logic has the ability to absorb the human experience,contain a large amount of data and infer the desired actionsrepresented with an accurate forecasting. Fuzzy logic controllersare based on heuristics and therefore able to incorporate humanintuition and experience. The advantages of using fuzzy logic forload-forecasting applications are it condenses large amount ofdata (Tables 1–4) into smaller set of variable rules, it contains theoperators’ experience, and it infers the correct actions fromhistorical data.

Load forecasting using fuzzy implementation is faster andmore accurate than the conventional forecasting methods thatdeals with rigid data and have long processing time. Conventionalload-forecasting systems in Jordan are normally based onstatistical modeling techniques. They have limited chances inpredicting the loads for abnormal days when some irregularevents occur, at abnormal weather or changes in temperature, andthe sudden changes in the consumption rate. Hence, theincorporation of meteorological effects into STLF model willundoubtedly provide improved load predictions.

Forecasts using fuzzy logic consistently resulted in a forecast-ing error that was 5% less than conventional statistical forecastingmethods. It can be concluded that the fuzzy forecasted valuesprovided by fuzzy implementation have more accuracy and betteroutcomes than conventionally forecasted results. Also, the presentfuzzy STLF model can save the utility significant sums of moneyby reducing the error in load predictions.

References

Chow, H., Wu, F., Momoh, J., 2005. Applied Mathematics for Restructured ElectricPower Systems. Springer, England.

Cirstea, M.N., Dinu, A., Khor, J.G., McCormick, M., 2002. Neural and Fuzzy LogicControl of Drives and Power Systems. Newnes, England.

Gabbi, G., Zanotti, G., 2003. Climate variables and weather derivatives: gasdemand, temperature and seasonality effects in the Italian case. Social ScienceResearch Network.

Garcıa, S.V., 2006. Energy efficiency and energy policy evaluation in the Spanishhotel sector. Master Thesis, Utrecht University, Utrecht, The Netherlands.

Gribik, P.R., Hogan, W.W., Popeii, S.L., 2007. Market-clearing electricity prices andenergy uplift. Harvard Report, USA.

Hobbs, B.F., Nelson, S.K., 1992. A nonlinear bilevel model for analysis of electricutility demand-side planning issues. Annals of Operations Research 34,255–274.

Jingfei, Y., 2006. Power system short term load forecasting. Master Thesis,Technische Universitat, Darmstadt University.

Jordan Investment Board, 2006. Annual Report.Khan, M.R., Abraham, A., 2003. Short term load forecasting models in Czech

Republic using soft computing paradigms. International Journal of Knowledge-Based Intelligent Engineering Systems 7 (4), 172–179.

Kodogiannis, V.S., Anagnostakis, E.M., 1999. A study of advanced learningalgorithms for short-term load forecasting. Engineering Applications ofArtificial Intelligence Journal 44 (9), 159–173.

Li-Chih, Y., Pan, M.C., 2007. Using adaptive network based fuzzy inference systemto forecast regional electricity loads. Energy Conversion and Management 49(2), 205–211.

Liao, G.C., Tsao, T.P., 2004. Application of fuzzy neural networks and artificialintelligence for load forecasting. Electric Power Systems Research 70 (3),237–244.

Maia, C.A., Gonc-alves, M.M., 2008. Application of switched adaptive system to loadforecasting. Electric Power Systems Research 78 (4), 721–727.

Mamlook, R., 2006. Fuzzy set methodology for evaluating alternatives to comparebetween different power production systems. Journal of Applied Science 6 (9),2117–2125.

Mamlook, R., Al-Jayyousi, O., 2003. Fuzzy sets analysis for leak detection ininfrastructure systems: a proposed methodology. Journal of Clean TechnologyEnvironmental Policy 6 (1), 26–31.

Mamlook, R., Badran, O., 2007. Fuzzy sets implementation for the evaluation offactors affecting solar still production. Desalination 203, 394–402.

Mamlook, R., Akash, B., Nijmeh, S., 2001. Fuzzy sets programming to performevaluation of solar systems in Jordan. Energy Conversion and Management 42(14), 1717–1726.

Negnevitsky, M., 2005. Artificial Intelligence. Pearson Education, England.NEPCO, 2006. Annual Report. Amman, Jordan.Pandian, S., Duraiswamy, K., Rajan, C., Kanagaraj, N., 2006. Fuzzy approach for

short term load forecasting. Electric Power Systems Research 67 (6–7),541–548.

The MathWorks. Fuzzy Logic Toolbox. Fuzzy Inference System. The MathWorks,Natick, MA, 1984–2007. See also /http://www.mathworks.com/access/help-desk/help/toolbox/fuzzyS.

Tong, F., 2007. Capacity demand and climate in Ekero—development of tool topredict capacity demand under uncertainty of climate effects. Master Thesis,Stockholm, Sweden.