2017 Fourth International Conference on Image Information Processing (ICIIP)
Short Term Load Forecasting Using Artificial NeuralNetwork
Saurabh SinghDepartment ofElectrical Engineering
National Institute of TechnologySrinagar, India
E-mail:[email protected]
Shoeb HussainDepartment Of Electrical Engineering
University of KashmirSrinagar, India
Email: [email protected]
Mohammad Abid BazazDepartment Of Electrical Engineering
National Institute of TechnologySrinagar, India
Email: [email protected]
159
Abstract-Short term load forecasting is required for powersystem planning, operation and control. It is used by utilities,system operators, generators, power marketers. In this paper,load forecasting has been done using ANN (Artificial NeuralNetwork). As load profile is different for weekdays andweekends, so for better forecasting performance, training ofneural network has been done separately for weekdays andweekends. Accordingly forecasting is done separately forweekdays and weekends. Neural network toolbox with 20neurons has been used for forecasting load of NEPOOL region ofISO New England. Hourly temperature (Dry bulb), humidity(Dew point) and electricity load of NEPOOL region has beentaken from 2004 to 2008. ANN model is trained on hourly datafrom 2004 to 2007 and tested on out-of-sample data from 2008.The test set is used only for forecasting to test the performance ofthe model on out-of-sample data. Simulation results obtainedhave shown the comparison of actual and forecasted load data.Performance of forecaster is calculated using MAE, MAPE anddaily peak forecast error.
Keywords- Artificial neural network (ANN); ArtificialIntelligence (AI); Short term load forecasting (STLF); Meanabsolute percent error (MAPE); Mean absolute error (MAE)
I. INTRODUCTION
Load forecasting is a very important component in powersystem energy management system. There are threeforecasting horizons - short term, medium term and long term.STLF is hourly or sub-hourly forecasting of load forecastingstarting from next hour to next week. It is mainly required byelectric utilities for making unit commitment decisions, reducespinning reserve capacity, for generator type coordination todetermine least cost operation, for transmission line loading.interchange scheduling and energy purchase. In addition toutilities, other newly formed entities such as load aggregators,power marketers, and independent system operators also needgood quality of load forecasting for their operations. [1,2]
Accuracy has very significant economic impact. Even avery small fraction reduction in the forecasting error can resultin substantial saving. Accurate forecasting leads to amplesavings in operating and maintenance costs, increasedreliability of power supply and delivery system, and correctdecisions for future development. Overestimation in load
978-1-5090-6734-3/17/$31.00 ©2017 IEEE
forecasting leads to unnecessary increase in the reserve and theoperating costs. Underestimation of load forecasting results infailure to provide the required spinning and standby reserveand stability to the system, which may lead to collapse of thepower system network.
Various factors that affect STLF are geographical location,mix of customer in service area, weather conditions, seasonaleffect, time of the day, day of week and random disturbancesetc. Estimation of future load has been difficult up to now,especially for the days with extreme weather, on holidays andother anomalous days. With the recent development of newmathematical, data mining and artificial intelligence tools, it ispossible to improve the forecasting result. [3]
Different techniques have been developed for loadforecasting during the past few years. At first differentmathematical models have been proposed but they were unableto accurately model the weather parameters, they had lack ofrobustness for representing weekends and public holidays andwere computational intensive.[4] Regression models are able toanalyze the relation between load and the influencing factorsbut they require heavy computational efforts. With thedevelopment of artificial intelligence (AI) techniques researchworks have been carried out on the application of thesetechniques to the load forecasting problem as AI tools haveperformed better than conventional methods in short-term loadforecasting. [1]
AI techniques reported in literatures are expert systems,fuzzy inference, neural network, fuzzy-neural models. Amongthe different techniques on load forecasting, application ofANN to load forecasting in power system has received muchattention in recent years. ANN is becoming popular because ithas ability to learn complex and nonlinear relationships that aredifficult to model with conventional techniques. [5]
Neural network fitting tool of MATLAB has been used tocompute STLF for NEPOOL region (ISO New England).Historical hourly data of temperature, humidity and electricityload have been used for load forecasting.
This paper has been organized in five sections. Section IIpresents the overview of ANN. Section III describes the datapreprocessing and modeling of ANN. Results of simulation are
2017 Fourth International Conference on Image Information Processing (ICIIP)
presented and discussed in Section IV. Section V discusses theconclusion and future work.
The input to neural network is as follows:-
II. ANN FOR STLF
ANN fundamentally performs like a human brain. A neuralnetwork is a massively parallel-distributed processor made upof simple processing units, known as neurons. This networkconsists of number of layers containing neurons and weightsassociated with the connection between neurons where theinformation is passed in feed-forward manner. A model ofartificial neuron is shown in Fig. 1. Neurons of ANN consistsof 3 main components:- weights connecting the nodes, thesummation function within the node and transfer function. Inthis model there is I input layer, I hidden layer and I outputlayer, as this structure of MLP is proven to address almost anykind of non-linear relationship. The neural network consists oftwo-layer feed-forward network with sigmoid activationfunction in hidden neurons and linear in output neurons.[6]First the information is passed in feed forward manner viainput to output then weight is adjusted via back propagation.Adjustment of weight and bias is called learning which is doneby Levenberg-Marquardt back propagation algorithm.
~ Dry bulb temperature
~ Dew point temperature
~ Hour of the day
~ Day of the week
~ Working day or holiday/weekend indicator(O forholiday and I for working day)
~ Previous day same hour load
~ Previous week same day same hour load
~ Previous 24 hr average load
These hourly data is used as input and hourly load is used asoutput to train neural network. Separate training andforecasting has been done for working days and weekends.
B. ModellingThe default neural network fitting toolbox of matlab consistingthree layers and 20 neurons has been used. Activation functionis sigmoid for the hidden layer and linear for output layer. [8]
Inputs Weights Activat ion Funct ion Output
Where At is actual load, F, is forecasted load and N is numberof observation points.
IV. SIMULATION AND RESULTS
Now neural network is trained with above 8 inputs and oneoutput separately for working days, weekends only andweekends including holidays using training set (2004-2007).The training set is used for building the model (estimating itsparameters). This trained neural network is tested on testingdata of year 2008. The test set is used only for forecasting totest the performance of the model on out-of-sample data.
Then a plot is created to compare the actual load and thepredicted load and to compute the forecast error. In addition tothe visualization, the performance of the forecaster isquantified using metrics such as mean average error (MAE),mean absolute percent error (MAPE) and daily peak forecasterror.
1 N
- LI4-F;1N i=l
M.AP.E.
M.AE. =
: Artificial Neuron :L I
Xl
Fig . 1. Model of an Artificial Neuron
III. LOAD FORECASTING METHODOLOGY
Hourly raw data in form of dry bulb, dew point andelectricity load from ISO NEW ENGLAND has been takenfrom year 2004-2008[7]. Dry bulb is taken for temperatureand Dew point for humidity as load depends on both theseparameters.
A. Data preprocessingIn this paper data from 2004-2007 has been used for ANN
training and 2008 data for testing of trained neural network. Asthe load profile is different for weekdays and weekends,separate analysis has been done for both. As load at anyparticular hour is similar to some previous days loads sopreprocessing of data has been done to train neural network.
160
2017 Fourth International Conference on Image Information Processing (ICIIP)
A. Forecasting for weekdays
• Neural network is trained on weekdays data of 2004-2007and tested on 2008 weekdays data.
• Forecasting error as a function of weekdays is shown inFig.4.
• A comparison of actual and forecasted load data of oneweek and residual of that data is shown in fig.2.:-
MAP.E. = 1.38%MAE. = 214.66 MWhDailyPeakMAP.E. = 1.34%
Data & Model Prediction
Break d own of forecast error st at ist ic s by w eekd ay
15 ++=1=
+ *~
0 10 + t~
ro =1= +ii5
'*' '" + +0
'" !+ +w 1 + +
c
t ~ f~
i;; 50- I
0$ $ $ $ $
M on Tue W ed Thu Fr i05/1005109051080510705105108
Res iduals
,OO~,~====
-100005105108 05107 05108 05,m 05/10
Fig. 4. Box-plot of the error distribution of forecasted load as a function ofweekdays for year 2008.
Fig. 2. Comparison of actnal and forecasted load for a weekday of year 2008.
Forecasting error as a function of month is shown in Fig.5.
Fig. 5. Box-plot of the error distribution of forecasted load as a functionofmonth of weekdays for year 2008.
+
+
++:j:
*t++ '" $
+ t• T I
$ $ ~
Breakdown offorecast error statistics by m ont h
01 02 03 04 05 06 07 08 09 10 11 12
15
•
Brea k down o f forecast error stati stics by h our
15 ++
+ ++ + +
+ + + ++
- 10 ++ +
I + + ++
:j: :j: + :j:w + f+ + + +"!":j: +
i + :j: * + +",+:j:5 + * + ~ + :j: i $ * + $ , f ~ T *+ + +
c,+ 'f + r t 't
lT I T7T Tf il 1III t f T t
o $ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ S ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~1 23 4 5 6 78 9 1011 121 314 151 61 718 1920 21 22 23 24
Hour
• Plot of percent forecast errors by hour of day, day of weekand month of the year for weekdays has been shown tocheck performance of forecaster.
Forecasting error as a function of hour of the day is shown inFig. 3.
Fig. 3. Box-plot of the error distribution of forecasted load as a function ofhour ofthe day for weekdays of year 2008.
161
2017 Fourth International Conference on Image Information Processing (ICIIP)
B) Forecasting for weekends:-
Now neural network is trained on weekends data of 2004-2007and tested on weekends data of 2008.
• Forecasting error as a function of weekends is shown inFig. 8.
• A comparison of actual and forecasted load data of oneweekend and residual of that data is shown:-
MAP.E. = 1.40%MAE. = 202.22 MWhDaily Peak MAP.E. = 1.78%
Residuals
1::\(==#:··············]05/10108 05/11 05/12
Fig. 8. Box-plot of the error distribution of forecasted load as a function ofweekends for year 2008.
Breakdown of forecast error stat ist ics by weekend
+12
'*'10 :j:
~ +0
~
8 '*'ra
1+
U5
i0
W 6c
'" +0
'" 40- I2 $ $0
S8t Sun
• Forecasting error as a function of month is shown in Fig. 9.
05/1205/11
Data & Model Prediction
05/10108
Fig. 6. Comparison of actual and forecasted load for a weekend of year 2008
• Plot of percent forecast errors by hour of day, day of weekand month of the year for weekends has been shown tocheck performance of forecaster.
Forecasting error as a function of hour of the day is shown inFig. 7.
Breakdown of fo recast e rror statistics by h ou r
15 ++
+
:j:+
10- + +
Ui + :j:'= + + f* +$ + +LU + + 1
1 5
~ ~~ ~~ ~ ~ ~ ~ ~~ ~~ ~ ~ ~ ~~ ~ ~ ~~ ~ ~a1 2 3 4 5 6 7 8 9 10 11 121 3 14 15 161 7 181 9 2 0 2 1 2 2 2 3 24
H o u r
Fig. 7. Box-plot of the error distribution of forecasted load as a function ofhour ofthe day of weekends for year 2008.
Breakdown offorecast e rror stat is tics by m onth
15 ++
+
10:j:
+.~
.~ *Ui + ++ :j: t $ *en + +
;I; -t- + + +~ + *
+
f$ + i~ 5 +
~ * + t T
~T
c, + r * I T ; I It T T
~ $$o ~ ~ $ $ $ $ $ ~ $0 1 02 03 04 05 0 6 07 08 09 1 0 11 12
Fig. 9. Box-plot of the error distribution of forecasted load as a function ofmonth of weekends for year 2008.
C) Then the neural network is trained using weekends(including holidays) data and tested upon it.
M.A.P.E. = 1.73%MAE. = 245.73 MWhDaily Peak MAP.E. = 1.82%
162
2017 Fourth International Conference on Image Information Processing (ICIIP)
The testing error for different testing days is shown in Table ITable I
Hourlv Load Forecast SummaryTesting Days MAPE MAE(MWh)
Weekdays(Mon-Fri) 1.38 214.66Weekends(Sat, Sun) 1.39 202.22
Weekends(including holidays) 1.73 245.73
• To forecast the load for any particular day take raw data fromdatabase. Create the predictor matrix and then predict loadusing trained neural network. For example load profile of 3April 2008 is forecasted below. This 24 hour (Day ahead)forecasting is required by system operator for various energytransactions.
Load Forecast Profile for 03-Apr-2008
V. CONCLUSION AND FUTURE WORK
This paper presents hourly short-term electricity load forecastusing artificial neural network (ANN) in NEPOOL region(ISO New England). Performance of ANN is calculated byM.A.P.E. Separate analysis has been done for weekdays andweekends to improve forecasting accuracy. A M.A.P.E. of1.38% for weekdays and 1.39% for weekends has beenobtained which shows good prediction with less error inforecasting.
Other parameters that affect short term load are weatherparameters like precipitation, wind velocity and customerclass (Industrial, Residential and commercial). If theseparameters will also be taken into account, result offorecasting can be further improved.
18
17
16
$' 15::;;:00 140
E-:il 130---'
12
11
100
, ,I I I I_____________ ~ L. ~ _ _ _ _ _ _ _ _ _ _ _ _ _L. _ _ _
I I I I
I I I II I I I
, ", ", "- - - - - - - - - - - - - ~ - - - - - - - - - - - - -~ - - - - - - - - - - - - - - - - - -'- - - - - - - - - - --, ,, ,, ,, ,, ,-------------,------ -------:--------------:--------------:------ ------
, , ,, , ,, , ,, , ,I I I I
- - - - - - - - - - - - - 1 - - - - - - - - - - - - -:- - - - - - - - - - - - - - ~ - - - - - - - - - - - - - -:- - - - - - - - - - --
, , ,, , ,, , ,, , ,I I I I- - - - - - - - - - - - -, - - - - - - - - - - - - -, - - - - - - - - - - - - -, - - - - - - - - - - - - - -, - - - - - - - - - - --, ,, ,, ,, ,, ,I I I I-------------,- ------------,..-------------,--------------,..------------I I I II I I I
-- -j--------------f-------------i-------'I---+-----'--;-N-eu-ra-::-IN-:-e-t lI I I II I I II I I I
Fig.l 0 Load profile of 03 April 2008
25
163
References
[I] Park, D.C., El-Sharkawi, M. A., Marks, R. J.,Atlas, L.E., &Damberg,M.J.(l991), Electric Load Forecasting Using an Artificial NeuralNetwork. IEEE Transactions on Power Systems, Vol.3, No.2, pp. 442449.
[2] G. Gross, F. D. Galiana, 'Short-term load forecasting', Proceedings ofthe IEEE, 1987,75(12), 1558 -1571.
[3] Paras Mandai, Tomonobu Senjyu, Atsushi Yona, Jung-Wook Park andAnurag K. Srivastava, "Sensitivity Analysis of Similar Days Parametersfor Predecting Short-Term Electricity Price", IEEE Trans. Power Syst.,E-ISBN: 978-1-4244-1726-1, pp. 568-574, September 2007.
[4] Mohsen Hayati and Yazdan Shirvany, "Artificial Neural NetworkApproach for Short Term Load Forecasting for Illam Region",International Journal of Electrical, Computer, and Systems EngineeringVolume I, Number 2, 20071SSN 1307-5179.
[5] K.Y. Lee, Y.T. Cha and I.H. Park, "Short Term Load Forecasting UsingAn Artificial Neural Network", IEEE Transactions on Power Systems,Vol I, No I, February 1992.
[6] A. K. Mahalanabis, D. P. Kothari, S.l. Ahson, Computer AidedPowerSystem Analysis and Control, Tata McGraw Hill PublishingCompany limited, New Delhi,1988.
[7] Website for historical load data. http://www.iso-ne.com/.[8] Neural Network overview from Neural Netwok toolbox.