You are on page 1of 6

ANN-based Short-Term Load Forecasting

in Electricity Markets
Hong Chen Claudio A. Cañizares Ajit Singh
University of Waterloo
Department of Electrical & Computer Engineering
Waterloo, ON, Canada N2L 3G1

Abstract—This paper proposes an Artificial Neu- in practice for a long time, showing the forecasting accu-
ral Network (ANN)-based short-term load forecasting racy that is system dependent. These traditional methods
technique that considers electricity price as one of the can be combined using weighted multi-model forecasting
main characteristics of the system load, demonstrating techniques, showing adequate results in practical systems
the importance of considering pricing when predicting [10]. However, these methods cannot properly represent
loading in today’s electricity markets. Historical load
data from the Ontario Hydro system as well as pricing the complex nonlinear relationships that exist between the
information from the neighboring system are used for load and a series of factors that influence it, which are typ-
testing, showing the good performance of the proposed ically dependent on system changes (e.g., season or time
method. of day).
Keywords: Short-term load forecasting, electricity Modern load forecasting techniques, such as expert sys-
markets, spot prices, Artificial Neural Networks tems [11], Artificial Neural Networks (ANN) [1, 12, 13, 14],
(ANN) fuzzy logic [15], wavelets [16], have been developed re-
cently, showing encouraging results. Among them, ANN
methods are particularly attractive, as they have the abil-
I. Introduction ity to handle the nonlinear relationships between load and
Short-Term Load Forecasting (STLF) is basically aimed the factors affecting it directly from historical data and
at predicting system load with a leading time of one hour to without having to select a given model. Different types
seven days, which is necessary for adequate scheduling and of ANNs have been applied to STLF; for example, multi-
operation of power systems. STLF traditionally has been layer feed-forward with one hidden layer [12], recurrent
an essential component of Energy Management Systems [14], and functional links [13]. In this paper, the more
(EMS) [1, 2, 3], as it provides the input data for load flow common multi-layer feed-forward ANN is used for STLF.
and contingency analysis. The paper is structured as follows: In Section II, the
load characteristics and the factors that influence it are
With the worldwide deregulation of the power industry, discussed. The ANN forecasting model and procedure used
load forecasting is becoming even more important, not only here are described in Section III. In Section IV, the histor-
for system operators, but also for market operators, trans- ical data from Ontario Hydro system is used to test the
mission owners, and any other market participants, so that proposed STLF technique. Finally, in Section V, the main
adequate energy transactions can be scheduled, and ap- contributions of this paper are highlighted and possible
propriate operational plans and bidding strategies can be research directions are discussed.
established. Thus, load forecasting has also become an im-
portant component of energy brokerage systems [4]. In this
new context, high forecasting accuracy and speed are re- II. Load Characteristics
quired not only for reliable system operation, but also for
adequate market operation, as both under-forecasts and Traditionally, system load at any given time is assumed
over-forecasts would result in increased operational costs to be a combination of four separate components, i.e.,
and loss of revenue. L = Ln + Lw + Ls + Lr (1)
To achieve high forecasting accuracy and speed, which
are the two most important requirements of STLF, it is where
important to analyze the load characteristics and identify L stands for the total system load;
the main factors affecting the load. In electricity markets,
in addition to the traditional load affecting factors such as Ln represents the “normal” part of the load, which is a
season, day type and weather, electricity price, which is set of standardized load shapes for each “type” of day
voluntary and may have a complicated relationship with that has been identified as occurring throughout the
system load, is also becoming an important factor influ- year;
encing the load. In this paper, the effect of pricing on the
STLF of the Ontario Hydro system is examined, showing Lw corresponds to the weather-sensitive part of the load,
that better results are obtained when this factor is consid- which is tightly coupled to the season of the year;
ered in the forecasting method. Ls stands for the special event part, which is the occur-
Various forecasting techniques have been applied to rence of an unusual or special event leading to a sig-
STLF to improve accuracy and efficiency. In general, nificant deviation from the typical load behavior;
STLF techniques can be classified as either traditional or
modern [5]. Traditional statistical load forecasting tech- Lr corresponds to a random part, which is an “unex-
niques, such as regression [6], time series [7], pattern plained” component usually represented as zero mean
recognition [8], Kalman filters [9], etc., have been used white noise.
x1 Forecasted Load
Wj1 Y1 Y2 Y3 Ym

x2 Wj2 ... Output layer

xk Wjk fj Oj
...
Wjn ... Hidden layer
xn
Fig. 1. Mathematical model of an ANN neuron.

... Input layer


In the present competitive electricity markets, system
load may also be significantly affected by prices. Since elec- X1 X2 X3 Xn
tricity must be produced and consumed instantaneously,
and there are only limited transfer capabilities of transmis- Load affecting factors
sion systems [17], electricity prices vary depending on place
and time, presenting relatively high variations as compared Fig. 2. Schematic of the three layer feed-forward ANN for STLF.
to other commodities. Consumers, especially price-elastic
consumers, can adjust their consumption behavior accord-
ing to the price information to achieve maximum benefits. 2. Network Architecture The three-layer fully con-
Hence, price should be factored in the relationship between nected feed-forward neural network depicted in Fig. 2 is
system load and its influencing factors, i.e., used here; it includes an input layer, one hidden layer
and an output layer [19]. Signal propagation is allowed
L = f (day,weather,special,price,random) (2) only from the input layer to the hidden layer and from
the hidden layer to the output layer. Input variables come
where f (·) is a highly nonlinear function that is difficult to from historical data corresponding to the factors that affect
represent explicitly, and hence forecasting the system load the load. The outputs are the desired forecasting results,
accurately with traditional statistical methods is a rather which in this case are m = 24, i.e., one for each hour of
complex problem. the day.
The number of inputs, the number of hidden nodes,
transfer functions, scaling schemes, and training methods
III. STLF Modeling and Procedure affect the forecasting performance and hence need to be
chosen carefully.
ANN-based methods are a good choice to study the
STLF problem, as these techniques are characterized by 3. Training ANN training basically consists on de-
not requiring explicit models to represent the complex re- termining the network parameters such as weights and
lationship between the load and the factors that determine others, that allow achieving the desired objective based
it. on the available training sets. Usually, multi-layer feed-
This section discusses the ANN model and the procedure forward neural networks are trained in a supervised man-
used here for STLF. ner. Back-propagation is used as the training method here
[18], which is an iterative procedure that has three steps
A. Modeling during each iteration:
1. Neuron model ANNs are made up of a number 1. Forward: The outputs are calculated for given inputs.
of simple and highly interconnected Processing Elements
(PE), called neurons, as depicted in Fig. 1 [18]. Its math- 2. Backward: The errors at the output layer are prop-
ematical model is expressed as agated backwards toward the input layer, with the
partial derivatives of the performance with respect to
 the weights and biases calculated in each layer.
Oj = fj (wjk xk ) (3)
k
3. Weight adjustment: A multivariate nonlinear numeric
optimization algorithm finds the weights that mini-
mize the error based on the gradient.
where
Training stops when the performance has been mini-
Oj is the output of a neuron; mized to the goal, the performance gradient falls below
a minimum gradient, the maximum number of epochs is
fj is a transfer function, which is differentiable and non- reached, or the maximum amount of time has been ex-
decreasing, usually represented using a sigmoid func- ceeded.
tion, such as a logistic sigmoid, a tangent sigmoid, The error function used in the back-propagation training
etc.; process is the sum-squared error, i.e.,
1 
wjk is an adjustable weight that represents the connection E= (tpj − Opj )2 (4)
strength; 2 p j

xk is the input of a neuron. where tpj and Opj are the target output and the actual
output j for input pattern p, respectively.
Input variable selection follows:
(k) (k)
Xi = (Xi − X̃i )/Si
(k) (k) (7)
Oi = (Oi − Õi )/SOi
Data Pre-processing
where Si and SOi are the estimates of the standard devia-
tion of input and output i, respectively, and X̃i and Õi are
Scaling the average values of the corresponding input and output.
The error function has different minimum values for the
Traing data set two different scaling schemes.
Training 4. Training Each layer’s weights and biases are ini-
tialized when the neural network is set up. The network
adjusts the connection strength among the internal net-
Simulation work nodes until the proper transformation that links past
inputs and outputs from the training cases is learned. Data
windows are used for training and moved one day ahead.
Post-Processing 5. Simulation Using the trained neural network, the
forecasting output is simulated using the input patterns.
6. Post-Processing The neural network output need
Error analysis de-scaling to generate the desired forecasted loads. If nec-
essary, special events can be considered at this stage.
Fig. 3. ANN-based demand forecasting procedure. 7. Error Analysis As characteristics of load vary,
error observations are important for the forecasting pro-
cess. Hence, the following Mean Absolute Percentage Er-
ror (MAPE) ε and Root Mean Square Error (RMSE) σ
The learning function used in the training process is are used here for after-the-fact error analysis:
a gradient descent with momentum weight/bias function,
which allows to calculate the weight change for a given N
neuron. It is expressed as 1  |Xt − Xf |
ε= ∗ 100 (8)
N i=1 Xt
dW = mc ∗ dWprev + (1 − mc ) ∗ lr ∗ gW (5)

 N
1 
σ=
where dWprev is the previous weight change, gW is the
weight gradient with respect to the performance, lr is the (Xt − Xf )2 (9)
N i=1
learning rate, and mc is the momentum. Different learning
rates and momentum affect the convergence properties. where Xt is the actual load and Xf is the forecasted load.
B. Forecasting Procedure
IV. Test Results
The STLF procedure for the chosen ANN model is
shown in Fig. 3. The historical data of Ontario system load for May 1999
1. Input Variable Selection Input variables such as was used for testing the proposed ANN-based STLF. Due
load, day type, temperature and spot prices of the previ- to the demerger of Ontario Hydro in early 1999, Ontario
ous day, and day type, temperature and spot prices of the electricity prices may be assumed to have had no ma-
forecasting day are initially chosen. jor impact on the load of interest. However, due to the
large power exchange between Pennsylvania-New Jersey-
2. Data Pre-processing Improperly recorded data Maryland (PJM) and Ontario, and the fact that PJM has
and observation error are inevitable. Hence, bad and ab- operated a large energy spot market for a while, one may
normal data are identified and discarded or adjusted using reasonably assume that these prices may have had a sig-
a statistical method to avoid contamination of the model nificant impact on the overall Ontario Hydro system load
[3]. for the period of interest. Thus, PJM electricity spot price
3. Scaling Since the variables have very different data for the corresponding time period is used here as the
ranges, the direct use of network data may cause conver- price input for the forecasting process.
gence problems. Two scaling schemes are used and com- A neural network with 100 inputs, 48 hidden nodes, and
pared. In the first scheme, all input Xi and output Oi 24 outputs, with the second scaling scheme as previously
variables are scaled to be in the [0, 1] range; hence, the described, is used here. A tangent sigmoid function was
input and output variables are scaled as follows: chosen as the transfer function for the hidden layer, and
a linear function for the output layer. This particular
(k) (k) (k)
ANN was chosen based on several tests with different num-
Xi = Xi /max(Xi ) ber of inputs, hidden nodes, transfer functions and scaling
(k) (k) (k) (6) schemes.
Oi = Oi /max(Oi ) The training performance for both a learning rate lr and
a momentum mc set at 0.6 is shown in Fig. 4. The fast
where k is the index of input and output vectors/patterns. convergence without oscillation is due to the relative large
In the second scheme, the input and output variables values of lr and mc , which were chosen based on multiple
are scaled to be in the [−c, c] range, where c is a positive tests carried out with different values of these two param-
number. The inputs and outputs in this case are scaled as eters.
Performance is 0.00999021, Goal is 0.01
1
10

4
x 10
1.7
forecasted demand
actual demand

0 1.6
10
Training−Blue Goal−Black

1.5

demand (MW)
1.4

−1
10

1.3

1.2

−2
10
0 100 200 300 400 500 600 700 800 900 1.1
928 Epochs 0 5 10 15 20 25
hour

Fig. 4. Training performance of the chosen ANN. Fig. 6. Forecasting results for Saturday May 29, 1999.
4
x 10
1.55
forecasted demand
actual demand
1.5

1.45

1.4
demand (MW)

1.35

0.035
1.3
0.03

1.25 0.025

0.02
error

1.2
0.015

1.15 0.01

0.005
1.1
0 5 10 15 20 25
hour 0
25
20 30
25
Fig. 5. Forecasting results for Monday May 24, 1999. 15
20
10 15
5 10
5
0 0
hour
date

The STLF results for Monday May 24, 1999 and Satur-
day May 29, 1999 are depicted in Figs. 5 and 6; observe the Fig. 7. Absolute percentage errors for the May 1999 load data set.
satisfactory forecasted results. The adequate performance
of the proposed STLF is also illustrated in the absolute
percentage error plot for each hour on each day of May
1999 shown in Fig. 7; the MAPE is 0.0048 (0.48%) and
the maximum absolute error is 0.03 (3%).
If price effects are not considered, the number of iter-
ations required to achieve the same forecasting accuracy
practically doubled, as shown in Table I. If the ANN pa- TABLE I
Comparison of the Number of Iterations
rameters obtained from the training process in both cases,
i.e., with and without price consideration, are used for fu-
ture STLF, better forecasting results can be obtained by lr ,mc without price with price
including price factor, as shown in Fig. 8. In other words, lr =0.6 1790 1079
the PJM energy spot market prices do have an influence mc =0.4
on the Ontario Hydro’s system load, and better overall lr =0.6 2278 1065
STLF performance can be achieved when the price effect mc =0.5
is considered. lr =0.6 1711 927
mc =0.6
V. Conclusions lr =0.6 4170 1131
mc =0.7
An ANN-based STLF method that uses a three layer
feed-forward neural network and a back-propagation train-
ing method is presented here, so that electricity prices
can be considered as one of the main factors affecting the
4
x 10
2 [5] I. Moghram and S. Rahman, “Analysis and Evaluation of Five
Short-Term Load Forecasting Techniques,” IEEE Trans. Power
Systems, Vol. 4, No. 4, Oct. 1989, pp. 1484–1491.
1.9

[6] A. D. Papalexopoulos and T. C. Hesterberg, “A Regression-


1.8
Based Approach to Short-Term System Load Forecasting,”
IEEE Trans. Power Systems, Vol. 5, No. 4, Nov. 1990, pp. 1535–
1547.
1.7
[7] M. T. Hagan and S. M. Behr, “The Time Series Approach to
demand (MW)

1.6
Short-Term Load Forecasting,” IEEE Trans. Power Systems,
Vol. 2, No. 3, Aug. 1987, pp. 785–791.

1.5 [8] A. S. Dhdashti, J. R. Tudor, and M. C. Smith, “Forecasting


of Hourly Load By Pattern Recognition: A Deterministic Ap-
proach,” IEEE Trans. Power Apparatus and Systems, Vol. 101,
1.4
No. 9, Sept. 1982, pp. 3290–3294.
actual demand
forecasted (without price)
1.3 forecasted (with price) [9] J. Toyada, M. Chen, and Y. Inoue, “An Application of State
Estimation to Short-Term Load Forecasting, I and II,” IEEE
Trans. Power Systems, Vol. 89, Sept. 1970, pp. 1678–1688.
1.2
0 5 10 15 20 25
hour [10] H. Chen and J. Liu, “A Weighted multi-model Short-term Load
Forecasting System,” Proc. IEEE Int. Conference on Power
Fig. 8. Comparison of the forecasting results for June 1, 1999. This System Technology, New York, NY, Vol. 1, 1998, pp. 557–561.
data is not used in the ANN training.
[11] S. Rahman and R. Bhatnagar, “An Expert System Based Al-
gorithm for Short-Term Load Forecast,” IEEE Trans. Power
Systems, Vol. 3, No. 2, May 1988, pp. 392–399.
load in deregulated markets. The load-price relationship is [12] C. N. Lu, H. T. Wu, and S. Vemuri, “Neural Network Based
highly nonlinear and difficult to model directly; hence, by Short Term Load Forecasting,” IEEE Trans. Power Systems,
considering price as another input to a neural network, it Vol. 8, No. 1, Feb. 1993, pp. 337–342.
is shown here that it is possible to readily account for the
[13] P. K. Dash, H. P. Satpathy, A. C. Liew, and S. Rahman, “A
impact of price on system load, without the need for deter- Real-time Short-Term Load Forecasting System Using Func-
mining an actual nonlinear function to represent this rela- tional Link Network,” IEEE Trans. Power Systems, Vol. 12,
tionship. The historical load data for the Ontario Hydro No. 2, May 1997, pp. 675–680
system as well as the prices from the neighboring electricity
[14] J. Vermaak, “Recurrent Neural Networks for Short-Term Load
market were used in this paper for testing, demonstrating Forecasting,” IEEE Tran. Power Systems, Vol. 13, No. 1, Feb.
that price does have an impact on the performance of a 1998, pp. 126–132.
STLF technique.
Since the relationship between electricity price and load [15] S. E. Papadakis, “A Novel Approach to Short-Term Load Fore-
casting Using Fuzzy Neural Network,” IEEE Tran. Power Sys-
is complex and dynamic, further research is needed to tems, Vol. 13, No. 2, May 1998, pp. 480–492.
study how different customers’ price response character-
istics and locations affect load forecasting. STLF could [16] T. Zheng, A. A. Girgis, and E. B. Makram, “A Hybrid Wavelet-
be performed considering smaller areas and various large Kalman Filter Method for Load Forecasting,” Electric Power
Systems Research, Vol. 54, No. 1, April 2000, pp. 11–17.
consumers to achieve higher forecasting accuracy.
[17] F. C. Schwapper, W. C. Caramanis, R. D. Tabors, and R. E.
Bohn, Spot Pricing of Electricity, Kluwer Academic Publisher,
VI. Acknowledgements 1988.

The authors would like to thank Mr. Derek Beamer of [18] C. Lau, Neural Networks Theoretical Foundations and Analysis,
Ontario’s Independent Electricity Market Operator (IMO) IEEE Press, 1992.
for providing data and valuable suggestions. [19] R. J. Schalkoff, Artificial Neural Networks, McGraw-Hill Inc.,
This research has been possible thanks to funding pro- 1997.
vided by NSERC and the E&CE Department of the Uni-
versity of Waterloo. Hong Chen received her Bachelor (1992) and Master(1995) de-
gree in Electrical Engineering from Southeast University in China.
References She worked in NARI (Nanjing Automation Research Institute,
P.R.China) from 1995 to 1998, and was engaged in developing EMS
power application software. Now she is a Ph.D. student in the De-
[1] A. D. Papalexopoulos, S. Hao, and T. M. Peng, “An Implemen- partment of Electrical & Computer Engineering at the University
tation of a Neural Network Based Load Forecasting Model for of Waterloo and doing research on electricity markets and computer
the EMS,” IEEE Trans. Power Systems, Vol. 9, No. 4, Nov. applications in power systems.
1994, pp. 1956–1962.
Claudio A. Cañizares received in April 1984 the Electrical Engi-
[2] H. Chen, “A Practical On-line Predicting System for Short- neer diploma from the Escuela Politécnica Nacional (EPN), Quito-
Term Load,” East China Electric Power, Vol. 24, No. 3, March Ecuador, where he held different teaching and administrative po-
1996. sitions from 1983 to 1993. His MS (1988) and PhD (1991) de-
grees in Electrical Engineering are from the University of Wisconsin–
[3] H. Chen, “An Implementation of Power System Short-Term Madison. Dr. Cañizares is currently an Associate Professor at the
Load Forecasting,” Power System Automation, China, Dec. University of Waterloo, E&CE Department, and his research activi-
1997. ties are mostly concentrated in studying stability, modeling and com-
putational issues in ac/dc/FACTS systems.
[4] I. Slutsker, K. Nodehi, S. Mokhtari, K. Burns, D. Szymanski and Ajit Singh completed his B.Sc. degree in Electronics and Com-
P. Clapp, “Market Participants Gain Energy Trading Tools,” munication Engineering (1979) at BIT, India, and M.Sc. (1986) and
IEEE Computer Applications in Power, Vol. 11, NO. 2, April Ph.D. (1991) degrees in Computing Science at University of Alberta,
1998, pp. 47–52. Canada. From 1980 to 1983, he worked at the R & D department of
Operations Research Group, the representative company for Sperry
Univac Computers in India. From 1990 to 1992, he was involved with
the design of telecommunication systems at Bell-Northern Research,
Ottawa. He is currently an Associate Professor at Department of
Electrical and Computer Engineering, University of Waterloo. Dr.
Singh has published several research papers in the areas of software
engineering, network computing, database systems, and artificial in-
telligence.

You might also like