Electrical Power and Energy Systems 31 (2009) 13–22

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

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

Electricity price forecasting in deregulated markets: A review and evaluation

Sanjeev Kumar Aggarwal 1, Lalit Mohan Saini 1, Ashwani Kumar *
Department of Electrical Engineering, National Institute of Technology, Kurukshetra, Haryana, India

a r t i c l e i n f o a b s t r a c t

Article history: The main methodologies used in electricity price forecasting have been reviewed in this paper. The fol-
Received 24 February 2007 lowing price-forecasting techniques have been covered: (i) stochastic time series, (ii) causal models, and
Received in revised form 25 August 2008 (iii) artificial intelligence based models. The quantitative analysis of the work done by various authors has
Accepted 14 September 2008
been presented based on (a) time horizon for prediction, (b) input variables, (c) output variables, (d)
results, (e) data points used for analysis, (f) preprocessing technique employed, and (g) architecture of
the model. The results have been presented in the form of tables for ease of comparison. Classification
Keywords:
of various price-influencing factors used by different researchers has been done and put for reference.
Price forecasting
Stochastic time series
Application of various models as applied to different electricity markets is also presented for
Regression models consideration.
Neural networks Ó 2008 Elsevier Ltd. All rights reserved.
Deregulated markets

1. Introduction uncertainties [2]. Therefore price-forecasting tools are essential

for all market participants for their survival under new deregulated
Under restructuring of electric power industry, different partic- environment. Even accurate load forecasts cannot guarantee prof-
ipants namely generation companies and consumers of electricity its and the market risk due to trading is considerable because of ex-
need to meet in a marketplace to decide on the electricity price treme volatility of electricity prices.
[1]. In the current deregulated scenario, the forecasting of electric- Seeing the importance of price forecasting, authors felt that
ity demand and price has emerged as one of the major research there is a need of comprehensive survey at one place so that future
fields in electrical engineering [2]. A lot of researchers and acade- researchers in this area can easily get information about the cur-
micians are engaged in the activity of developing tools and algo- rent state of the research. Although, a few attempts have already
rithms for load and price forecasting. Whereas, load forecasting been made in this direction [5–9], but only qualitative aspects of
has reached advanced stage of development and load forecasting price forecasting have been addressed, and there is no such work
algorithms with mean absolute percentage error (MAPE) below that has presented the quantitative analysis of the papers pub-
3% are available [3,4], price-forecasting techniques, which are lished so far. Authors of [5] documented the key issues of electric-
being applied, are still in their early stages of maturity. In actual ity price modeling and forecasting, and reviewed the price models
electricity markets, price curve exhibits considerably richer struc- adapted from financial assets. Importance of the price-forecasting
ture than load curve [5] and has the following characteristics: high problem, key issues and some techniques developed so far have
frequency, nonconstant mean and variance, multiple seasonality, been reported in [6]. Various price-forecasting techniques from in-
calendar effect, high level of volatility and high percentage of unu- put and output variables perspective have been discussed and
sual price movements. All these characteristics can be attributed to comparison of results of five different techniques has been pre-
the following reasons, which distinguish electricity from other sented in [7]. Time series forecasting procedures have been dis-
commodities: (i) non-storable nature of electrical energy, (ii) the cussed in [8]. An overview of price-forecasting papers and
requirement of maintaining constant balance between demand general procedure for price forecasting has been presented in [9].
and supply, (iii) inelastic nature of demand over short time period, So an overall assessment of the price-forecasting algorithms is still
and (iv) oligopolistic generation side. In addition to these, market required.
equilibrium is also influenced by both load and generation side In order to investigate the state of price-forecasting methodol-
ogies, a review of 47 papers published during 1997 to November
2006, has been done based on the following parameters: (i) type
* Corresponding author. Tel.: +91 9416366091; fax: +91 1744 238050.
E-mail addresses: [email protected] (S.K. Aggarwal), lmsaini@rediffmail.
of model, (ii) time horizon for prediction, (iii) input variables
com (L.M. Saini), [email protected] (A. Kumar). used, (iv) output variables, (v) analysis of results, (vi) data points
1
Tel.: +91 9416828819/9416137773. used for analysis, (vii) preprocessing employed, and (viii) model

0142-0615/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2008.09.003
14 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22

architecture. All the papers have been classified in three main librium model. Study of game theory models in itself is a major
categories. It has been observed that forecasting errors are still area of research and has been kept outside the scope of this pa-
high from risk management perspective and test results are per. A detailed discussion on game theory models can be found
difficult to compare with each other. Most of the electricity in [11].
markets are at the stage of infancy, so the researchers have to
make predictive analysis based on the small data set available with 2.2. Simulation models
them. Little work has been done in the direction of price spike pre-
diction. All the work has been quantified and put in the form of These models form the second class of price-forecasting tech-
tables. niques, where an exact model of the system is built, and the solu-
This paper is organized as follows: In Section 2, a short intro- tion is found using algorithms that consider the physical
duction to price-forecasting methodologies is given. In Section 3, phenomenon that governs the process. Then, based on the model
factors affecting electricity prices, as considered by different and the procedure, the simulation method establishes mathemati-
authors in their respective models, have been classified and dis- cal models and solves them for price forecasting. Price forecasting
cussed. The various features of the time series and causal models by simulation methods mimics the actual dispatch with system
have been outlined in Section 4. Artificial intelligence (AI) based operating requirements and constraints. It intends to solve a secu-
methods have been elaborated in Section 5. Data-mining models rity constrained optimal power flow (SCOPF) with the entire sys-
are discussed in Section 6. Section 7 deals with locational marginal tem range. Two kinds of simulation models have been analyzed
price (LMP) forecasting models. Section 8 deals with the different in this paper. One is market assessment and portfolio strategies
techniques as they are applied by the researchers to different elec- (MAPS) algorithm developed by GE Power Systems Energy Consult-
tricity markets. Discussion and key issues are given in Section 9. ing [12] and the other is UPLAN software developed by LCG Con-
Section 10 concludes the review. sulting [13].
MAPS is used to capture hour-by-hour market dynamics while
2. Price-forecasting methodologies simulating the transmission constraints on the power system. In-
puts to MAPS are detailed load, transmission and generation units’
Numerous methods have been developed for electricity price data. Where as the outputs are complete unit dispatch information,
forecasting and most of these algorithms are same as used for load LMP prices at generator buses, load buses and transmission flow
forecasting and especially short-term load forecasting (STLF). Time information. UPLAN, a structural multi-commodity, multi-area
horizon varies from hour ahead to a week ahead forecasting. The optimal power flow (MMOPF) type model, performs Monte Carlo
price-forecasting models have been classified in three sets [10] simulation to take into account all major price drivers. UPLAN is
and these three sets have been further divided into subsets as used to forecast electricity prices and to simulate the participants’
shown in Fig. 1. behavior in the energy and other electricity markets like ancillary
service market, emission allowance market. The inputs to MMOPF
2.1. Game theory models are competitive bidding behavior, generation units’ data, the trans-
mission network data, hydrological conditions, fuel prices and de-
The first group of models is based on game theory. It is of great mand forecasts. These are almost comparable to the input
interest to model the strategies (or gaming) of the market partic- variables of MAPS. The outputs are forecast of prices and their
ipants and identify solution of those games. Since participants in probability distribution across different energy markets. The dy-
oligopolistic electricity markets shift their bidding curves from namic effect of drivers on market behavior has also been captured.
their actual marginal costs in order to maximize their profits, Both UPLAN and MAPS may be used for long as well as short range
these models involve the mathematical solution of these games planning.
and price evolution can be considered as the outcome of a power Simulation methods are intended to provide detailed insights
transaction game. In this group of models, equilibrium models into system prices. However, these methods suffer from two draw-
take the analysis of strategic market equilibrium as a key point. backs. First, they require detailed system operation data and sec-
There are several equilibrium models available like Nash equilib- ond, simulation methods are complicated to implement and their
rium, Cournot model, Bertrand model, and supply function equi- computational cost is very high.

Electricity Price Forecasting

Models

Game Theory Time Series

Models Simulation Models
Models

Parsimonious Artificial Intelligence Regression or Causal

Stochastic Models based models Models

Neural Network
Data-mining Models
based Models

Fig. 1. Classification of price-forecasting models.

 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22 15

2.3. Time series models the basis of correlation analysis on each of these independent vari-
ables with the price (dependent) variable.
Time series analysis is a method of forecasting which focuses on
the past behavior of the dependent variable [14]. Sometimes exog- 2.3.3. Artificial intelligence (AI) models
enous variables can also be included within a time series frame- These may be considered as nonparametric models that map
work. Based on time series, there are further three types of models. the input–output relationship without exploring the underlying
process. It is considered that AI models have the ability to learn
2.3.1. Parsimonious stochastic models complex and nonlinear relationships that are difficult to model
Many stochastic models are inspired by the financial literature with conventional models. These models can be further divided
and a desire to adapt some of the well known and widely applied into two categories: (i) artificial neural network (ANN) based mod-
in practice approaches. In this paper, univariate discrete type mod- els and (ii) data-mining models.
els like autoregressive (AR), moving average (MA), autoregressive
moving average (ARMA), autoregressive integrated moving aver- 2.3.3.1. ANN based models. ANNs are able to capture the autocor-
age (ARIMA), and generalized autoregressive conditional heteros- relation structure in a time series even if the underlying law
kedastic (GARCH) have been considered. These are discrete time governing the series is unknown or too complex to describe.
counterparts corresponding to the continuous-time stochastic Since quantitative forecasting is based on extracting patterns
models. Purely finance-inspired stochastic models involving cer- from observed past events and extrapolating them into the fu-
tain characteristics of electricity prices, like price spikes and mean ture, thus ANN may be assumed to be good candidates for this
reversion, have been kept outside the scope of this review. A dis- task [16]. The available NN models are: (i) multilayer feed for-
cussion on these models can be seen in [5]. ward NN (FFNN), (ii) radial basis function network (RBF), (iii)
Stochastic time series can be divided into stationary process support vector machine (SVM), (iv) self-organizing map (SOM),
and non-stationary process. The basic assumption of stationarity (v) committee machine of NNs, and (vi) recurrent neural network
on the error terms includes zero mean and constant variance. In (RNN).
AR, MA and ARMA models conditions of stationarity are satisfied;
therefore they are applicable only to stationary series. ARIMA mod- 2.3.3.2. Data-mining models. Recently, data-mining techniques like
el tries to capture the incremental evolution in the price instead of Bayesian categorization method, closest k-neighborhood categori-
the price value. By the use of a difference operator, transformation zation, reasoning based categorization, genetic algorithm (GA)
of a non-stationary process into a stationary process is performed. based categorization, have gained popularity for data interpreta-
The class of models where the constant variance assumption does tion and inferencing. All those models using data-mining tech-
not need to hold is named heteroskedastic. Thus GARCH model niques have been covered in the category of data-mining models
considers the conditional variance as time dependent. In all these in this work.
models price is expressed in terms of its history and a white noise
process. If other variables are affecting the value of price, the effect 3. Factors influencing electricity prices
of these variables can be accounted for using multivariate models
like TF (transfer function) and ARMA with exogenous variables The factors influencing spot prices may be classified on the basis
(ARMAX) models. As electricity price is a non-stationary process, of: C1 – market characteristics, C2 – nonstrategic uncertainties, C3
which exhibits daily, weekly, yearly and other periodicities. There- – other stochastic uncertainties, C4 – behavior indices, and C5 –
fore, a different class of models that have this property, designated temporal effects. The different input variables, along with the class
as seasonal process model, is used. they belong to, used by different researchers are presented in
Table 1. There are as many as 40 variables used by different
2.3.2. Regression or causal models researchers. Most of the researchers have utilized past experience
Regression type forecasting model is based on the theorized in selecting the input variables for their respective model and
relationship between a dependent variable (electricity price) and choice of best input variables for a particular model is still an open
a number of independent variables that are known or can be esti- area of research.
mated [15]. The price is modeled as a function of some exogenous The widely used input variable is the electricity price of previ-
variables. The explanatory variables of this model are identified on ous days. Researchers have used as much as past 1–7, 14, 21, 28,

Table 1
Factors influencing electricity prices

Class Input variable Time period whose data is used as input

C1 (1) Historical load f(load); (d  m, t), m = 1, 2, 3, 4, 7, 14, 21, 28
(2) System load rate, (3) imports/exports, (4) capacity excess/shortfall (d, t), (d, t  1), (d  1, t), (d  2, t), (d  7, t)
(5) Historical reserves (d, t  2), (d, t  1), (d, t)
(6) Nuclear, (7) thermal, (8) hydro generation, (9) generation capacity, (10) net-tie flows, (11) MRR, (12) (d, t)
system’s binding constraints, (13) line limits
(14) Past MCQ (market-clearing quantity) (d  1, t)
C2 (15) Forecast load (d, t  2), (d, t  1), (d, t)
(16) Forecast reserves, (17) temperature, (18) dew point temperature, (19) weather, (20) oil price, (21) gas (d, t)
price, (22) fuel price
C3 (23) Generation outages, (24) line status, (25) line contingency information, (26) congestion index (d, t)
C4 (27) Historical prices f(price); (d  m, t  n), m = 0, 1, 2, 3, 4, 5, 6, 7, 8, 14, 21,
28, 364 and n = 0, 1, 2, 3, 4.
(28) Demand elasticity, (29) bidding strategies, (30) spike existence index, (31) ID flag (d, t)
C5 (32) Settlement period, (33) day type, (34) month, (35) holiday code, (36) Xmas code, (37) clock change, (d, t)
(38) season, (39) summer index, (40) winter index

C1 – market characteristics, C2 – nonstrategic uncertainties, C3 – other stochastic uncertainties, C4 – behavior indices, C5 – temporal effects, d – day, t – settlement period
number of the day.
Note: The serial number of input variables given here are used in the input variable column of Tables 2 and 4 for respective input variables used by different researchers.
16 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22

364 days price lags to capture the complete seasonal/calendar 4. Methodologies based on stochastic time series and causal
variations namely daily, weekly and yearly variations. As price is models
strongly correlated with demand, next most often used input var-
iable is demand. Most authors have used the projected demand of Twelve research papers can be covered in this category, three
independent system operator (ISO), of the concerned electricity are causal models [19,28,29] and nine are stochastic time series
market, as an input variable, a few have predicted the demand first models [30–38]. In Ref. [28], a regression-based model for electric-
and then used it as input variable for the price-forecasting model ity price was derived based on the assumption that power con-
[17,18]. Many researchers have also used historical load data as in- sumption and market prices are stochastic processes. Statistical
put variables. Authors of Ref. [19] have included functions of price results for price model coefficients were shown. Vucetic et al.
and forecasted load in their model. Instead of load, Li and Wang [19] have assumed a piece-wise stationary price time series having
[20] have used system load rate (SLR) as input variable so as to multiple regimes with stable price–load relationship in each re-
take the effect of the rate at which load is changing on the output. gime. These regimes, in the price series, were discovered with
Capacity excess or surplus is total available capacity minus the re- the help of a regime discovery algorithm and price was modeled
quired capacity at peak hour. This has been used by most of the by applying separate regression model for each regime. In Ref.
researchers as input variable, because it may affect the price sig- [29,35,36] the price series has been decomposed into detailed
nificantly in case surplus goes below certain threshold level and and approximation parts using wavelet transform (WT). Future
thereby prompting some major participants to utilize this period behavior of decomposed series was predicted by applying appro-
as an opportunity to exercise their market power. Since tempera- priate model in the wavelet domain and finally inverse WT was
ture is the main exogenous variable that affects the system load, used to generate price prediction in the time domain.
authors of Ref. [18,21,22] have used temperature as input variable In Ref. [29], second order regression polynomial of forecasted de-
in their respective models. To take the effect of inflation and cost mand was applied to predict detailed components. In Ref. [35], the
of fuel prices on electricity price, fuel and oil prices have also been future behavior of all the constitutive series has been predicted by
used as input variables in [21–24]. Must run ratio (MRR) is the applying ARIMA to each of the series. In Ref. [36], both load and price
generation concentration index (an indicator of oligopolistic series were decomposed. Price and historical load data’s approxi-
nature of the market), which has been used as an input variable mate part was used by multivariate time series for price approxi-
in [25,26] and has been shown to have considerable impact on mate coefficients prediction and price detail coefficient part used
market price. by univariate time series to forecast price detail coefficients. No-
In order to understand the market state, instead of overall gen- gales et al. [30] have developed two models. The first one was a dy-
eration capacity, Gonzalez et al. [10] have used hourly production namic regression (DR) model, which relates spot price to its own
capacity of different technologies like hydro, thermal and nuclear lagged values and actual demand values. In the TF model, the rela-
as input variables and also reported the use of input variables like tionship between price and demand has been established through
participants’ pricing strategies, production costs, aggregated sup- a TF term and a disturbance term that follows an ARMA process.
ply functions, generation companies’ shares etc. in their model In Ref. [31], an ARIMA based forecasting model was presented. Cuar-
but no improvement in the accuracy was observed. ID flag, used esma et al. [32] have demonstrated a comparison of the perfor-
in [27], is an indicator for presence of peak price (price volatility) mance of 50 linear univariate time series AR and ARMA models.
in the neighborhood of the predicted settlement period. Multiple Seasonal process ARIMA model has been proposed in Ref.
seasonalities related to daily, weekly, monthly and yearly periodic- [33,34]. In both these papers, a stationary price time series has
ities have also been utilized as input variables as is evident from been obtained by filtering out non-periodic trend component and
the Table 1. Kian and Keyhani have used demand elasticity and periodical component and then price profile has been predicted
bidding strategies as explanatory variables of a regression-based by applying ARIMA. Further accuracy improvement has been
model [28]. achieved through successive error correction method. In Ref.

Table 2
Main characteristics of time series models

Paper Model type Input variables (serial numbers as per Variable Preprocessing employed Model identification Parameter
column 2 of Table 1) segmentation and validation estimation
[29] Second order 27, 15 SS WT – –
polynomial
[32] AR, ARMA 27 SS, 24 hourly – – –
Series
[37] ARMA, ARMAX, 27, 1, 15 24 hourly LT, Normalization – MLF
AR, ARX series
[38] GARCH 27, 1 SS LT ACF, PACF MLF
[30] (1) DR, (2) TF 1, 27 SS LT, outliers have been removed ACF, PACF MLF
[31] ARIMA 1, 27, 8 SS LT ACF, PACF MLF
[33] Seasonal process 27 SS Elimination of periodic and non- ACF, PACF RA
periodic trend component
[34] Seasonal process 27 SS Elimination of periodic and non- ACF, PACF RA
periodic trend component
[35] ARIMA 27 SS WT ACF, PACF MLF
[36] Multivariate 27, 15 SS WT – –
ARMA
[28] MLR 17, 22, 27, 15, 9, 28, 29 SS – Statistical tools LSE
[19] Nonlinear 27, 15, f(price), f(load) SS – – LSE
regression

DR, dynamic regression; LT, log transformation; LSE, least square estimation; MLF, maximum likelihood function; MLR, multiple linear regression; RA, regression analysis; SS,
single series; TF, transfer function; WT, wavelet transform.
 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22 17

Table 3 Authors of Ref. [42] initially reported the use of FFNN in price
Forecasting performance comparison of time series models profile forecasting that tried 12 different combinations. In Ref.
Paper Data Predicted period Level of accuracy Time Output [27], the raw price data was pre-processed by a front-end proces-
used horizon sor (based on fuzzy logic) representing the features of Saturday,
(days) Sunday and public holidays. The predictor was a FFNN trained by
[29] 1 1 week DMAPE 2.5–11.11% 1 DA PP BP that predicted price profiles corresponding to weekends and
[32] 442 45 days WMAE 3–7%. 1–7 DA PP public holidays. Szkuta et al. have also used BP trained FFNN
[37] 272 4 weeks WMAPE 3–11.1% 1 DA PP
[38] 147, 105 12 months WMAPE 9–11% 1 DA PP
[43]. Different FFNN models for weekdays and weekends were pre-
[30] 81, 135, 2 weeks, 1 week DMAPE 3–5% 1 DA PP sented in [23]. Yao et al. [44] initially utilized WT for the decompo-
92 sition of price and load data series into detailed and approximation
[31] 145, 85, 3 and 11 weeks, WMAPE 8–20% 1 DA PP parts and then RBF network was used for predicting the approxi-
73, 92 1 and 3 weeks (average 11%)
mate part and whereas, the detailed part was predicted by a
[33] 50, 50 2 sets of 10 days MaxAE 1.21–4.36, 36– 1 DA AvP
58 weighted average method. In Ref. [45], NN model was used for pre-
[34] 10 2 different days DPE 1.5%. Hourly PE 1h PP, CI dicting the price and fuzzy model predicted price ranges using lin-
0.1–5.23%. ahead ear programming. Two models, applicable only to working days,
[35] 48 4 weeks of 4 WMAPE 5–27% 1 DA PP were compared in Ref. [41]. First one is a RNN model and second
seasons
is a k-weighted nearest neighbor (kWNN) algorithm based model,
[36] 27 1 week DAPE min 0.1–5.3%, 1 DA PP
DAPE max 52.2–98.7% which utilized a weighted-Euclidian norm to find days, which are
[28] 28 3 days – 1, 2, 3 PP nearest, in certain characteristics, to the forecast day. Genetic algo-
DA rithm (GA) was used for estimating the weights corresponding to
[19] 540 – MSE 53.7–93.9, R2 0.8– 1 DA PP
different nearest neighbors. Zhang et al. [46] have implemented a
0.68
cascaded NN structure using a non-cascaded FFNN with the pre-
AE, absolute error; MAE, mean absolute error; DMAE, daily MAE; WMAE, weekly dicted input (load and weather) expressed as the measured input
MAE; MAPE, mean absolute percentage error; DMAPE, daily MAPE; WMAPE, weekly
plus an additional error term representing the associated uncer-
MAPE; PE, percentage error; DPE, daily PE; APE, absolute percentage error; DAPE,
daily APE; MPE, mean percentage error; MSE, mean square error; RMSE, root mean
tainties. Gaussian RBF networks approximate input–output rela-
square error; AvP, average price; CI, confidence interval; DA, day ahead; PP, price tionships by building localized clusters and since unimportant
profile; R2, coefficient of determination. input factors may mislead local learning of RBF networks and
thereby poor generalization, therefore, a two-step training method
based on the inverses of standard deviations to identify and elim-
[34], in addition to price profile, confidence interval (CI) of price for inate unimportant input factors was developed in Ref. [22].
that period was also predicted. ARMA and ARMAX models were Chaos theory was applied to construct a phase space from past
used in [37]. In Ref. [38], univariate ARMA model with GARCH error data of electric price and load in Ref. [47] and RNN was used for
components was utilized and GARCH model with demand as exog- prediction. Rodriguez and Anders [48] have proposed a hybrid of
enous variable was also developed. NN and fuzzy logic known as adaptive-network-based fuzzy infer-
Main characteristics of different time series models are given in ence system (ANFIS) in which the output was obtained as a linear
Table 2 and forecasting performance comparison has been pre- combination of the input membership values of the input variables
sented in Table 3. It can be observed that log transformation of and the inputs. An adaptively trained NN has been proposed whose
the price time series has been adopted as a preprocessing tech- architecture can be changed during learning phase [40]. Input fac-
nique in most of the stochastic time series models. This has been tors for the NN were obtained using a price simulation method.
done to obtain more stable variance. The idea of variable segmen- Authors in [25,26] have used NN model for short-term forecasting
tation, i.e., framing the model as 24 separate hourly series, has been and a linear regression model for long term forecasting. The predic-
applied in Ref. [32,37]. It has been observed that an hour-by-hour tion of spot price was done in Ref. [39] using the method of nonlin-
modeling strategy for electricity spot prices improves significantly ear auto-correlated chaotic model, whose parameters were
the forecasting abilities of linear univariate time series models predicted based on a wavelet NN (WNN) having hidden layer with
[32]. Autocorrelation function (ACF) and partial autocorrelation wavelet function. To overcome the inadequacy of a single network,
function (PACF) are the preferred choice of researchers for model committee machine consisting of RBF and MLP networks has been
identification and estimation. Maximum likelihood function presented in Ref. [21]. Instead of simple averaging the outputs of
(MLF) is the most widely used parameter estimation technique. different networks, the method used the current input data and
the historical data to calculate weighting coefficients, for combin-
5. Neural network-based models ing predictions of different networks, in a weight calculator. Gonz-
alez et al. [10] have proposed a switching model based on the
In this category, 17 researchers have forecasted the price pro- input–output hidden Markov model (IOHMM) framework. The
file, while six have made point prediction like maximum price or model was based on the premises that each market may be repre-
average price and in one paper [39] the parameters of a chaos mod- sented by two states, one of them is hidden state, (characterized by
el have been forecasted. Authors of [17,40,41] have used 24 output the interaction among resources, demand, and participants strate-
nodes and all other papers have used one output node. Information gies), and the visible state, the power price series. NN has been
for NN models is given in Tables 4–6. In Table 4, information used to model state subnetwork and a dynamic regression process
regarding model used, preprocessing employed and input variables of input variables has been used to implement output subnetwork.
used by the different researchers has been presented. Forecasting Extended Kalman filter (EKF) learning has been used to train
performance comparison has been given in Table 5. NN models’ MLP networks by treating weights of a network as the state of an
architecture information and data used in different models, has unforced nonlinear dynamic system in Ref. [24]. By ignoring the
been compared in Table 6. It is evident from the Table 4 that the interdependencies of mutually exclusive weights from different
FFNN architecture, which is also known as multilayer perceptron neurons, a significantly lower computational complexity and stor-
(MLP), along with back propagation (BP) as the learning algorithm age per training instance was achieved. Rough set theory (RST) has
is the most popular choice among researchers for a price-forecast- been applied to the input data pattern in order to group and
ing problem. combine the similar data patterns in Ref. [49] and the resulting
18 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22

Table 4
Neural network models’ input variables and preprocessing employed

Paper NN model Learning Input variables (serial numbers as per Total number of input Preprocessing technique
algorithm column 2 of Table 1) factors
[43] MLP BP 27, 15, 16, 32, 33, 35, 34, 37, 36 15 –
[48] (1) MLP, (2) (i) BP, (ii) LM 15, 23, 4, 3 1 Outliers removed
FMLP
[27] MLP BP 27, 15, 32, 31 9, 9, 6 Feature extraction for different days. SR [0, 1]
[42] MLP BP 27, 15, 4, 32, 33 12 –
[23] MLP BP (CG) 27, 15, 3, 14, 22, 19, 32, 33, 38 9 SR [0, 1], outliers removed
[51] MLP BP 27, 33, 34 5 Preprocessing using NN to forecast max, min,
medium values of prices
[17] MLP BP 27, 1, 15 1 –
[18] MLP BP 27, 1, 15, 17, 32, 33 1 Similar days data using Euclidian norm
[40] MLP BP 27, 1, 15, 5, 16, 32, 33, 24, 26, 13 – Outliers removed
[25] MLP BP 27, 15, 11 – –
[26] MLP BP 27, 15, 11 – –
[41] MLP BP 27 – –
[20] MLP AFSA 27, 2 6 WT, variable segmentation
[10] MLP, DRM IOHMM 27, 1, 15, 6, 7, 8 MLP – 5, DRM – 4 –
[49] MLP BP 27, 1 32 Similar patterns found using RST
[50] MLP GDR 27 19 –
[39] MLP BP 27 – Noise filtration using Fourier wave filter
[47] RNN – 27, 1, 15 17 –
[44] RBF – 27, 1, 33 12 WT, different model for each weekday
[45] MLP LM 27 – SR [0.1–0.9]
[52] MLP BP 27 4 ACF
[21] CM – 27, 1, 15, 4, 17, 20, 21 RBF – 23, MLP – 55 SR [0, 1]
[46] MLP BP, QN for CI 27, 1, 15, 4 56 –
[24] MLP EKF 27, 15, 4, 20, 21, 33, 35 50 –
[22] RBF 2 stage training 15, 4, 17, 20, 21, 33 23 SR [0, 1]

AFSA, artificial fish swarm algorithm; ACF, autocorrelation function; BP, back propagation (first order gradient learning algorithm); CG, conjugate gradient; CI, confidence
interval; CM, committee machine; DRM, dynamic regression model; EKF, extended Kalman filter; FMLP, fuzzy MLP; GDR, generalized delta rule; IOHMM, input–output
hidden Markov model; LM, Levenberg Marquardt algorithm; MLP, multilayer perceptron; QN, quasi-Newton; RBF, radial basis function; RST, rough set theory; SR, scaling
range; WT, wavelet transform.

Table 5
Forecasting performance comparison of neural network models

Paper Output Training data (days) Predicted period Time horizon Level of accuracy
[43] PP 203 1 week 1 time period ahead Daily AvE 2.18–11.09
[48] PP 14/28 1 day, 30 days 1 DA DMAPE 20–38%.
[27] PP 180, 180, 2 60 days, 60 days 1 DA DMAPE 8.93–12.19%
[42] PP 77 1 week 1 DA Average DMAPE 11.57–12.86%
[23] PP, QP 363, 404, 131 1 month 1 DA DMAE 1.19–1.76 (training and validation only)
[51] PP 1095, 730 2 sets of 2 days 1, 2, 3 DA Error less than 1c€ in 85% cases
[17] LP, PP 28 2 different weeks 1 DA MAPE without spike 8.44%, with spike 15.87%
[18] LP, PP 90 1 week, 1 month 1–6 h ahead WMAPE 10.69–25.77%, monthly MAPE 9.75–20.03%.
[40] PP, zonal PP, PDF 7–56 1 week Short term WMAPE 11–13%
[25] PP, CI – 1 week 1 DA DMAPE 10–20%
[26] PP, CI – 1 week 1 DA WMAPE 15.5%
[41] PP 20 2 sets of 3 months 1 DA RNN: AvPE 12–15%, kNN: AvPE 9–11%
[20] PP – 1 week 1 DA DMAPE 3.5–5.16%
[10] PP, PDF 182 1 week, 92 days 1 h ahead WMAPE 15.83%
[49] PP 30 1 month 1 h ahead DMAPE 6.04%
[50] PP 48 4 weeks of 4 seasons 1 DA Average weekly MAPE = 7.5%
[39] PCM, PP – 10 days 1 time period ahead APE 8%
[47] PP – 2 days 1, 25, 49 h ahead DMPE 2.22–8%
[44] PP 60 1 week 1 DA Average AE 4–7.5%
[45] Max Price, Range 60 – 1 DA Overall RMSE 9.23%
[52] AvP 228, 221, 214, 207, 144 7–91 days m – ahead, m = 7, 14, 21, 28, Average MAPE 8.22–9.12%
91
[21] OPHAP 365 6 months 1 DA Monthly MAPE 7.74–19.85%
[46] OPHAP, C.I. 426 5 months 1 DA Average monthly MAPE 8.8%, one-sigma CI coverage
66.6%
[24] OPHAP, C.I. 427 11 and 2 months 1 DA MAPE 11.1%, one-sigma CI coverage 68%.
[22] OPHAP 427 12 months 1 DA Average MAPE 11.9%

AvP, average price; AvE, average error; AvPE, average percentage error; DA, day ahead; LP, load profile; OPHAP, on-peak hour average price; PP, price profile; PCM, parameters
of chaotic model; PDF, probability density function; QP, quantity profile; SDE, standard deviation of error; WMPE, weekly mean percentage error.
Note: Abbreviations of all accuracy criterion are same as Table 3.

patterns were used to train the NN. A fuzzy neural network (FNN) price series) of the input space has been performed in hidden layer
having higher learning capability has been proposed in Ref. [50]. In and defuzzification process in the single node of the output layer.
FNN, the fuzzyfied classification process (internal decomposition of In Ref. [18], historical days that are similar in nature to a forecast
 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22 19

Table 6 ing module separates the two price signals. The NN-wavelet mod-
Neural network models’ architecture ule predicts the normal price and the possibility of price spikes at
Paper No. of neurons Activation No. of parameters Settlement specific occasions. If a specific occasion is forecasted to have price
function periods spike then spike-forecasting module is activated. The range of price
[43] [15-15-1] * 241 48 spike can be predicted through data-mining techniques such as
[48] [(1,2)-(4,8,12)-(1)] S/L, FMF/L ANN – 13 to 49, 24 categorization algorithm. The value of price spike is estimated
ANFIS – 24 using k-nearest neighboring approach. Hourly PE (percentage er-
[27] [9-7-4-1], [9-7-4-1], [6- S/S 107, 107, 33 48
4-1]
ror) varied from 1% to 31% in most of the test cases, whereas in
[42] [12-8-5-1] S/S 155 48 one case it was reported to be 49%. In [56], SVM and probability
[23] [4-6-2] TS/L 44 24 based classification algorithm was combined with normal price-
[51] [1-2:4-1:14], [10-3-1] * 243 24
*
forecasting method to determine probability of price spike. Then
[17] [72-15-24], [48-15-24] 1479, 1119 24
a Bayesian classifier was used for range of the forecasted price
[18] [5/10-*-1], [4/9-*-1] * * 48
[40] [25-40-24], [73-100- * 2024, 9824, 24 spike and k-nearest neighboring approach for value of price spike.
24], [121-150-24] 21924 A spike existence index was used as an input variable to utilize the
*
[25] [3-2-1] 11 24 characteristic of price spikes that tend to occur together in a short
[26] [3-2-1] * 11 24
* period. Forecasted hourly PE varies from 5.47% to 20% in most of
[41] [24-24-24] 1200 24
* * * the cases, whereas in one case it is reported to be 49%. A hybrid nu-
[20] 24
[10] [5-*-1] S/L * 24 meric method that integrates a Bayesian statistical method and a
[49] * * * – Bayesian expert (BE) has been proposed in Ref. [57] for price spike
[50] [19-19-1] FC/L 1102 24 prediction. Price series was classified into three classes of price
[39] * WF/* * 48
* * * spikes, normal price and lower price using Bayesian classification
[47] 24
[44] [12-*-1] GRBF * 48 approach and a BE combined with SVM has been used to forecast
[45] [7-7-1] L/L 64 24 electricity price spikes, normal price and lower price.
[52] [4-7-1] S/L 43 *
[21] 6 clusters, [55-8-1] * 283, 457 24
[46] [56-8-1] * 465 24 7. Methods for LMP prediction
[24] [50-*-1] * 500 24
[22] 6 clusters * 283 24
In a power system, when the available least-cost energy cannot
, not reported; DA, day ahead; FC, fuzzified classified function; FMF, fuzzy mem- be delivered to load in a transmission-constrained area, higher-cost
bership function; GRBF, Gaussian radial basis function; L, linear function; S, sigmoid generation units have to be dispatched to meet that load. In this sit-
function; TS, tan sigmoid; WF, wavelet function.
uation, the price of energy in the constrained area is higher than the
day were identified based on a weighted-Euclidian norm method unconstrained market-clearing price (MCP). LMP is defined as the
and then a NN, trained with this similar days’ input data, forecast price of lowest-cost resources available to meet the load, subject
the price by modifying the price curve obtained by averaging three to delivery constraints of the physical network and is made up of
similar price days. A regression model has been applied to deter- three components, (i) energy cost component, (ii) transmission
mine weighted factors in weighted-Euclidian norm method. congestion component, and (iii) marginal loss component. The con-
In Ref. [51], preprocessing was done to forecast maximum, min- gestion and loss components are different for different locations
imum, medium values of the price using three auxiliary NNs and and the energy cost component is identical for all the nodes. For a
then five principal NNs were used to forecast hourly prices. market, there is only one system marginal price (SMP), whereas,
Georgilakis [17] has reported an adaptively trained MLP-BP, in there is an LMP involving the line flow constraints and other secu-
which main NN predicted the hourly prices using forecasted load rity constraints at each node/area in a market. Three works pertain-
information of an auxiliary NN. In Ref. [20], WT has been used to ex- ing to LMP forecasting have been considered [58–60]. In Ref. [58],
tract approximate price signals and then these signals fed to an arti- an EKF learning based NN model for forecasting zonal LMPs has
ficial fish swarm algorithm (AFSA) based NN to map influences of been reported. Congestion components have been estimated by
nonlinear factors. In Ref. [52], ACF has been applied to the price forecasting differences between zonal LMPs and hub LMPs. Quanti-
time series to find out correlation between different periods of fication of transmission outages has been done based on a heuristic
the series and NN was used for price prediction. A moving cross val- method to feed into NN as input. On-peak average day-ahead and
idation method was employed for finding the best architecture of on-peak real-time LMPs have been predicted in log form. Overall
the FFNN. MAPE varies from 8% to 20%. A fuzzy reasoning and RNN based
method has been proposed in Ref. [59]. Quantification of contingen-
6. Data-mining models cies has been done based on fuzzy rules to feed into NN as input in
the form of a variable called lmp (ratio of the LMP over the hub
Five papers have been considered in this section. Two working LMP). Three different RNN models for weekday, Saturday and Sun-
day models have been proposed in Ref. [53]. One of them is kWNN day have been used. In [58,59], one transaction period ahead LMP
algorithm combined with GA and the other is dynamic regression for an area has been predicted. In [60], a Fuzzy-c-means (FCM)
model in which least square estimation (LSE) method has been and RNN based method has been used. FCM is used to classify the
used for estimation of coefficients. A hybrid of Bayesian-based clas- transaction periods into three clusters according to load levels:
sification and AR method that does not need any training has been peak, medium and off-peak load. In total nine RNNs were developed
presented in [54]. In this, a clustering algorithm, other than the k- to forecast nine different combinations of three clusters according
means and the convergent k-means, has been used to predict out- to load and three classes based on type of day.
put probability density function (PDF) and an AR model captures
the output change trend. Monthly MAPE in this method varies 8. Forecasting methodologies from electricity markets
from 9.96% to 13.69%. In Ref. [55], normal price has been predicted perspective
using wavelet and NN based model and price spike using a data-
mining framework. Bayesian classification and similarity searching Researchers have developed various forecasting tools covering
techniques have been used to mine the database. The preprocess- most of the deregulated markets. In [61], model has not been
20 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22

Table 7 (iii) inelastic nature of electricity demand over short term. Authors
Price-forecasting research and electricity markets in Ref. [10] have tried to incorporate the effect of first reason on the
Serial Market Total Paper no. market’s hidden states by using generation levels of different tech-
no. papers nologies as input variables. In [25,26], MRR has been included as
1 PJM electricity 6 [12,20,36,58–60] an input variable to capture oligopolistic nature of the market.
market Garcia et al. [38] have shown the effect of market volatility on
2 California electricity 13 [13,17,19,23,28,30,31,33,34,37,38,40,45] the performance of ARIMA and GARCH models, whereas; effect of
market
3 New England 7 [21,22,24,46,47,54,58]
volatility on the performance of the other models has yet not been
electricity market reported adequately in literature.
4 Ontario electricity 1 [48] The selection of input feature is a key issue for the success of
market any forecasting technique. Principal component analysis (PCA),
5 Spanish electricity 8 [10,30,31,35,38,41,50,53]
correlation analysis, genetic algorithm (GA), sensitivity analysis,
market
6 Victoria electricity 2 [19,43] spectrum analysis techniques can be used for this purpose. Authors
market, NEM in Ref. [24,40,43] have performed sensitivity analysis to show the
7 Queensland 2 [55,56] effect of input variable variation on output. Ref. [48] has performed
electricity market feature selection using correlation analysis. An analytical method,
8 UK power pool 4 [27,29,42,44]
which can select minimum number of effective input features,
9 European energy 3 [32,51,52]
exchange (Leipzig) has yet not been reported.
10 Electricity markets of 3 [25,26,39,57] Most of the time series models are univariate models, whereas
China few models involving structural approach are available. The prob-
11 Korean power 1 [34]
lem with the time series models is the assumption of stationarity,
exchange
whereas price series exhibits a high degree of non-stationarity. It
can be observed from Section 4 that, research is moving in the
direction of development of more sophisticated hybrid and non-
applied to any market but the work is significant because it ex- stationary models involving some kind of preprocessing of data
plores the possibility of signal-processing techniques (Fourier in order to attain stable mean and variance for price series. Garcia
and Hartley transform) as a preprocessing and filtering tool to et al. [38] have developed a non-stationary time series model
bring out hidden patterns in the price signal. Authors of Ref. based on GARCH with reasonable degree of success.
[30,31,38,58] have tested their models on more than one electricity From risk management perspective, distribution of prices is
market, whereas, others have confined themselves to only one more important than the point prediction; six papers have covered
market. An analysis of power prices across 14 different electricity this aspect as shown in Tables 2 and 5.
markets has been outlined in Ref. [62] and shown how price There is no benchmark for checking the continued out-perfor-
evolution is different in different markets and therefore large mance of a single model over other models. Most of the results re-
variations exist in price-forecasting accuracy achieved by different ported by different researchers cannot be put in a single
models across different electricity markets. framework because of diversity in their presentations (Tables 3
Spanish electricity market [63], PJM [64] and New England elec- and 5). Authors of [50] have compared the performance of FNN
tricity market [65] are the markets, which have caught the atten- price-forecasting model with 6 other models and proved its
tion of most of the researchers. These are based on standard superiority. Persistence method [17,40] and naïve method [35]
market design (SMD) structure, which is basically a two-settle- are some of the reported methods, which can be used as bench-
ment market comprising a day-ahead market and a real-time intra- mark for testing the effectiveness of any new forecasting
day market. Most of the statistical models have been applied on methodology.
market data from these markets. Whereas; Ontario electricity mar- A price signal exhibits much richer structure that load series,
ket [66] and National electricity market of Australia (NEM) [67] fol- although signal-processing techniques, like WT, are good candi-
low a single settlement real-time structure. Only a few researchers dates for bringing out hidden patterns in price series after decom-
have applied their models on data from these markets. Apart from posing price series into better behaved signals, probability of loss
that, California electricity market [68] is also one of the largely of valuable information remains. An analytical method for noise re-
studied markets in the world for the well-known problems that moval is yet to be reported.
it faced in the second half of 2000. The information regarding No single available model has been applied across data from lar-
status of research in different electricity markets is presented in ger number of markets. There is little systematic evidence as yet
Table 7. that one model may explain the behavior of price signal in different
electricity markets, which is an indicator of participants’ collective
9. A discussion and key issues in designing a price-forecasting response to uncertainties, on a consistent basis.
system Four papers, [23,30,40,48] have presented the results of their
respective models with and without removing spikes and observed
Designing a price-forecasting model is a complex task as is evi- that prediction quality was improved by removing the outliers. On
dent from the literature review presented in previous sections. the other hand, price spike prediction is relatively a new and
Variations in input variable selection, forecasting horizon, prepro- important area because price spikes have the capacity to signifi-
cessing to be used, model selection, parameter estimation and cantly affect the profitability of both suppliers and customers. In
accuracy assessment have been reported, but few guidelines to three papers [55–57], emphasis has been given to predict price
help the new designer. The key issues involved in formulating a spike.
price-forecasting problem are as follows: Most of the researchers have concentrated on price forecasting
The electricity markets are highly volatile in nature and the rea- in day-ahead markets following SMD structure and case for real-
sons for spot price volatility are: (i) at any particular point in time, time electricity markets is relatively under investigated. There is
plants of different technologies with different heat rate curves are a need to make more research efforts in other markets as well; this
in operation making the aggregate supply–price curve complex will help in interpreting and understanding the price evolution in
and intraday variable in nature, (ii) oligopolistic supply side, and different electricity markets in a better perspective.
 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22 21

10. Conclusions [13] Deb R, Albert R, Hsue Lie-Long, Brown N. How to incorporate volatility and risk
in electricity price forecasting. Electricity J 2000(May):1–16.
[14] Box GEP, Jenkins GM, Reinsel GC. Time series analysis: forecasting and control.
An overview of different price-forecasting methodologies is pre- Pearson Education; 2004.
sented and key issues have been analyzed. Quantification of vari- [15] Moghram I, Rahman S. Analysis and evaluation of five short-term load
forecasting techniques. IEEE Trans Power Syst 1989;4(4):1484–91.
ous features of research papers pertaining to price forecasting
[16] Haykin S. Neural networks – a comprehensive foundation. 2nd ed. Prentice
has been done. Broadly short-term price-forecasting models con- Hall; 1999.
centrate on any of three output variables – average price, peak [17] Georgilakis PS. Market clearing price forecasting in deregulated electricity
markets using adaptively trained neural networks, SETN 2006. Lecture notes in
price, and price profile. Of these three types, price profile forecast-
artificial intelligence LNAI 3955. Berlin, Heidelberg: Springer-Verlag; 2006. p.
ing is more common and has been reported across different elec- 56–66.
tricity markets using different models. A mathematical model of [18] Mandal P, Senjyu T, Funabashi T. Neural networks approach to forecast several
price Y(t) can be represented in the form of the following equation: hour ahead electricity prices and loads in deregulated market. Energy Convers
Manage 2006;47:2128–42.
YðtÞ ¼ yp ðtÞ þ yðtÞ þ ys ðtÞ, where, yp(t) is a component which de- [19] Vucetic S, Tomsovic K, Obradovic Z. Discovering price – load relationships in
pends primarily on the time of day and on normal working condi- California’s electricity market. IEEE Trans Power Syst 2001;16(2):280–6.
tions of load and supply pattern for the particular day. The term [20] Li C, Wang S. Next-day power market clearing price forecasting using artificial
fish-swarm based neural networks, ISNN 2006. Lecture notes in computer
y(t) is an additive residual term describing influences due to load science LNCS 3972. Berlin, Heidelberg: Springer-Verlag; 2006. p. 1290–5.
and supply pattern deviations from normal and random correlation [21] Guo Jau-Jia, Luh PB. Improving market clearing price prediction by using a
effects. Usually such effects are small compared to time of day committee machine of neural networks. IEEE Trans Power Syst
2004;19(4):1867–76.
component when deviations are moderate from the normal condi- [22] Guo Jau-Jia, Luh PB. Selecting input factors for clusters of Gaussian radial basis
tions. The third component ys(t) depends upon the complex strat- function networks to improve market clearing price prediction. IEEE Trans
egies adopted by the participants and the most difficult to model. Power Syst 2003;18(2):655–72.
[23] Gao F, Cao X, Papalexopoulous A. Forecasting power market clearing price and
Different price-forecasting methodologies can be categorized in
quantity using a neural network method. IEEE Power Eng Soc Summer Meet
three major categories. Among these categories, a statistical model 2000;4:2183–8.
tries to capture the effect of price-influencing factors on price by [24] Zhang L, Luh PB. Neural network-based market clearing price prediction and
confidence interval estimation with an improved extended Kalman filter
analyzing the past data. Univariate models like ARIMA predict
method. IEEE Trans Power Syst 2005;20(1):59–66.
the future price values based on only price series data itself, [25] Hu Z, Yu Y, Wang Z, Sun W, Gan D, Han Z. Price forecasting using integrated
whereas; multivariate linear models like DR, TF and nonlinear approach. In: Proceedings of IEEE international conference on electric utility
models like ANN can consider the effect of exogenous variables deregulation, restructuring and power technologies. Hong Kong; April 2004. p.
28–31.
as well. The DR and TF algorithms have been found to be more [26] Hu Z, Yang L, Wang Z, Gan D, Sun W, Wang K. A game theoretic model for
effective than the ARIMA models. Although DR and TF models have electricity markets with tight capacity constraints. Electric Power Energy Syst
also shown good performance over nonlinear models in some case 2008;30(3):207–15.
[27] Wang A, Ramsay B. A neural network based estimator for electricity spot
studies, but in some cases ANN models have given better results. pricing with particular reference to weekend and public holidays.
Moreover, recent variations achieved in ANN and ARIMA using fuz- Neurocomputing 1998;23:47–57.
zy logic and WT hold promise as well. In conclusion, there is no [28] Kian A, Keyhani A. Stochastic price modeling of electricity in deregulated
energy markets. In: Proceedings of the 34th Hawaii international conference
systematic evidence of out-performance of one model over the on system sciences; 2001. p. 1–7.
other models on a consistent basis. This may be attributed to the [29] Chang-il Kim, In-Keun Yu, Song YH. Prediction of system marginal price of
reason that history of electricity markets is relatively short and electricity using wavelet transform analysis. Energy Convers Manage
2002;43:1839–51.
large differences in price developments exist in different power
[30] Nogales FJ, Contreras J, Conejo AJ, Espinola R. Forecasting next-day electricity
markets. It is hoped that with better computational tools at the dis- prices by time series models. IEEE Trans Power Syst 2002;17(2):342–8.
posal of researchers price evolution in electricity markets will be [31] Contreras J, Espinola R, Nogales FJ, Conejo AJ. ARIMA models to predict next-
day electricity prices. IEEE Trans Power Syst 2003;18(3):1014–20.
better understood over a period of time.
[32] Cuaresma JC, Hlouskova J, Kossmeier S, Obersteiner M. Forecasting electricity
spot-prices using linear univariate time-series models. Appl Energy
References 2004;77:87–106.
[33] Zhou M, Yan Z, Ni Y, Li G. An ARIMA approach to forecasting electricity price
[1] Schweppe FC, Caraminis MC, Tablors RD, Bohn RE. Spot pricing of with accuracy improvement by predicted errors. In: Proceedings of IEEE power
electricity. Kluwer Academic Publishers; 1988. engineering society general meeting, vol. 1; 6–10 June 2004. p. 233–8.
[2] Bunn DW. Forecasting loads and prices in competitive power markets. Proc [34] Zhou M, Yan Z, Ni YX, Li G, Nie Y. Electricity price forecasting with confidence-
IEEE 2000;88(2):163–9. interval estimation through an extended ARIMA approach. IEE Proc Generation
[3] Abdel-Aal RE. Modeling and forecasting electric daily peak loads using Trans Distribution 2006;153(2):187–95.
abductive networks. Electric Power Energy Syst 2006;28:133–41. [35] Conejo AJ, Plazas MA, Espinola R, Molina AB. Day-ahead electricity price
[4] Mandal P, Senjyu T, Urasaki N, Funabashi T. A neural network based several- forecasting using the wavelet transform and ARIMA models. IEEE Trans Power
hour-ahead electric load forecasting using similar days approach. Electric Syst 2005;20(2):1035–42.
Power Energy Syst 2006;28:367–73. [36] Haiteng XH, Niimura T. Short-term electricity price modeling and forecasting
[5] Bunn DW, Karakatsani N. Forecasting electricity prices. Review paper. London using wavelets and multivariate time series. In: IEEE power systems
Business School; 2003. conference and exposition PES, no. 1; 10–13 October 2004. p. 208–12.
[6] Amjady N, Hemmati M. Energy price forecasting – problems and proposals for [37] Weron R, Misiorek A. Forecasting spot electricity prices with time series
such predictions. IEEE Power Energy Mag 2006(March–April):20–9. models. In: Proceedings of international conference on the European
[7] Li G, Liu Chen-Ching, Lawarree J, Gallanti M, Venturini A. State-of-the-art of electricity market EEM. Lodz, Poland; 10–12 May 2005.
electricity price forecasting. In: International symposium of CIGRE/IEEE PES; [38] Garcia RC, Contreras J, Akkeren M, van Garcia JBC. A GARCH forecasting model
5–7 October 2005. p. 110–9. to predict day-ahead electricity prices. IEEE Trans Power Syst
[8] Contreras J, Santos JR. Short-term demand and energy price forecasting. In: 2005;20(2):867–74.
Proceedings of IEEE MELECON. Benalmadena, Malaga, Spain; 16–19 May 2006. [39] Wu W, Zhou Jian-Zhong, Yu J, Zhu Cheng-Jun, Yang Jun-Jie. Prediction of spot
p. 924–7. market prices of electricity using chaotic time series. In: Proceedings of the 3rd
[9] Niimura T. Forecasting techniques for deregulated electricity market prices. international conference on machine learning and cybernetics. Shanghai; 26–
Power Eng Soc General Meet IEEE 2006;18–22(June):1–6. 29 August 2004. p. 888–93.
[10] Gonzalez AM, San Roque AM, Garcia-Gonzalez J. Modeling and forecasting [40] Yamin HY, Shahidehpour SM, Li Z. Adaptive short-term electricity price
electricity prices with input/output hidden Markov models. IEEE Trans Power forecasting using artificial neural networks in the restructured power markets.
Syst 2005;20(1):13–24. Electrical Power Energy Syst 2004;26:571–81.
[11] Bajpai, P, Singh SN. Bidding and gaming in electricity market: an overview and [41] Lora AT, Santos Jose R, Santos Jesus R, Ramos JLM, Exposito AG. Electricity
key issues. In: Proceedings of national power system conference (NPSC), market price forecasting: neural networks versus weighted-distance k nearest
Chennai; 2004. p. 338–346. neighbors, DEXA 2002. Lecture notes in computer science LNCS 2453. Berlin,
[12] Bastian J, Zhu J, Banunaryanan V, Mukherji R. Forecasting energy prices in a Heidelberg: Springer-Verlag. p. 321–30.
competitive market. IEEE Comput Appl Power 1999(July):40–5. [42] Wang A, Ramsay B. Prediction of system marginal price in the UK power pool
using neural networks. Proc IEEE Int Conf Neural Networks 1997;4:2116–20.
22 S.K. Aggarwal et al. / Electrical Power and Energy Systems 31 (2009) 13–22

[43] Szkuta BR, Sanabria LA, Dillon TS. Electricity price short-term forecasting using [54] Li E, Luh PB. Forecasting power market clearing price and its discrete PDF using
artificial neural networks. IEEE Trans Power Syst 1999;14(3):851–7. a Bayesian-based classification method. In: Proceedings of the power
[44] Yao SJ, Song YH, Zhang LZ, Cheng XY. Prediction of system marginal price by engineering society winter meeting, IEEE 2001. p. 1518–23.
wavelet transform and neural network. Electric Mach Power Syst 2000;28: [55] Lu X, Dong ZY, Li X. Electricity market price spike forecast with data mining
537–49. techniques. Electric Power Syst Res 2005;73:19–29.
[45] Niimura T, Ko Hee-Sang, Ozawa K. A day-ahead electricity price prediction [56] Zhao JH, Dong ZY, Li X, Wong KP. A general method for electricity market price
based on a fuzzy-neuro autoregressive model in a deregulated electricity spike analysis. In: IEEE power engineering society general meeting, IEEE 2005,
market. In: Proceedings of the 2002 international joint conference on neural vol. 1. p. 1286–93.
networks, vol. 2, nos. 12–17; 2002. p. 1362–6. [57] Wu W, Zhou J, Mo L, Zhu C. Forecasting electricity market price spikes based on
[46] Zhang L, Luh PB, Kasivisvanathan K. Energy clearing price prediction and Bayesian expert with support vector machine, ADMA 2006. Lecture notes in
confidence interval estimation with cascaded neural networks. IEEE Trans artificial intelligence LNAI 4093. Berlin, Heidelberg: Springer-Verlag. p. 205–12.
Power Syst 2003;18(1):99–105. [58] Ma Y, Luh PB, Kasiviswanathan K, Ni E. A neural network based method for
[47] Yang H, Duan XZ. Chaotic characteristics of electricity price and its forecasting forecasting zonal locational marginal prices. In: Power engineering society
model. In: Proceedings of the IEEE Canadian conference on electrical and general meeting, IEEE 2004, vol. 1; 6–10 June 2004. p. 296–302.
computer engineering, CCECE 2003. p. 659–62. [59] Hong YY, Lee Chuan-Fang. A neuro-fuzzy price forecasting approach in
[48] Rodriguez CP, Anders GJ. Energy price forecasting in the Ontario competitive deregulated electricity markets. Electric Power Syst Res 2005;73:151–7.
power system market. IEEE Trans Power Syst 2004;19(3):366–74. [60] Hong YY, Hsiao CY. Locational marginal price forecasting in deregulated
[49] Lee Jeong-Kyu, Park Jong-Bae, Shin Joong-Rin, Lee KY. A system marginal price electricity markets using artificial intelligence. IEE Proc Generation Trans
forecasting based on an artificial neural network adapted with rough set Distribution 2002;149(5):621–6.
theory. In: Power engineering society general meeting, IEEE 2005, vol. 1; 2–16 [61] Nicolaisen JD, Richter Jr CW, Sheble GB. Price signal analysis for competitive
June 2005. p. 528–33. electric generation companies. In: Proceedings of the conference on electric
[50] Amjady N. Day ahead price forecasting of electricity markets by a new fuzzy utility deregulation and restructuring and power technologies 2000. London,
neural network. IEEE Trans Power Syst 2006;21(2):887–96. UK; April 2000. p. 4–7.
[51] Gareta R, Romeo LM, Gil A. Forecasting of electricity prices with neural [62] Li Y, Flynn PC. Deregulated power prices: changes over time. IEEE Trans Power
networks. Energy Convers Manage 2006;47:1770–8. Syst 2005;20(2):565–72.
[52] Pao Hsiao-Tien. A neural network approach to m-daily-ahead electricity price [63] Website of system operator of Spanish electricity market. <www.omel.com>.
prediction, ISNN 2006. Lecture notes in computer science LNCS 3972. Berlin, [64] Website of PJM electricity market. <www.pjm.com>.
Heidelberg: Springer-Verlag. p. 1284–9. [65] Website of New England electricity market. <www.iso-ne.com>.
[53] Lora AT, Santos Jesus R, Santos Jose R, Exposito AG, Ramos JLM. A comparison [66] Website of independent electricity system operator of Ontario, Canada.
of two techniques for next-day electricity price forecasting, IDEAL 2002. <www.ieso.com>.
Lecture notes in computer science LNCS 2412. Berlin, Heidelberg: Springer- [67] Website of Australia’s national electricity market. <www.nemmco.com>.
Verlag. p. 384–90. [68] Website of California independent system operator. <www.caiso.com>.