IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO.

2, MAY 2006 887

Day-Ahead Price Forecasting of Electricity Markets

by a New Fuzzy Neural Network
Nima Amjady, Member, IEEE

Abstract—In this paper, an efficient method based on a new

fuzzy neural network is proposed for short-term price fore-
casting of electricity markets. This fuzzy neural network has
inter-layer and feed-forward architecture with a new hyper-
cubic training mechanism. The proposed method predicts hourly
market-clearing prices for the day-ahead electricity markets. By
combination of fuzzy logic and an efficient learning algorithm, an
appropriate model for the nonstationary behavior and outliers of
the price series is presented. The proposed method is examined on
the Spanish electricity market. It is shown that the method can
provide more accurate results than the other price forecasting
techniques, such as ARIMA time series, wavelet-ARIMA, MLP,
and RBF neural networks.
Index Terms—Fuzzy neural network, price forecast.

I. INTRODUCTION

I N MANY countries around the world, the power industry

is changing from the centralized and vertically integrated
structure to an open market environment with the aim of pro-
Fig. 1. The 1999 weekday noon PJM prices.

viding a competitive framework. This process is known as re- and pay-as-bid pricing. Under the uniform pricing structure, the
structuring. The main objective of an electricity market is to de- marginal bid block sets the market-clearing price (MCP). How-
crease the cost of electricity through competition [1]. Instead ever, in the pay-as-bid (discriminatory) pricing structure, every
of considering the electrical energy as a public service, it is winning block gets its bid price as its income. The pricing mech-
considered as a commodity in the market environment. How- anism can affect the competition, efficiency, consumer surplus,
ever, the electricity market has its own complexities due to dif- and total revenue of the players in the electricity markets. A
ferences from most other commodities. The electrical energy whole discussion about this matter can be found in [3]. This
cannot be considerably stored, and the power system stability re- paper is focused on the MCP prediction in the uniform pricing
quires constant balance between generation and load. On short structure, which is the most commonly accepted structure of the
time scales, most users of electricity are unaware of or indif- electricity markets around the world. The MCP prediction can
ferent to its price. These two facts enforce the extreme price be performed for 1–168 h (one week) ahead; however, its usual
volatility or even price spikes of the electricity market, for in- forecast horizon is one day ahead [3], [4]. It is noted that after
stance, the price spikes of the Pennsylvania-New Jersey-Mary- the day-ahead market, where the major part of the total energy
land (PJM) market in 1999, shown in Fig. 1 [2]. is traded, subsequent short-term market mechanisms (such as
Generally, there are two different types of trading in the elec- intraday markets, ancillary reserves, and real-time markets) can
tricity markets: the pool and bilateral contracts. In the pool, the be executed in order to provide the final balance between power
producers and consumers submit their own bids, including a set generation and demand. However, this paper is focused on the
of quantities with their prices. Then the set of clearing prices MCP prediction in the day-ahead markets.
for the next day is announced by the market operator. Bilateral The remaining sections of this paper are organized as follows.
contract is an agreement between a seller and a buyer on a cer- In Section II, an overview of the problem, i.e., the MCP predic-
tain amount of electrical energy at a certain fixed price. Bilateral tion and its previous solution methods, are presented. Then the
contract is a tool for the companies to hedge against the risk of rationale behind the proposed technique is explained. In Sec-
pool price volatility. Pool type and bilateral contract transactions tion III, details of the technique are described. Obtained numer-
can be simultaneously seen in any modern electricity market. ical results are presented and discussed in Section IV. A brief
Another essential characteristic of the electricity markets is review of this paper and the future research are in Section V.
the pricing mechanism, which can be in the form of uniform
II. PROBLEM DESCRIPTION
Manuscript received July 12, 2005; revised November 6, 2005. Paper no. The price forecast is a key piece of information in the elec-
TPWRS-00426-2005. tricity markets. In the pool-type power markets, the market par-
The author is with the Department of Electrical Engineering, Semnan Uni-
versity, Semnan 35195-363, Iran (e-mail: amjady@tavanir.org.ir). ticipants must express the price of their offered quantities. The
Digital Object Identifier 10.1109/TPWRS.2006.873409 bids are accepted in the ascending order of price until the total
0885-8950/$20.00 © 2006 IEEE
888 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

Fig. 2. Price series for Week 43 (November) of the Spanish market. Fig. 3. Week 5 (February) of the Californian market.

demand is met. So, forecast of the pool price is important for rectly forecast the MCP by analyzing the effective parameters
a company to adjust its own price/production schedule. In the on the market price, such as production costs and strategic be-
bilateral contracts, the agreed price of buyer and seller is also havior of the market participants. An important group of these
based on MCP predictions. Thus, a good knowledge of future methods is based on the game theory and auction theory [5].
pool prices helps to valuate more accurately bilateral contracts. Another group consists of fundamental or structural models.
However, MCP prediction is difficult. MCP is a nonstationary These models based on traditional cost models have been de-
series with nonconstant mean and variance. This series is also veloped for centralized systems and adapted to liberalized mar-
encountered with significant outliers, i.e., unusual prices (espe- kets [5]. The second category of the MCP prediction methods
cially in periods of high demand), due to unexpected or uncon- tries to forecast MCP without analyzing the underlying phys-
trolled events in the electricity markets [4]. In [5], it has been ical processes in detail. These methods, based on the black box
shown that a significant nonstationary component can be de- models, analyze price evolution by means of statistical data. The
tected in electricity price time series due to its multiple sea- methods of the second category, usually based on the time series
sonalities (related to daily, weekly, and monthly periodicities) or neural network (NN), are more common for MCP prediction
and to its switching nature related to discrete changes in par- than the methods of the first category, due to their more flexi-
ticipants’ strategies (e.g., agents decide to switch from a con- bility and ease of implementation.
servative attitude to a more aggressive or risky one). In [6], Time-series models have been already applied to forecast
other complexities of the hourly price series, such as high-fre- commodity prices, such as oil and natural gas [8]. Early ap-
quency changes and calendar effect (e.g., weekends and holi- plications of the time-series models in the power system were
days), have been explained. Two examples of the price series related to STLF. With the restructuring process, autoregres-
are shown in Figs. 2 (Spanish market, 2002) and 3 (Californian sive (AR) models were used to predict weekly prices, like in
market, 2000). The horizontal axis in both figures is in terms of the Norwegian system [9]. In [8], a more efficient time-se-
hours. Price values, i.e., vertical axis, in Fig. 2 are in terms of ries model, i.e., autoregressive integrated moving average
c Euro/KWh (cE/KWh) and in Fig. 3 are in terms of $/MWh. (ARIMA), has been used for price forecasting of the Spanish
These figures show the above-mentioned characteristics, which and Californian electricity markets. In [6], two other time-se-
are sufficient to produce a highly stochastic time series, creating ries models, including dynamic regression and linear transfer
many complexities in forecasting its future values. function models, have been proposed for short-term price
A lot of methods have been developed for MCP prediction forecasting. The main drawback of these time-series models is
of the electricity markets, especially in the last decade. Some that they are usually based on the hypothesis of stationarity;
of these methods are basically the short-term load forecasting however, the price series violate this assumption. To solve
(STLF) methods. However, MCPs are usually more volatile than this problem, another kind of time-series model known as au-
hourly loads, and so, the MCP prediction is more complex than toregressive conditional heterocedastic (ARCH)-type models,
the STLF. This is due to the fact that MCP is dependent on the assuming time-dependent variance, have been developed [5],
hourly loads and some other stochastic signals, such as equip- [10]. Although these models have been successfully applied
ment outages and fuel prices. For instance, in [7], projected to other commodity markets, however, their application to
surplus, Henry hub gas price, and Nymex oil price have been electricity prices encounters difficulty [5]. Some other MCP
used in addition to hourly loads and temperatures to forecast prediction methods such as econometrics, which are based on
the MCPs. Thus, uncertainty of hourly loads and the above sto- the time-series models, are not discussed more, since they also
chastic signals are combined, resulting in a higher level of un- encounter the above-mentioned problems.
certainty in the MCP. Although some promising results have been obtained from
In general, the MCP prediction methods can be divided into the time-series techniques for MCP prediction, however, most
two main categories. The methods of the first category try to di- of these methods are linear techniques and so cannot appropri-
AMJADY: DAY-AHEAD PRICE FORECASTING OF ELECTRICITY MARKETS 889

ately track the hard nonlinear behavior of the MCP signal. For (IOHMM) architecture, has been presented. Although decom-
instance, the hourly load signal can have sudden changes due to position of the MCP signal can improve its forecasting accuracy,
rapid variations of the temperature. These nonlinear behaviors however, these methods suffer from some problems. Usually de-
are more serious in the case of price signal, as seen from Figs. 2 composition of the MCP signal, including high-frequency com-
and 3. Thus, the time-series techniques may encounter large un- ponents, to a finite number of constitutive series or blocks lose
expected prediction errors. Forecast errors more than 30% in the some information, and some interdependencies may be ignored.
case of MCP prediction have been reported [4], [8]. In [6], this Besides, identification of the MCP states, assignment of them
problem is stated in another way. The time-series techniques are to the constitutive blocks and composition of the forecasts are
successful in the areas where the frequency of the data is low, difficult tasks based largely on the probabilistic and uncertain
such as weekly patterns. Rapid variations and high-frequency information.
changes of the target signal can be problematic for these tech- To solve the above problem (inadequacy of a single NN to
niques. Thus, there is a need for more efficient forecasting tools construct a global model for the MCP signal), instead of using
capable of tracking the hard nonlinear behaviors of the MCP multiple forecasting tools, a new fuzzy NN (FNN) with higher
signal. learning capability than NNs is proposed. An appropriate tool
In recent years, another kind of forecasting method based on to soften the high-frequency changes of the MCP is fuzzy
the artificial intelligence techniques, and especially NNs, has logic. Besides, different parts of the MCP series have different
been proposed by the researchers for price forecasting of the input–output mapping functions (different functional relation-
electricity markets. Fuzzy regression models that relate prices ships), which in [5] is referred to as the piece-wise time series.
and demands have been applied to the Californian market by From state–space viewpoint, MCP can be described by a non-
Nakashima et al. [6]. Ramsay et al. [11] have proposed a hybrid linear and time-dependent model. Here, to solve the problem
approach based on NNs and fuzzy logic, with examples from of time dependency of the MCP model, the idea of division
the England-Wales pool market and daily mean errors around of the input vector space (in place of MCP) into subspaces
10%. Szkuta et al. [12] have proposed a three-layered NN with is proposed, such that in each subspace, a time-independent
back-propagation (BP), showing results from the Victorian elec- model for the MCP can be considered. The proposed FNN
tricity market, with daily mean errors around 15%. In [13], a partitions the input space based on the concept of hypercube.
This partitioning process is a part of learning phase of the
similar NN has been applied for the Californian market by Gao.
FNN. So, no external composer/decomposer of the MCP (with
Nicolaisen et al. have presented Fourier and Hartley transforms
the above-mentioned difficulties) is required, and a smooth
[14] as filters to the price data inputs of a NN. Rodriguez and
partition in terms of input space can be obtained. Finally, the
Anders compared an NN owning two learning algorithms with
nonlinear behavior of the MCP signal is modeled by the NN
a neuro-fuzzy system for forecasting of hourly MCP in the On- part of the FNN.
tario energy market, where the later showed better prediction
accuracy than the former [15].
III. PROPOSED FNN
Although NNs are universal approximators and therefore can
theoretically emulate any function, in practice, it is often very To introduce the proposed FNN, its NN part and fuzzification
hard for an NN to construct a mapping function for the non- method are described. The NN part of the proposed FNN has an
stationary and highly nonlinear MCP signal. In [7], it has been inter-layer and feed-forward architecture [18], which consists
shown that in view of reasons such as insufficient input–output of three layers: input layer, hidden layer, and output layer.
data points or too many tunable parameters, in reality, a single Input layer nodes do not process their inputs. These nodes
NN often misrepresents part of the nonlinear input–output rela- just propagate input features of the training or test sample
tionship of the MCP. In [16], it has been shown that traditional to the hidden layer. Thus, each node of the
learning algorithms such as BP and Newton’s methods may not hidden layer obtains a vector of weighted inputs, as follows:
be effective to train NNs for predicting nonstationary processes,
such as MCP series. Finally, in [5], it is concluded that it is often (1)
very hard for an NN to extract a global model for the whole data
of the MCP series. where is input vector of hidden node , and indicates
To solve this problem, recently, some researchers try to de- weight between input node and hidden node . The hidden
compose the complex signal of MCP, then forecast its compo- nodes of the proposed FNN perform fuzzification process as
nents separately, and finally combine these forecasts to obtain follows:
the MCP prediction. In [4], the authors decomposed the MCP
series through the wavelet transform and assigned an ARIMA (2)
model to each constitutive series. In [16], the authors used a kind
of decoupled extended Kalman filter (DEKF) to train an MLP
NN for the MCP prediction. In their DEKF, the input weights (3)
to a neuron are grouped, and the weight covariance matrix is
assumed to be block diagonal. This method ignores the interde-
pendencies of weights from different neurons. In [7] and [17],
the idea of committee machine, a parallel structure of different
NNs, has been presented. In [5], a similar approach, including (4)
the MLP NNs in the form of input–output hidden Markov model (5)
890 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

In the above relations, is output of hidden node , indi-

cates element of vector , and is number of hidden nodes.
indicates hypervolume corresponding to hidden node .
Consider the cube in the conventional space (three-dimen-
sional). The volume of this cube is

(6)

where and indicate upper and lower bands of

cube on the axes . If the dimension of input space increases,
i.e., , then the conventional cube becomes a subspace
with rectangular boundaries of the -dimensional input space.
This subspace is called hypercube. From a mathematical view-
point, we can extend the concept of volume, calculating by (3), Fig. 4. Membership function O versus input I for one-dimensional input
for this -dimensional input subspace. Here, this extension of space (I becomes scalar).
volume concept is called hypervolume. One hypercube is as-
signed to each hidden node of the FNN (totally hyper-
cubes). Thus, the FNN can partition the input vector space into Thus, the learning algorithm of the FNN must perform two
subspaces in the form of hypercube. Each input with input fea- tasks, each one involved with a separate set of adjustable pa-
tures indicates a point in the -dimensional rameters.
input vector space. in (2) represents the membership func- 1) Classification of the input vector space into hypercubes,
tion of the FNN, in the form of proportion of two hypervol- based on the observations, i.e., training samples. This part
umes, measuring the closeness of the current input to the hy- of the learning process is involved with the adjustment of
percube . When the current input is in the the tunable parameters of ,
hypercube , based on (3) and (4), the hypervolume be- . As seen from
comes equal to . However, when the current input (1), acts as a scaling factor for input feature when
goes outside the rectangle, becomes less than one, and it re- presented to class . The learning algorithm by means of
duces as the current input becomes farther from the rectangle. these weights can locate observations in the input space,
So, each hypercube implements a fuzzy class in the input vector when observed from class . and indicate
space, and the hidden layer nodes of the FNN perform a fuzzi- upper and lower bounds of the hypercube for input fea-
fied classification of the input space. Assume that we can imple- ture .
ment each functional relationship of the MCP (mapping func- 2) Implementation of each functional relationship of MCP
tion from the input features to MCP) in one in one class. This part of the learning process is involved
class, i.e., one hypercube of the input space. Then, the FNN by with the adjustment of tunable parameters ,
receiving the current input can rank its classes according to the . In this part, output of hidden nodes or classes
distance between the input and each class. Closer classes have of the FNN are tuned based on the input features to match
higher ranks and so more contribution in the final output of the the desired output.
FNN. In other words, responses of different classes are com- The applied membership functions are continuous and soft
bined based on the degrees of closeness to construct the final changing (see Fig. 4). From a mathematical viewpoint, these
output or out, which presents the MCP prediction. This combi- functions are well conditioned, except at some boundaries,
nation (defuzzification) is performed in the single node of the which are unavoidable. So, the partial derivative of these mem-
third layer or output layer of the FNN bership functions can be easily computed. This characteristic
makes the motivation to use the generalized delta rule (GDR)
method [18] in the learning algorithm to adjust the tunable
(7) parameters of the FNN

where indicates weight between hidden node and output (8)

node. It is noted that in the combination relationship of (7),
measures the degree of closeness of the current input to class ,
and indicates the response of the class to the input. It (9)
is noted that if we use a crisp classification, then only one class
is selected for each input. However, by fuzzified classification, where and are two successive iterations; indicates
all classes can contribute in the final output of the FNN by the total training error of the FNN; and and represent the
degree of closeness to the input. So, the fuzzified classification learning rate and momentum rate of the learning algorithm cor-
permits the FNN to employ all its acquired knowledge for re- responding to parameter , respectively. Mathematical details
plying to each input. of computation of (8) and (9) can be found in the Appendix.
AMJADY: DAY-AHEAD PRICE FORECASTING OF ELECTRICITY MARKETS 891

It is noted that the proposed method does not decompose market shows even higher dispersion, which causes more uncer-
the MCP signal. Thus, loss of accuracy due to high-frequency tainty in periods of high demand, producing less accurate fore-
components of MCP does not occur. Besides, in the proposed casts. In [5], it has been shown that there are different functional
method, there is no need for an external composer/decomposer. relationships between the input variables and marginal price in
Instead of the MCP decomposition, the method decomposes the this market. Thus, it is seen that the Spanish electricity market
input space, where each functional relationship of the MCP can is a real-world case study with sufficient complexity.
be implemented in one subspace (hypercube). Identification of Training samples of the FNN have one output feature dedi-
these functional relationships (task 1) and their implementation cated to forecast MCP for the next hour. Different input features
(task 2) are performed automatically and internally in the FNN for MCP prediction have been proposed in the previous works.
so that the forecast error of the MCP is minimized. The pro- These input features contain different lags of the price as the
posed method is relatively similar to a finite element method, auto-regression part, and exogenous variables including hourly
since they both decompose the domain of input variables. How- system demand, hourly generation, surplus generation, fuel
ever, the finite element methods are used to solve the differential costs, generation companies’ shares, aggregated supply func-
equations, while the proposed method is used for the MCP pre- tions, day indicator, etc. as the cross-regression part [4]–[8],
diction. [11], [12], [17]. Although use of exogenous variables can
The only remaining part of the proposed method is selection increase the accuracy of the MCP prediction, in this paper, only
of the number of hidden nodes of the FNN, i.e., the number the auto-regression part is considered. In [4] and [5], it has been
of hypercubes of the input space. In the MCP decomposition shown that for the Spanish electricity market, some of these
methods, the number of constitutive blocks is usually selected exogenous variables, such as fuel costs, generation companies’
based on trail and error and/or experience. For instance, in [5], shares, and aggregated supply functions have no significant
some architectures were examined, and the best one among effect on the accuracy of the MCP prediction. Moreover, this
them (four-state model) was selected. In [4], four constitutive paper focuses on the advantages of the proposed FNN as an
series (one approximation and three details) were considered. MCP prediction method. Thus, for the sake of simplicity and
In the committee machine presented in [7], two NNs (one MLP clear comparison, no exogenous variables are considered for
and one RBF) were considered. Instead, the following rule is the FNN. Selection of the best input features for the MCP
used here. In the proposed hypercubic modeling, each input prediction, dependent on the structure of the electricity market,
feature presents one dimension of the input vector space. There is a challenging task, which can be a matter for future research.
is a known rule in linear algebra that the maximum number of Different sets of lagged prices have been proposed as input
linearly independent vectors in the -dimensional vector space features for the MCP prediction in the Spanish electricity
is . Each hypercube in the -dimensional input space can be market [5], [6], [8]. Bearing in mind the daily and weekly pe-
constructed by a linear combination of the vectors of this space, riodicity and trend of the MCP signal, the set of
and so, it can be finally represented by a vector in the space.
Thus, for the maximum discrimination and maximum covering
of the input space, the best choice for the number of hypercubes is considered in this paper as the input
is , i.e., . If the number of hypercubes is less than features. The selected input features are lagged prices recom-
the dimension of the input space, then some vectors in the space mended in [6]. , i.e., MCP of hour is the single output
cannot be constructed by combination of hypercubes, which feature of the training samples. When MCP of a hour is fore-
means that some parts of the input space cannot be covered casted, it is used as for the MCP prediction of the next
by the hypercubes. If is selected more than , then the hour, and this cycle is repeated until the MCP of the whole
hypercubes become linearly dependent, which increases the forecast horizon (here, the next 24 h) is predicted. As described
risk of over-fitting and computation burden. in the previous section, the number of hidden nodes of the
FNN is selected equal to the number of input nodes, i.e.,
. The learning rates are selected 0.9, and
IV. NUMERICAL RESULTS the momentum rates are selected 0.7, recommended values
for the GDR training mechanism [18], [19]. Larger values of
The proposed method is examined on the day-ahead elec- the learning and momentum rates may accelerate the learning
tricity market of mainland Spain, commonly used as the test process, but at the same time, convergency of the GDR may be
case in many price forecasting papers [4]–[6], [8], [11]. Some lost. Especially larger values of the momentum rate may result
price forecasting methods, such as the ARIMA time series, are in the oscillation of the training error. Besides, based on (8), the
routinely used by the power industry in Spain. The electricity larger rates mean larger variations of the adjustable parameters
market of mainland Spain is a duopoly with a dominant player. in each step, and so, the resolution of the GDR for finding some
This results in price changes related to the strategic behavior optimum points may decrease. More details about this matter
of the dominant player, which are hard to predict [4]. In [6], the can be found in [18] and [19].
Spanish and Californian markets have been compared, and it has Now training period must be selected. If functional rela-
been concluded that a higher proportion of outliers and a lesser tionships (not numerical values) of a signal slowly varies with
degree of competition results in more volatility of the Spanish time, then a long history of the signal can be considered as
market, which in turn makes this market less predictable. More- the training period, which results in a large number of training
over, it has been shown that during peak hours, the Spanish samples. Moreover, obtained results from one training phase,
892 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

such as weights of an NN, can be used for many forecasts. For TABLE I
instance, functional relationships of hourly load, especially in WEEKLY MEAN ERRORS FOR FOUR WEEKS OF 2002
the developed countries, slowly varies with time. Thus, many
training samples from the past history of this signal, such
as two years ago, can be used for the NN or FNN training.
Besides, the training phase can be repeated once every month
or even once every season. However, as described previously,
functional relationships of the MCP varies much more rapidly
than the hourly load. So, the training period of MCP is more
limited than the hourly load. On the other hand, if the training TABLE II
period is selected to be very short, then the NN or FNN can VARIANCE OF PREDICTION ERROR FOR FOUR WEEKS OF 2002
not derive all functional relationships of the MCP due to the
small number of training samples. In [4], the last 48 days has
been proposed as the training period of the ARIMA time series
for the MCP prediction in the Spanish electricity market. This
training period is used here for the FNN training. Thus, the
FNN is trained with training samples in
the form shown above. Then it forecasts the MCP of the next
24 h (one day ahead), which are unseen for the FNN. The been quoted from [4], and so, they are in terms of . The
training phase of the FNN is executed for each day separately. weekly error of the FNN in terms of and WME are pre-
It is noted that the selected input variables and training period sented in the columns 4 and 5 of Table I, respectively (
(including 1152 training samples) have been considered in all values are less than WME values). However, in all remaining
examinations of this paper. examinations of this paper, only DME and WME values are
Real data of the Spanish electricity market in year 2002 [20] presented as daily and weekly errors, respectively. As an index
are considered in the first examination. Obtained results for four of uncertainty, weekly error variance of the examined
weeks corresponding to the four seasons of year 2002 are pre- methods for the considered weeks is presented in Table II. The
sented in Table I (the last row indicates average error of these definition of [4] for , consistent with (11), has been con-
four weeks). In this manner, representative results for the whole sidered for the results of Table II, since variance values of the
year are provided. According to [4], the fourth week of Feb- ARIMA and wavelet-ARIMA methods have been quoted from
ruary, May, August, and November (months 2, 5, 8, and 11) are this reference
selected for winter, spring, summer, and fall seasons, respec-
tively. In Table I, the test weeks are indicated in column 1, re- (12)
spectively. The weekly mean error (WME) of a forecast method
can be defined as follows:s
In the wavelet-ARIMA method, the MCP series is decom-
posed through wavelet transform into four constitutive series,
WME (10)
including one approximation and three details. Then one
ARIMA model is assigned to each constitutive series. In the
where and are actual and forecasted price of hour ARIMA method, one time series is allocated to the whole MCP
, respectively. Similarly, by consideration of daily period or 24 series. From Table I, it can be seen that the accuracy of the
h instead of weekly period or 168 h in (10), daily mean error wavelet-ARIMA is better than the ARIMA, and accuracy of the
(DME) can be obtained. In [4], weekly and daily mean errors proposed FNN is better than both. The nonstationary behavior
called and have been defined in another way as fol- of the constitutive series of the wavelet-ARIMA is less than the
lows: MCP series, and so, the accuracy of this method is more than
the ARIMA method. The proposed FNN presents a better solu-
tion than the wavelet-ARIMA, since the FNN does not involve
(11)
decomposition/composition of the complex MCP signal. In-
stead, it smoothly classifies the input vector space by means of
where or weekly average of actual prices has been fuzzified classification. Thus, different functional relationships
used for denominator of , instead of , to avoid ad- of the MCP can be better recognized, without requiring any
verse effect of prices close to zero. Similarly, is defined approximation or external composer/decomposer. The Spanish
[4]. On the other hand, DME and WME are more sensible since electricity market has more unstable behavior in the summer
they reflect daily and weekly mean absolute percentage error and fall seasons than the winter and spring seasons due to the
(MAPE), respectively. Besides, WME for a week has an av- strategic behavior of the dominant player in the market [4].
erage of seven DME values corresponding to seven days of that Thus, all represented methods encounter higher forecast errors
week. However, is not the weekly average of the cor- for the summer and fall weeks than the winter and spring weeks.
responding values. Results of the ARIMA and wavelet- As seen from Table I, the accuracy of the proposed method for
ARIMA methods, indicated in columns 2 and 3 of Table I, have all examined weeks is better than the wavelet-ARIMA method.
AMJADY: DAY-AHEAD PRICE FORECASTING OF ELECTRICITY MARKETS 893

Fig. 5. Hourly prices (dark), price forecasts (grey), and absolute value of
forecast errors (grey, at the bottom) for the fall’s test week.

Besides, in the less stable periods, the difference between fore-

cast error of the FNN and wavelet-ARIMA method increases.
This characteristic of the proposed FNN is useful for the utility
companies to encounter with these periods in an electricity
market. As seen from Table II, the weekly error variances of
the FNN are also smaller than those of the other techniques,
especially in the less stable periods. Obtained results from the
FNN for the fall’s test week (with the highest forecast error) are
shown in Fig. 5. It can be seen that even for the worst test case,
the FNN follows the trend of the price curve, and its accuracy
is reasonable.
In Fig. 6, the proposed FNN is compared with the MLP and
RBF NNs for the same test weeks of Table I. The RBF networks
are effective in exploiting local data characteristics, and MLP
networks are good at capturing global data trends [7]. These two
NNs, representative of different forecast methods, have been
considered in this examination. In Fig. 6, bar charts from 1 to
7 are DME and 8 indicates WME. The FNN, MLP, and RBF
NNs had the same training period and training samples as de-
scribed for the results of Table I. Parameters of the MLP and
RBF NNs were selected, as recommended in [7]. As seen from
the obtained results, each of the MLP and RBF outperforms the
other in some test periods. However, the WME of the FNN is
less than both NNs in all test weeks. The average of four WME
values for the MLP and RBF is 10.97% and 10.6%, respectively,
compared to 7.83% for the FNN. While the MLP and RBF NNs
try to construct a global mapping function for the whole MCP
series, the FNN smoothly partitions the input vector space. The
functional relationship of the MCP in each subspace can be con-
structed much easier than a global model. The FNN, MLP, and
RBF had the same training time, about 5 min, on a Pentium P4
2.8-GHz personal computer with 512-MB RAM memory. In the
written codes for the FNN, MLP, and RBF, the training time is
measured, and at the end of 5 min, the used learning algorithm is
terminated. Obtained values for the adjustable parameters (such Fig. 6. Daily and weekly mean errors of the MLP (light grey), RBF (dark
as weights) in the last executed iteration are considered for the grey), and FNN (dark). a) Winter test week. b) Spring test week. c) Summer test
week. d) Fall test week.
MCP prediction of the 24 h ahead. This training time has been
also considered for the ARIMA and wavelet-ARIMA methods
on a slightly different hardware [4]. This short training time per- effective in extraction of the data trend and so the accuracy of the
mits to execute the training phase separately for each day, i.e., forecast method. Training error of 28 learning phases corre-
the forecast method can be trained with the earliest data. This is sponding to four considered weeks is in the range of
894 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

TABLE III TABLE IV

MINIMUM, AVERAGE, AND MAXIMUM OF DME IN FOUR MONTHS OF 2002 DME OF AUGUST WEEK (%)

TABLE V
DME OF NOVEMBER WEEK (%)
to . is defined as the average error of all training
samples

(13)
TABLE VI
STATISTICAL MEASURES OF AUGUST AND NOVEMBER WEEKS
where is the number of training samples (here, 1152).
To give a better insight about the overall accuracy of the pro-
posed method in a long run, minimum, average, and maximum
DME of the FNN in four months of 2002 (months 2, 5, 8, and
11) are presented in Table III. The minimum DME of February,
i.e., 3.08%, occurred in the selected week or fourth week of
this month (rounded to 3.1% in the fourth day of winter test
week of Fig. 6). The average of DME values of this month, i.e.,
4.95% (indicating monthly mean error of February), is slightly hourly system demands as the input variables. Besides, exact de-
larger than the WME of the selected week, i.e., 4.8%. In May, mand values have been used for the DR and TF methods, while
the monthly average of DME values is approximately the same in practice, at least the demand value of future hours is forecast.
as the WME of the spring test week (5.62% versus 5.6%). The On the other hand, the proposed FNN uses only the auto cor-
maximum DME of August, i.e., 15.41%, occurred in the fourth relation information like the previous examination. In Table I
day of the summer test week, rounded to 15.4% in Fig. 6. The and Fig. 6, it has been shown that the FNN with the same input
average DME of this month is less than WME of the summer test data can provide more accurate forecasts for MCP than the other
week (9.90% versus 10.2%). In November, the average DME is mentioned methods. From Tables IV–VI, it is seen that the FNN
also less than the WME of the fall test week but with less differ- with less input data can provide MCP prediction accuracy com-
ence (10.54% versus 10.7%). As seen from Table III, prediction parable to the DR and TF methods. These comparisons reveal
accuracy of the FNN is also acceptable for the four considered the forecast capability of the FNN. Besides, the obtained values
months. of SD, as an index of the model uncertainty [6], indicate less un-
In Tables IV–VI, the proposed FNN is compared with dy- certainty of the FNN model than the DR and TF methods. The
namic regression (DR) and ransfer function (TF) methods for setup time of the DR and TF methods, measured on a slightly
day-ahead price forecasting of the Spanish electricity market. different hardware [6], is about the training time of the FNN.
The DR and TF methods are based on the time-series analysis,
and their mathematical details can be found in [6]. Results of V. CONCLUSION
these methods, including DME, WME, and standard deviation
of error (SD), shown in Tables IV–VI have been quoted from High-quality MCP prediction would help utilities submit ef-
this reference. SD values in Table VI have been computed ac- fective bids with low risks and make good bilateral transaction
cording to the following relation: decisions. However, MCP prediction is a difficult task due to
characteristics of the MCP signal, such as high-frequency com-
ponents, nonstationary behavior, multiple seasonality, calendar
effect, high volatility, high percentage of outliers, hard nonlinear
(14) behavior, etc. At the same time, utility companies usually have
limited information. Thus, there is an essential need for an effi-
cient and robust forecasting method for the companies that trade
Two weeks have been considered for this comparison. The first in the electricity markets. This subject matter motivates much
one, considered in Table IV, is the fourth week of August 2000 research, especially in the last decade. Recently, the researchers
(from August 21–27) with low demand, and the second one, noticed that to efficiently solve this problem, the MCP signal
considered in Table V, is the third week of November 2000 (from should be decomposed. However, high-frequency changes of
November 13–19) with high demand [6]. As seen from DME the MCP plus its hidden and complex states make difficult the
and WME values in Tables IV–VI, the FNN has comparable decomposition of this signal.
and even slightly better prediction accuracy than the DR and TF In this paper, another decomposition way based on the
methods. As described in [6], both the DR and TF methods used input vector space has been proposed for this problem. The
a set of lagged prices (similar to that of the FNN) and a set of proposed method has been implemented by means of a new
AMJADY: DAY-AHEAD PRICE FORECASTING OF ELECTRICITY MARKETS 895

FNN. The FNN internally and smoothly partitions the input

space in the form of a fuzzified classification. Obtained results (A8)
from the Spanish electricity market have been presented and
compared with those of DR, TF, and ARIMA time series,
wavelet-ARIMA, MLP, and RBF networks. It has been shown (A9)
that the proposed method can provide reasonable forecasts
of MCP by means of limited information consisting of a set
(A10)
of lagged prices. The research work is under way in order to
select more appropriate input features, including both lagged
prices and exogenous variables, based on the electricity market
structure. The spectrum analysis and sensitivity analysis may By means of (2), it can be written that
be good candidates for this purpose. Besides, MCP prediction
in the real-time markets can be a matter for future research.
(A11)
APPENDIX
Here, we compute the partial derivatives of the error func- According to (3), we have
tion with respect to the adjustable parameters of the FNN,
i.e., . The GDR algorithm uses
the energy criterion for the total training error (A12)

(A1) (A13)

where , , and indicate actual output, desired output, According to (4), it can be written that
and training error of the training sample , respectively; is
the number of training samples. By using (A1), it can be written
that If

(A14)
(A2)
If

By using the chain rule, we have

(A15)
(A3) If

(A4)
(A16)

(A5)
By substituting from (A14) to (A16) in (A13),
can be obtained. Then by substituting
Similarly, for weights between input layer and hidden layer, it in (A10), the partial derivative of can be obtained.
can be written that To compute the partial derivative of , by a process
similar to (A6)–(A10), it can be written that

(A17)

(A6)
(A18)

(A7) (A19)
896 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 21, NO. 2, MAY 2006

According to (3) and (4), we have REFERENCES

[1] J. Contreras, A. J. Conejo, S. de la Torre, and M. G. Munoz, “Power
engineering lab: Electricity market simulator,” IEEE Trans. Power Syst.,
vol. 17, no. 2, pp. 223–228, May 2002.
(A20) [2] M. Davison, C. L. Anderson, B. Marcus, and K. Anderson, “Develop-
ment of a hybrid model for electrical power spot prices,” IEEE Trans.
If Power Syst., vol. 17, no. 2, pp. 257–264, May 2002.
[3] Y. S. Son, R. Baldick, K. Lee, and S. Siddiqi, “Short-term electricity
market auction game analysis: Uniform and pay-as-bid pricing,” IEEE
Trans. Power Syst., vol. 19, no. 4, pp. 1990–1998, Nov. 2004.
[4] A. J. Conejo, M. A. Plazas, R. Espinola, and A. B. Molina, “Day-ahead
electricity price forecasting using the wavelet transform and ARIMA
(A21) models,” IEEE Trans. Power Syst., vol. 20, no. 2, pp. 1035–1042, May
2005.
If [5] A. M. Gonzalez, A. M. San Roque, and J. G. Gonzalez, “Modeling and
forecasting electricity prices with input/output hidden Markov models,”
IEEE Trans. Power Syst., vol. 20, no. 1, pp. 13–24, Feb. 2005.
[6] F. J. Nogales, J. Contreras, A. J. Conejo, and R. Espinola, “Forecasting
next-day electricity prices by time series models,” IEEE Trans. Power
(A22) Syst., vol. 17, no. 2, pp. 342–348, May 2002.
[7] J. J. Guo and P. B. Luh, “Improving market clearing price prediction
If by using a committee machine of neural networks,” IEEE Trans. Power
Syst., vol. 19, no. 4, pp. 1867–1876, Nov. 2004.
(A23) [8] J. Contreras, R. Espinola, F. J. Nogales, and A. J. Conejo, “ARIMA
models to predict next-day electricity prices,” IEEE Trans. Power Syst.,
vol. 18, no. 3, pp. 1014–1020, Aug. 2003.
[9] O. B. Fosso, A. Gjelsvik, A. Haugstad, M. Birger, and I. Wangensteen,
By substituting from (A20) and “Generation scheduling in a deregulated system. The Norwegian case,”
from (A21) to (A23) in (A19), the partial derivative of IEEE Trans. Power Syst., vol. 14, no. 1, pp. 75–81, Feb. 1999.
[10] R. C. Garcia, J. Contreras, M. V. Akkeren, and J. B. C. Garcia, “A
can be obtained. Then by substituting GARCH forecasting model to predict day, ahead electricity prices,” in
in (A17), the partial derivative of can be obtained. Proc. Work. Applied Infrastructure, Berlin, Germany, October 11, 2003.
Similarly, to compute the partial derivative of , [11] A. J. Wang and B. Ramsay, “A neural network based estimator for elec-
tricity spot-pricing with particular reference to weekend and public hol-
it can be written that idays,” Neurocomputing, vol. 23, pp. 47–57, 1998.
[12] B. R. Szkuta, L. A. Sanabria, and T. S. Dillon, “Electricity price short,
term forecasting using artificial neural networks,” IEEE Trans. Power
Syst., vol. 14, no. 3, pp. 851–857, Aug. 1999.
(A24) [13] F. Gao, X. Guan, X.-R. Cao, and A. Papalexopoulos, “Forecasting power
market clearing price and quantity using a neural network method,” in
Proc. IEEE Power Eng. Soc. Summer Meeting, vol. 4, Seattle, WA, Jul.
(A25) 2000, pp. 2183–2188.
[14] J. D. Nicolaisen, C. W. Richter Jr., and G. B. Sheblé, “Price signal anal-
ysis for competitive electric generation companies,” in Proc. Conf. Elec-
trical Utility Deregulation Restructuring Power Technologies, London,
According to (3) and (4), we have U.K., Apr. 4–7, 2000, pp. 66–71.
[15] C. P. Rodriguez and G. J. Anders, “Energy price forecasting in the On-
tario competitive power system market,” IEEE Trans. Power Syst., vol.
19, no. 1, pp. 366–374, Feb. 2004.
(A26) [16] L. Zhang and P. B. Luh, “Neural network-based market clearing price
prediction and confidence interval estimation with an improved extended
Kalman filter method,” IEEE Trans. Power Syst., vol. 20, no. 1, pp.
If 59–66, Feb. 2005.
[17] X. Yao and Y. Liu, “Making use of population information in evolu-
tionary artificial neural networks,” IEEE Trans. Syst., Man, Cybern. B,
Cybern., vol. 28, no. 3, pp. 417–425, Jun. 1998.
[18] N. Amjady, Introduction to Intelligent Systems. Semnan, Iran: Semnan
Univ. Press, 2002.
(A27) [19] Y. H. Pao, Adaptive Pattern Recognition and Neural Net-
If works. Reading, MA: Case Western Reserve Univ., 1989.
[20] OMEL. Market operator of the electricity market of mainland Spain.
[Online] Available: http://www.omel.es/frames/es/index.jsp.

(A28)
If Nima Amjady (M’97) received the B.Sc., M.Sc., and Ph.D. degrees in electrical
engineering from Sharif University of Technology, Tehran, Iran, in 1992, 1994,
and 1997, respectively.
(A.29) Presently, he is an Associate Professor with the Electrical Engineering De-
partment, Semnan University, Semnan, Iran. He is also a Consultant with the
National Dispatching Department of Iran. His research interests include secu-
rity assessment of power systems, reliability of power networks, load and price
From (A25)–(A29), can be obtained. Then from forecasting, and application of artificial intelligence methods to the problems of
(A24), is determined. power systems.