Electricity Price Forecasting Using ELM-Tree Approach

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395 -0056
Volume: 03 Issue: 01 | Jan-2016 www.irjet.net p-ISSN: 2395-0072
© 2016, IRJET | Impact Factor value: 4.45 | ISO 9001:2008 Certified Journal | Page 1209
Electricity Price Forecasting Using ELM-Tree Approach
Mr. Vrushab Gosalia 1, Mr. Pradip Andhare 2, Mr. Prajwal Patil 3, Mr. Arunoday Chorage 4,
Mrs. Renuka Gound 5
1234Student, Department of Information Technology, PCCOE, Maharashtra, India
5 Assistant Professor, Department of Information Technology, PCCOE, Maharashtra, India
---------------------------------------------------------------------***------------------------------------------------------------------
Abstract - In power markets there is no regularity in
forecasting of electricity price which is the most
essential tasks & basis for any decision making. In the
competitive electricity markets forecasting of the
electricity price is critical for consumers and producers
of electricity for planning their operations and also to
maintain the risk of electricity. Forecasting also plays a
very important role in economic optimization of
electricity usage. Artificial intelligence (AI) along with
the ELM-Tree Approach has been applied in price
forecasting that is, the day ahead usage of the
electricity and will also predict the monthly bill of the
user as per there electricity usage. For handling the
complex interdependencies between the electricity
prices, historical load usage and various other factors
artificial neural networks are the best method that can
be used. This ANN model approach is used to forecast
the behaviors of the electricity market based on the
various factors such as historical prices, calendar
date,occasions and other information to forecast the
future prices of electricity and load.
Key Words: Artificial Neural Network, Electricity
Market, Price Forecasting, Decision Tree, Extreme
Learning Machine.
1. INTRODUCTION
RECENT changes in the electricity industries in several
countries have led to the low regulated and more
competitive energy markets.In this newly structure, price
of the electricity has the key of all the activities in the
market. Price forecasting, with dependable accuracy which
helps power suppliers for selling up rational offers in the
short term.
Generally there are 2 groups of forecasting models,
traditional models and advanced techniques viz as ANN,
Fuzzy logic [1].
Price forecasting models are thetime series and regression
analysis. In the recent years, artificial intelligent(AI)
methods are more commonly used for price forecasting.
Among these methods, ANN method is the simple and
powerful tool used for forecasting. The reason should be
the ANN to learn the difficult input-output relationship
through a supervised training process with the historical
data.
We know that many factors are impacting on the
electricity price, in which some of the factors are more
important than other are, as practically, we consider only
those which are more efficient factors. So, it will be useful
to study which factors affect on price and to up to what
extent. The factors, which affect the electricity price, are
line limit to load pattern, bidding patterns and generator
outages factors. In power market, the load pattern is an
effective attribute on the bidding behavior of Generating
Companies. Therefore we consider the historical price of
system and the load as the factors which affect the price.
The performance of the Artificial Neural Network ANN
approach is greatly affected by the selection of its inputs.
Historical price data and the load have been identified for
electricity price forecasting.
1.1 Price Forecasting By ANN
An ANN is a parallel distributed processor made up of the
simple processing units called neurons. The neurons have
the natural capability for storing experimental knowledge
and make it available for use [2].
Every neural network structure has to undergo the
training of phase with the available data or patterns. This
training / learning phase uses a suitable learning
algorithm. The prime objective of the learning algorithm is
to modified system data weights of the network is in an
orderfashion for attained a desired design objective and to
increase the accuracy of the learning stage minimize the
error. The working of ANN can be divided in two phases
one is training phase and other is recalling phase or
testing phase.
In training phase both the input pattern and its
corresponding target output is supplied to the network.
Input is given to the network at input node; the input layer
neuron processed the input by using activation function
and gives its output. The output from this layer is given as
input to the next level neuron, and so on up to the output
node. The links connected between neurons are having
some weight. These weights are updated by some learning
algorithm in training phase till the error between the
network output and actual output for that input data set or

pattern is minimized. The level of error depends on the
learning algorithm, quality of data and type of network.
Once the minimized error is obtained the other inputs are
given to trained network to get the output. This is the
recall phase or testing phase. The structure of the ANN for
training is shown in figure 1.
This describes the modules, which should be considered to
design a good neural network model for price forecasting
Fig -1: Training of ANN [4].
This describes the modules, which should be considered to
design a good neural network model for price forecasting
1.2 Input Selection
The aim of the input selection in the case of ANN is finding
optimal input parameters. Result of optimal inputs will
lead to smaller ANN with more accuracy and speed will
also be more. Electricity price affecting parameters can be
categorized into:-
Day Type
Historical price data
Amount of Demand.
1.3 Training
The ANN training process requires a set of examples with
proper network behavior. During the training of ANN, the
weights and biases of the Artificial Neural Networks are
repeatedly adjusted to for the purpose of minimizing the
network performance function. The training method
which is selected for the new ANN models is Levenberg-
Marquart back propagation (LMBP), this method updates
the weight and bias values according to L-M optimization.
LMBP is an improved version of Gauss-Newton method. To
deal with the additive noise in the training samples this
method has an extra regularization term. Existing back
propagation methods are often too slow for practical
problems in comparison to LMBP.
Neurons have nonlinear transfer function known as the
“tangent sigmoid" which is present in the hidden and
output layers. The inputs received by a tansig node are
passed through this function to produce an output. These
function generates outputs between -1 and +1.The inputs
of these function should also be in same range. That’s why,
it is necessary to limit the inputs and target outputs of the
artificial neural networks. Both the methods i.e. Mean-
standard deviation and min-max normalization methods
have been tested and min-max method has been selected.
This normalization method provides the advantage of
mapping the target output to the non-saturated sector of
tansig function. This process improves the accuracy of the
training [3].
1.3 Output and Hidden Layer
The ANN models have the output layer. The output is the
hourly price in the model of price forecasting, because of
which only one neuron is present in the output layer .
2. Survey of Existing Research Papers
[1] B.R. Szkuta, L.A. Sanabria, T.S. Dillon ,“Electricity
Price Short-Term Forecasting Using Artificial Neural
Networks”, IEEE Transactions on Power Systems
It presents the System Marginal Price (SMP) short-term
forecasting implementation using the Artificial Neural
Networks (ANN) computing technique. The described
approach uses the three-layered ANN paradigm with back
propagation. The retrospective SMP real-world data,
acquired from the deregulated Victorian power system,
was used for training and testing the ANN. The results
presented in this paper confirm considerable value of the
ANN based approach in forecasting the SMP.
The energy trading levels between market participants,
however, is largely dependent on the short-term price
forecasts. In the Victorian wholesale electricity market,
three levels of trading are recognized: bilateral long-term
contracts, short term forward trading and spot trading.
The bulk of trading is based on a combination of the three.
In this paper we limit this influence to two time intervals
ahead (time lag). The input data temporal set, used for
training the ANN that contains retrospective values of
variables is encapsulated in the vector. All input
parameters are known for all the time intervals included
in the training. The most popular ANN architecture is a
three-layered, feed-forward system with back-
propagation. The success of this configuration dwells in
the fact that it can learn from retrospective information in
a process called supervised learning. In supervised
Learning the network is trained using historical data
derived from the system the relationship between input
and output is to be determined. The back-propagation
learning algorithm, as mentioned earlier, is characterized
by two parameters, the training rate and the momentum.
These were used by the process of progressive refining

weights based on the RMS error minimization. In training,
the network learns by adjusting the weights.
[2]Kishan Bhushan Sahay, M. M Tripathi,” An Analysis
of Short-Term Price Forecasting of Power Market By
Using ANN ”, International Journal of Science.
A neural network approach for forecasting
short-term electricity prices. Almost Until the end of last
century, electricity supply was considered a public service
and any price forecasting which was undertaken tended to
be over the longer term, concerning future fuel prices and
technical improvements. Short-term forecasts have
become increasingly important since the rise of the
competitive electricity markets. In this new competitive
framework, short-term price forecasting is required by
producers and consumers to derive their bidding
strategies to the electricity market. Price forecasting in
competitive electricity markets is critical for consumers
and producers in planning their Operations and managing
their price risk and it also plays a key role in the economic
optimization of the electric energy industry. Accurate,
short-term price forecasting is an essential instrument
which provides crucial information for power producers
and consumers to develop accurate bidding strategies in
order to maximize their profit.
Fig -2: Neural network weight comparisons.
In this paper artificial intelligence (AI) had been applied in
short-term price forecasting that is, the day ahead hourly
forecast of the electricity market price. A new artificial
neural network (ANN) has been used to compute the
forecasted price in ISO New England market using
MATLAB R13. The data used in the forecasting are hourly
historical data of the temperature, electricity load and
natural gas price of ISO New England market. The
simulation results have shown highly accurate day-ahead
forecasts with very small error in price forecasting.
[3] Filipe Azevedo, Zita A , “Forecasting Electricity
Prices with Historical Statistical Information using
Neural Networks and Clustering Techniques” , IEEE
Electrical and Power System.
Factors such as uncertainty associated to fuel prices,
energy demand and generation availability, are on the
basis of the agents major concerns in electricity markets.
Facing that reality, price forecasting has an increasing
impact in agents’ activity. The success on bidding
strategies or on price negotiation for bilateral contracts is
directly dependent on the accuracy of the price forecast.
However, taking decisions based only on a single
forecasted value is not a good practice in risk
management. Artificial neural networks method is used
for finding the market price for a given period, with a
certain confidence level. Historical information was used
to train the neural networks and the number of neural
networks used is dependent of the number of clusters
found on that data. K-Means clustering method is used to
find clusters. A study case with real data is resented and
discussed in detail. At that time, electricity price evolution
was directly dependent on the government’s social and
industrial policy, and price forecasting was mainly focused
on the underlying costs namely, fuel prices and
technological innovation. Any price forecasting made on
that basis was tended to be over the long term. With
electricity markets deregulation and opening to
competition during 1990s, this changed dramatically.
Ownership on this activity sector become private rather
than public or a mixture of both. Competitive markets, like
pools or power exchanges, have been introduced for
wholesale trading. However, due to the instantaneous
nature of the product electricity and uncertainty
associated to fuel prices, energy demand, generation
availability and technical restrictions, are at the origin of
the high volatility that electricity prices exhibit. Also,
factors like the structure and the management rules of any
specific electricity market may introduce other sources of
price volatility. Due to electricity price volatility,
generators face price risks because they sell energy at
variable pool prices while their fuel and other costs may
be fixed. On the other hand, distributors/retailers also face
price risks because they supply most of their costumers at
an annual fixed tariff but they have to purchase electricity
at variable pool price. Facing this reality, electricity price
forecast is therefore extremely important for all electricity
markets players, namely for risk management.
Several works have been made for price forecast
on electricity markets. In those works, we can find a
variety of techniques used to achieve that goal. Koreneff et
al.
Proposed a general use modular time series
approach, with detached regression analysis assuming
that spot price follows a day, a week- and a year cycle.
Stochastic processes were also used to analyse time series
and to forecast electricity price. Electricity market players

are daily confronted with the necessity to make decisions.
Investment decisions are very difficult because they are
influenced by investor’s cognition, investment behaviour,
psychological limitation and even the global economic and
political environment
To hedge efficiently it is imperative to have access to good
price forecasts in order to develop biddings strategies and
negotiation skills. Taking decisions based only on a single
forecasted value is not always a good practice on risk
management, because there are several factors are on the
basis of price formation and some of them are
unpredictable.
3. ELM-TREE Approach
3.1 Extreme Learning Machine
Given a training set X that contains N distinct instances
with n inputs and m outputs, i.e.
X={(xi,yi)|xi=[xi1,...,xin]T∈Rn,yi=[yi1,...,yim]T∈Rm,i=1,...,N
}, the SLFNs with ˜N hidden nodes and activation function
g(x) are formulated as where wj=[wj1,...,wjn]T is the
weight vector connecting the input nodes and the jth
hidden node , bj is the bias of the jth hidden node,
βj=[βj1,...,βjm] T is the weight connecting the j-th hidden
node and the output nodes, and o I is the output of xi in
the network.
The standard SLNFs can approximate the N training
instances with zero error, i.e., there exists wj, bj, and βj,
such that:
. ∑j=1 βj g(wj · xi +bj ) = yi, i= 1, . . . , N.
Algorithm1: Extreme Learning Machine–ELM.
Input: Training Set{(xi,yi)|xi∈Rn,yi∈Rm,i=1,...,N};
activation function g(x);the number of hidden node N.
Output: Input weight wj, input bias bj, and output weight
β.
Steps:-
1. Randomly assign input weight wj and bi as bj
where
j=1,...,N.
2. Calculate the hidden layer output matrix H;
3. Calculate the output weight β=H†Y where H† is the
Moore–Penrose generalized inverse of matrix H [3].
3.2 Extreme Learning Machine with uncertainty
reduction.
In this section, a new classification model named ELM-
Tree is developed to handle the over-partitioning problem.
The key difference between ELM-Tree and traditional DT
lies in the determination of the leaf nodes. In the ELM-Tree
model, ELMs are embedded as the leaf nodes when certain
conditions are met. Thus, there are two key steps in the
construction of ELM-Tree, i.e., splitting a non-leaf node
and determining an ELM leaf node[3].
Input: Node X={(xi,yi)}N
i=1,where xi = {xi1,...,xin} ∈ Rn is the
ith instance, yi∈{1,...,m} is the class label of xi,m is the
number of classes, and Aj, j = 1,...,n is the j-th attribute.
Output: Two child nodes X1 and X2.
1. For each attribute Aj, sort its values x1 j,...,xnj in
ascending order, and the sorted values are recorded as x1j*
≤...≤x*Nj;
2. Get all the available cut-points of each Aj, i.e.,
cutij = x∗ij+x∗i+1,j/2 , i=1,2,···,N−1;
3. Calculate the information gain of each Aj and its
cut-point cutij :
Where Xij1= {xi ∈ X | xij≤ cutij} and Xij2∈ {xi ∈ X| xij> cutij }
are the two subsets of X divided by cutij, |X| is the size of X,
and I (cutij) is the expected information of cutij;
4. For each attribute Aj, select its optimal cut-point
cuti(j)j where
i(j)=argmax{Gain(X,cutij)};
i=1,...,N−1
5. Calculate the split information of the optimal cut-
point for each Aj:
Split(X,cuti(j)j)=
−
6. Calculate the gain ratio of Aj:

Ratio(X, |Aj) =
7. Select the optimal attribute Aj∗ where
j∗=argmax {Ratio(X,Aj)};
8. Split X into X1 and X2 by Aj∗ and its optimal cut-
point cuti(j∗)j∗ where X1={xi ∈ X |xij∗≤cuti(j∗)j∗} and X2 ∈ {xi
∈ X |xij∗>cuti(j∗)j∗}[3].
3.3. Construct an ELM-Tree
The induction of ELM-Tree is described in Algorithm 3.
There are two important parameters in Algorithm 3,i.e.,the
truth level threshold θ ∈ [0,1], and the uncertainty
coefficient
ε ∈ [0,1]. For implementation, we further modify the
uncertainty coefficient as ε × log2(m) for entropy-based
ELM-Tree and ε × ln(m) for ambiguity-based ELM-Tree,
where m is the number of classes in X.
Input: A training set with N instances, n attributes and m
classes; truth level threshold θ ∈ [0,1]; uncertainty
coefficient ε ∈ [0,1];
and integer parameter N* ∈ {1,...,N}.
Output: An ELM-Tree.
Steps:-
Ω is initialized as an empty set;
Consider the original training set as the root-node, and
add it to Ω;
while Ω is not empty do
Select one node from Ω, denoted by X;
if max{pi}m
i=1 > θ or |X| <N* then
Assign X a label, i.e., argmaxi {pi}m
i=1,and remove
it from Ω;
else
if I(cutij) < ε for i=1,...,N−1,j=1,...,n then
Train an ELM for X and remove it from Ω;
else
Split X into two child nodes X1 and X2
by Algorithm2;
Remove X from Ω, add X1 and X2 to Ω;
end
end
end [3].
4. PROPOSED SYSTEM
To forecast (predict) future power need, the electric
power production may be planned. This is called power
load adaption. To reduces the cost of electricity bill and
also to provide estimated usage of the load an accurate
load forecasting is very important. It can be divided into
long-term load forecasting, mid-term forecasting, short-
term forecasting. Midterm and long-term forecasting
mainly used in power factory to help them to forecast the
electricity load usage by their users, and the time ranges of
long term is from six months to twelve months and
midterm forecasting is from one month to six months
respectively. The short-term forecasting can be used in
generators macroeconomic control, power exchange plan
and so on. And the prediction is from one day to seven
days in the future, or a little longer time [1].
Factors influencing System Load usage : The Four major
categories of factors that influence system load are:-
1. Economic Factors: The economic environment in
which the utility operates has a clear effect on the electric
demand consumption patterns
2. Weather factors: Significant changes in load
pattern are due to metrological factors as most of the
utilities have large components of weather sensitive load
such as space heating, air conditioning and agricultural
irrigation
3. Random disturbances: These include loads such
as steel mills, wind tunnels whose operations can cause
large variations in electricity usage.
4. Time factor: The principal of time viz. seasonal
effects, weekly – daily cycle legal and religious holidays
play an important role in influencing load patterns[3].
Fig. 3: System Architecture.

4.1 Modules
The system architecture is divided into two modules:-
a. Data Pre-processing Module
b. Constructing ANN-Model Module
4.1.1 Data Preprocessing Module:-
Feature Extraction:-
The electricity market data comes in the form of a time
series, and does not provide any specific features for use
with ANN. Thus we have to create features from the
available past data to be used as inputs to the ANN. For the
purpose of the forecasting it is very much necessary to
consider both short and long-term forecasting and
sometimes seasonal pattern also become mandatory to be
consider. Because of seasonal patterns sudden changes
might be observed in the price of electricity. So to capture
the changes we sought of creating the features which are
relevant based on the historical data which can lasts for a
longer period. Also other features such as last year same
day, same hour, last week same day same hour price
fluctuation etc. can be created.
Feature Selection:-
For the purpose of robust forecasting system, feature
selection is a very important step.
Using the learning algorithm the estimated accuracy is
obtained based on which evaluation of the feature subset
is done. Cross validation is used for estimation of accuracy.
Unsupervised learning problems can also use cross-
validation which results in better clusters by some other
target or objective function is used instead of classification
accuracy
Fig. 4: Features Selection.
Incorporating Other Features:-
Using historical prices, features are generated which gave
good forecasting results, but for gaining more better
results we have to consider other features which are
indirectly associated with price data for electricity. The
parameters which can affect the load or price of electricity
is also considered.
Normalization
Normalization is compulsory in the ANN where the data
input is different from each other and ranging in different
measurements and scale. In this approach the defined
different input data that has different scales, thus the
normalize the data is used to achieve a consistent form.
Use map min max function available in Matlab to
normalize our input data. Map min max returns a
normalized matrix by normalizing the each row to the
range of provides minimum and maximum values. In our
method we normalize the data into the range of the input
(-1, 0, 1).
4.1.2 Constructing ANN Model Module
Many models of the ANN has been proposed for
purpose of classification and regression (forecasting)
problems in machine learning. Out of all the models, the
multilayer perceptron is the best known and most widely
used. Typically, a feed forward ANN contains unit
arranged in three layers, input, hidden and output. Each
layer contains a number of units which contain shared
data but these are not connected to each other. Our ANN
model we select three layer network with ten nodes in the
hidden layer. To perform the forecasting using a neural
network, 2 basic steps are requiring training and learning.
In the learning step of the neural network learns how to
reconstructing the input and output map by updating the
weight of inputs given and biases at the end of each
iterator. Back propagation is the most common learning
algorithm, in which at the end of iteration output error is
propagated back to input adjusting the weight and biase.
To overcome from the slow convergence rate of the back
propagation algorithm the main two parameters are
learning rate and momentum can be adjusted.
Many models of the ANN has been proposed for purpose
of classification and regression (forecasting) problems in
machine learning. Out of all the models, the multilayer
perceptron is the best known and most widely used.
Typically, a feed forward ANN contains unit arranged in
three layers, input, hidden and output. Each layer contains
a number of units which contain shared data but these are
not connected to each other. Our ANN model we select
three layer network with ten nodes in the hidden layer. To
perform the forecasting using a neural network,2 basic
steps are requiring training and learning.
In the learning step of the neural network learns how to
reconstructing the input and output map by updating the
weight of inputs given and biases at the end of each
iterator. Back propagation is the most common learning
algorithm, in which at the end of iteration output error is
propagated back to input adjusting the weight and biase.
To overcome from the slow convergence rate of the back

propagation algorithm the main two parameters are
learning rate and momentum can be adjusted.
5. CONCLUSIONS
System will have gained high relevance and will represent
one of the main sources of information that can bias one
common base for system user for bidding in the electricity
market. In particular scenarios, such as for generator
companies, forecasting of electricity price and
consumption level can be used as possible value for
finalizing the price level in this competitive market.
The application will provide:
1. Forecasting of Electricity price based on various
parameters such as calendar date, temperatures.
2. Report Generation mechanism on daily basis,
weekly basis.
REFERENCES
[1]M. Ranjbar, S. Soleymani, N. Sadati, and A. M. Ranjbar “
Electricity Price Forecasting Using Artificial Neural
Network”.
[2]Renuka D. Suryawanshi, D. M. Thakore “Decision Tree
Classification Implementation with Fuzzy Logic”.
[3 RanWanga, Yu-Lin Heb,∗,Chi-YinChowa, Fang-Fang Oub,
Jian Zhangb, “Learning ELM-Tree from big data based on
uncertainty reduction“.
[4] B.R. Szkuta, L.A. Sanabria, T.S. Dillon ,“Electricity Price
Short-Term Forecasting Using Artificial Neural Networks”,
IEEE Transactions on Power Systems.
[5] L. Sharma, M. Chakrawarti, A. Dutta, N. Adhikari ,
”Neural Network Based approach for Short-Term Load
Forecasting” , International Journal of Science.
[6] Filipe Azevedo, Zita A , “Forecasting Electricity Prices
with Historical Statistical Information using Neural
Networks and Clustering Techniques” , IEEE Electrical and
Power System.
[7] S. Saravanan, S. Kannan and C. Thangaraj , “Forecasting
India's Electricity Demand Using Artificial Neural
Network” , IEEE- International Conference on Advances In
Engineering, Science And Management.
[8] Ming-Tang Tsai, Chien-Hung Chen , ” A Forecasting
System of Electric Price Using the Refined Back
Propagation Neural Network.”
[9]J. Stuart McMenamin, Ph.D., Frank A. Monforte, Ph.D.
Christine Fordham, Eric Fox, Fredrick D. Sebold Ph.D., and
Mark Quan ” Statistical Approaches to Electricity Price
Forecasting”.

Electricity Price Forecasting Using ELM-Tree Approach

More Related Content

What's hot (14)

Similar to Electricity Price Forecasting Using ELM-Tree Approach (20)

More from IRJET Journal (20)

Recently uploaded (20)

Electricity Price Forecasting Using ELM-Tree Approach