FUZZY AND DECISION TREE APPROACH FOR FORECASTING ANALYSIS IN

POWER LOAD

P. Santhosh Kumar, K. Kumaran @Kumar,

Student, Department of CSE, Assistant Professor, Department of CSE,
Pondicherry Engineering College, Pondicherry Engineering College,
[email protected] [email protected]

ABSTRACT
The most important challenges in electric load forecasting is to find the accurate
electricity load forecasting. Because, it is volatile in nature and has to be consumed immediately.
Fuzzy Decision Tree is applied to predict the annual electricity requirement in India. Population
and Per Capital gross domestic product (GDP) are taken as input variables and the electricity
consumption is predicted output variable. Past 30 years of historical data has been used for
training and 4 year of data is used for testing the fuzzy decision Tree. Comparatively to measure
the accuracy of data has been made with Artificial Neural Network and Fuzzy Decision Tree
model using Mean Absolute percentage error (MAPE). The results states that proposed Decision
Tree model has given high performance and less error rate than the Artificial Neural Network
model.
Keywords: Load forecasting, fuzzy logic, decision tree.

I. INTRODUCTION
Nowadays, the demand for electricity is increasing tremendously. Hence, it’s important
to predict the power consumption based on the statistics of population growth and gross
domestic product values. Neural Networks can be used to extract the patterns and detect the
trends that are too complicated by various techniques. To solve the forecast problem using fuzzy
decision tree model and applied this technique in real world dataset provided in the web[13,14].
1.1 Type of forecasting:
Short Term load forecasting (STLF):
 Short term load forecasting is to predicting and forecast the electric power consumption
using in per day. STLF is important role for utility operations and real-time control for security
functions.
Medium Term load forecasting:
Medium term load forecasting is to predicting and forecast the electric load for one day to
one year.
Long term load forecasting:
Long term load forecasting is to predicting and forecast the electric power consumption
using more than one year. LTLF is important role in generation, transmission and distribution
operations.

II. RELATED WORK

2.1 ARTIFICIAL NEURAL NETWORK

Artificial Neural Network are real world problem in modeling complex. A neural
network is a machine learning approach that is inspired by brain at particular learning task [1,2].
ANN has fewer connection for processing elements (neurons) to solve specific problems. A
neural network can be trained and adjusting the values of weights to perform particular function
between elements so that a particular input leads to a specified target output[1,3]. Therefore, the
network is adopted, based on a comparison of the output and the target. In ANN model [2],[3][4]
depicts the values with the number of hidden layers and the number of neurons present in the
hidden layer. This calculated value provides the error rate for the electricity consumed for the
particular time period.

2.2 FUZZY LOGIC

A fuzzy logic [6] system is to handle numerical data and linguistic knowledge. It is a non
linear mapping of an input data vector into a scalar output. i.e. mapping numbers into numbers.
A fuzzy linear regression model [6], [7] for summer and winter seasons is developed.
The estimation fuzzy problem for the model is turned out to linear optimization problem, fuzzy
linear regression, and is solved using the linear programming based simplex method. The
parameters of fuzzy, the fuzzy load for one day ahead is predicted. Results of predication of 24
hour load ahead for a utility company are presented. It has been found using such fuzzy model; a
reliable operation for the electric power system could be obtained.

III. EXISITING MODEL

Input Layer Hidden Layer Output Layer

Figure 1: Structure of ANN model
An ANN as a system consists of fewer connection of elements to solve particular task.
These interconnecting elements in an ANN are known as neurons [9], [10]. Each neuron may
have several inputs and only one output. The output of the input layer is send to the hidden layer,
which in turn sent to the output layer. The multilayer perceptron’s with back propagation
learning algorithm [5] is used in ANN model because it is suitable model for long term
forecasting, which helps in the non linearity incomplete or noisy in the input data. The network
can be trained by adjusting the weight between the neurons. Therefore, the neural network is
accepted, based on a similarity of the target and the output. Back-propagation is a regular method
of training multilayer artificial network. The inputs are passed to the network, then the output is
calculated based on the network selected and the output is compared with the targets. If the
output obtained output is not closer to the target value then the learning rule is used to adjust the
weights between the neurons in order to get network outputs closer to the targets. Fig. 1 shows
the structure of ANN model.

IV. PROPOSED MODEL

The development of Fuzzy Decision Trees (FDT) differs from traditional decision trees in
two respects: it uses splitting criteria based of fuzzy restrictions and its inference procedures are
different [ 12]. We are adopting the following three steps, adapted from, for developing our FDT:
(i) Fuzzifying the training data; (ii) Inducing and pruning the generated tree; and (iii) Applying
fuzzy rules represented by the tree for forecasting [12].
i. Fuzzifiying Training data

The analysis based on 30 years yearly electric load. The data covered GDP and
population from that we extract the six sector of electric load.

Based on careful analysis of data and the information elicited from the experts of six
sector. we selected the trapezoidal membership function due to its simplicity and
implementation. Therefore we designed the suitable membership function and the set of
linguistic terms for each of the variable.

Splitting dataset into training data and test data

Analyzing training data by a decision tree algorithm

and generating decision tree

Estimating the accuracy of obtained decision tree using

test data

Accuracy is
considered
acceptable?
Y

Y
Applying decision tree to future data for predict the
output
 Figure 2: The Process of Decision Tree Generation

When fuzzify these rules, the result will be forecasted and to be particular instance. From
analysis of load we assigned the minimum and maximum load. The forecasted load was
categorized using numerical value from 0.00(minimum) to1.50(maximum).
Given points X1,x2 , x3, x4as shown in figure 1, the membership function for the various variable
is defined by the equation.

1.0

X1 x2 x3 x4
Figure 3: Trapeziodal membership function template

X2-x X 1≤ x ≤ X 2
X2-X1

µv (x) = 1.0 X 2≤ X ≤ X3
X4-x X 3 ≤ x ≤ X4
X4-X3

ii. Inducing and puring the fuzzy decision tree

The procedure FDT is comprise three elements: (i) selection of splits at every new node
of the tree, (ii) a rule for determining when a node should be considered as the terminal and (iii)
a rule for assigning labels to identified terminal nodes.
A number of heuristics have been proposed for building and pruning fuzzy decision trees [12].
We are adapting the method developed in [12] due to itssimplicity and the availability of
software for its implementation.The algorithm for building the tree is presentedas follows:

Fuzzy Decision Tree Algorithm

G=GetData();[gdp,population]

TreeBulit=F

While!TreeBulit {

XRoot=Weight

While!EndofNodes() {

Compute the data

PKM=∑|G|j=1 fk+1(XjM,µvkc(yj))

PM=∑|Dc|k=1pkM

Computing the std information

IM=-∑|Dc|k=1(pkM * log pkM)

pkM pkM

Nodes I {

search the set of remaining attribute from v-vi by

Computing IsMVi

Selecting attribute Vi such that info gain GiM is maximal

If(XjM>0.0 has unique claisfication) or ( vM=v))

TreeBulit=T

Split M into |Di| sub nodes by making

Child M|Vpi get samples defined by XM|Vpiand

Computing new membership using fuzzy restrictions

Leading to M|Vpi using equation

 XjM|Vpi= f1 (f0(ej,Vpi),XjM)

Variable/ parameter Description

V ={V1, V2, V4, V5, V6, V7} The set of seven fuzzy input variables
i i i
D = D 1D 2,D 3 The set of terms defined over variableVi.
vip The fuzzy term p for variable Vp,e.g. μTempLow
ui∈Ui The crisp data for the variableVi.
Dc The fuzzy terms for the Fload decisionvariable.
E = {ej|ej =u1j, u2j, . . . ,unj, yj} The training example database, eachitem in the database is
. viewed as aevent
W = {wj} The confidence weight.
wj The weight of ej∈E.
M Nodes in the fuzzy decision tree.
FM The set of fuzzy restrictions on thepath leading toM.
VM The set of attributes appearing onthe path leading toM.
XM = XM j The set of memberships in M for allthe training examples.
M|vip The particular child of node M createdby using V i to

The set of M’s children whenVi∈V − V N is used for the

splitM.
SMV
split.
PM The total example count for node M.

The example count for decisionvck∈Dc in nodeM.

M
I Information measure for node M.
PNk
GMi= IM− ISMVI The information gain when usingVi inM.
μ(.) : X → [0.0, 1.0] A mapping from X to[0.0, 1.0].

In addition to the algorithm, the parameters, variables, and functions used in our FDT design are
listed .The training data were generated from the yearly electric load for n-1 data, as well as the
nth were used as a test data. In this model the entire database (E) is divided into two sets: (i)the
training and (ii) test sets. The training set is further divided into two other disjoint sets. One of
these sets was used for building the tree while the other was used for pruning the resulting tree.
The procedure of Fuzzy decision tree utilizes the training data and added in a top-down fashion,
until the tree nodes are successfully to stopping criteria are met. The resulting tree is optimized
by a pruning procedure which acts in a bottom-up version, to remove irrelevant parts of the tree.
To deal with missing values, an example is split into all children if the need feature value is not
available. The percentage of the event with unknown values is reduced by attribute utilization.
If Puiunknown denotes the total count of examples in node M with unknown values for Vi, then
the split function fr() at splitting point r is defined as:

fr(ei, [Vi is vip]) = 1.0 if uij unknown,

|Di|

μvip(uij) otherwise
.

Note that |Di| is the cardinality of set Di, which is the number of elements in the set. In our model
to extract the electric by GDP and Population.

V. EXPERIMENTAL RESULT

30 year historical data of power consumption, Population and GDP has been used for
prediction. The historical data for population ,GDP and power consumption was taken from
webpage [13],[14],[15]. Table 1 shows comparison of the existing ANN model and proposed
FDT forecasted values for the year 1986, 2000, 2006, and 2010. In ANN model the hidden
layer eighth neurons gives the best result among other neurons. FDT model is constructed and
implemented using the same input GDP and Population which is used in ANN model. To
evaluate the performance of proposed FDT model results are compared with neural networks
with back propagation algorithm. Mean Absolute Percentage Error (MAPE) is used for the
performance metrics.

MAPE=Actual value – Forecasted Value *100

Actual value

Table 1 Power Consumption: Actual Vs Forecasted

Year Actual output(BkWh Forecasted %error

 Output(BkWh)

ANN FDT ANN FDT

1986 123.09 123.26 122.58 0.14 0.41

2001 316.60 314.98 316.54 0.51 0.018

2006 415.29 415.82 415.36 0.13 0.016

2010 612.64 613.59 612.35 0.16 0.047

Average absolute error 0.94 0.49

Table 1 shows power consumption of actual and forecasted value using artificial neural network
and FDT model for the year 1986, 2001, 2006 and 2010 and also it shows the error rate of actual
and forecasted value. Table 2 shows power consumption of future years upto 2019 using ANN
and FDT model

Figure 4. Power Consumption: Actual Vs Forecasted

Table 2 Forecasted Future Power Consumption: ANN and FDT model

Forecasted Output (BkWh)
Year
ANN FDT

2016 716.4 912.4

 2017 758.38 979.2

2018 802.52 1048.0

2019 848.0 1118.5

VI. CONCLUSION
The power consumption of India is forecasted by using the social factors: GDP and
Population. The two model FDT and the ANN model of multilayer perceptron (MLP) with back
propagation network of different layer with different node have been studied and implemented
using 30 years historical data and it has been tested with 4 years historical data. The proposed
fuzzy Decision Tree model is implemented using the same historical data which is used in the
ANN model. The result of both models is validated using the MAPE. The MAPE obtained by
ANN model is 0.94 and the proposed fuzzy Decision Tree Model is 0.49. To conclude based on
obtained result that the consumption using the fuzzy Decision Tree model is efficient and more
accurate than the ANN model.

REFERENCES
[1].S.Saravanan, S.Kannan, and C. Thangaraj, “Forecasting India’s Electricity Demand Using
Artificial Neural Network “, IEEE International conference 2012.
[2].K. Hornik, M. Stinchcombe, and H. White, “Multilayer feed forward networks are universal
approximators,” in the proceeding of the IEEE Transaction on Neural Networks, vol. 2, no. 5,
pp. 359–366, Oct 1995.
[3]. K. Hornik, “Approximation capabilities of multilayer feed forward networks,” in the
proceeding of the IEEE Transaction on Neural Networks, vol. 4, no. 2, pp. 251– 257, Jul
1991.
[4]. M. Leshno, V. Y. Lin, A. Pinkus, and S. Schocken, “Multilayer feed forward networks with
a no polynomial activation function can approximate any function,” in the proceeding of the
IEEE Transaction on Neural Networks, vol. 6, no. 6, pp. 861–867, Aug 1993.
[5]. Manish Sankar, B. Yeganarayana, DeepakKhemani“Backpropagation learning algorithms
for classification with fuzzy mean square error” Elsevier 1998.

[6]. RustumMamlook,OmarBadran ,EmadAbdulhadi, “A fuzzy inference model for short-term

load forecasting ” a Middle East University for Graduate Studies, Elsevier 2009.
[7]. D K Ranaweera, N F Hubele and G GKarady, ”Fuzzy logic for short term load forecasting”
in Elsevier transacting on Electrical & Energy Systems Vol.18 No, pp.215-222,1996.
[8]. A. Ghanhari, A. Naghavi, S.F. Ghaderi, and M. Sabaghian, "Artificial Neural Networks and
regression approaches comparison for forecasting Iran's annual electricity load," Proceeding
International Conference on Power.
[9]. BahmanKermanshahi, Hiroshi Iwamiya, Enviroinmental Engineering,” Up to year 2020
load forecasting using neural nets”, Elsevier 2002.
[10].ParasMandal, TomonobuSenjyu, Naomitsu Urasaki, ToshihisaFunabash,”A neural
network based several-hour-ahead electric load forecasting using similar days approach”
Elsevier 2006.
[11]. Chih-chouchiu, deborah f. cook ,jen-lung kao and yu-chaochou ” Combining a neural
network and a rule-based expert system for short-term load forecasting ” Elsevier 1997.
[12] .Stephan Trenn “Short-term electric load forecasting in uncertain domain” in the proceeding
of the IEEE Transaction on Neural Networks, Vol.19. No.5, MAY 2011.
[13]. http://www.populstat.info/
[14]. http://www.indexmundi.com/india/gdp_(purchasing_po wer_parity).html
[15]. http://www.iasri.res.in/agridata/12data/chapter1/db2012 tb1_13.pdf