1. 1
Abstract: - Demand forecasts are important for energy
suppliers, electricity generation, transmission,
distributors. Since the electricity demand is volatile in
nature; it cannot be stored and has to be consumed
instantly. The aim of this study detail with electricity
consumption towards forecast future projection of
demand. The input variables used are Population, GDP,
gross national income, Industry, Consumer price index
and Per capita power consumption. The approach
method using artificial neural network (ANN) based on
Feed-Forward neural network, trained by using
incremental backpropagation (BP) algorithm.
Forecasting capability is evaluated by calculating the
Mean Square Error (MSE).
Index Terms: Electricity Demand, Feed-Forward N.N,
backpropagation, Root Mean Square Error.
I. INTRODUCTION
Expanding access to energy means including 1.2
billion people: 0.6 billion that still have no access to
electricity (81% of whom live in the rural areas) and
0.4 billion that only has access to unreliable electricity
networks. We need machine technique and practical
approaches, as a mechanism that receives inputs
directly and transmits to development, electricity plays
a central role in both fighting poverty and addressing
climate change. Medium and long-term load fore-
casting using Artificial neural network algorithm has
achieved satisfactory results [2-5]
.
The implementation tasks can be divided into four
headings:
I. DATA ENCODING AND NORMALIZATION
II. ARTIFICIAL NEURAL NETWORK
III. TRAINING
IV. COMPUTING ACCURACY
This subdivision is data pre-processing and sent to
Feed-Forward neural network and training using
incremental back-propagation is used to predict the
following year electricity load and it is implemented
in visual studio 2015 community using C# language.
II. DATA ENCODING AND NORMALIZATION
One of the essential keys to working with artificial NN
is data encoding and normalization.
After all data has been collected we may have some
data in non-numeric data and high magnitude data.
After all data has been converted to Encoding non-
numeric output data to numeric values by using 1-of-
N encoding, encoding non-numeric input data to
numeric values by using 1-of-(N-1) encoding and
apply normalization to high magnitude using any one
out of two different types of normalization, Gaussian
normalization and min-max normalization.
Fig. 1: Data normalization
In general, the min-max normalized value for some
value x is (x - min) / (max - min)— very simple, the
Gaussian normalized value for some value x is (x -
mean) / standard deviation.
III. ARTIFICIAL NEURAL NETWORK
An artificial neural network will accept one or more
inputs and produces one or more outputs. The basic
neural network input to process (hidden layer) and
process to output computation is known the feed-
forward NN mechanism.
Electricity Demand Forecasting Using ANN
K Naren Chandra (15MCC1023) *
Department of Computer Science Engineering, VIT University, Chennai-600128, India
Phone: 9790723239*, E-mail:kattlanaren.chandra2015@vit.ac.in

2. 2
Fig. 2: Neural Network Architecture
Understanding the feed-forward NN mechanism is
important to know how to create NN, which will make
predictions.
A NN computes in several stages to get output and the
feed-forward NN mechanism is explained by using
Fig. 2, there are six inputs each line connecting one
node to another represents a weight constant. The
weight is labeled IHWEIGHT [0] [0] it represents the
weight from input GDP to hidden 1 and the weight in
the lower right corner is labeled HOWEIGHT [12] [0]
it represents the weight from hidden 7 to output. All
hidden and output layer nodes has an arrow pointing
to it this are called the bias values.
The first stage in the feed-forward NN mechanism is
to compute for all hidden nodes. The value of hidden
1 node and is computed as the product of each input
value and its associated weight are summed then
associated bias value is added.
From Fig. 3 Pre-activation sum function applied by
the activation function, it will be explained in detail
next section, but for now it's enough to say that the
activation function is the hyperbolic tangent function,
which is usually abbreviated tanh.
Hidden 1 = tanh (pre-activation sum)
After all, hidden node output values have been
computed, these values are used as inputs to the output
nodes and it’s computed slightly differently from the
hidden nodes. The preliminary output sums, before
activation, for output nodes are computed the same
way as hidden node sums, the activation function for
the output layer is called the softmax function.
For 1ouput:
output = exp (pre-activation output sum) / exp (pre-
activation output sum)
For n output:
output = exp (pre-activation output sum) / sum of all
sum of all= exp (pre-activation output 1 sum) +…...+
exp (pre-activation output n sum)
Fig. 3: Activation Functions
If we look at Figures 2, we will see these are the final
output values computed by the neural network. Where
exp function is the math constant e = 2.71828. . . raised
to the xth power. Notice the output values sum is 1.0,
which is not a coincidence that is the point of using the
softmax function.
A. Activation Functions
In above the hyperbolic tangent function for hidden
layer node activation and the softmax function for
output layer node activation. There is a third common
activation function called the logistic sigmoid function
In general, the hyperbolic tangent function is the best
choice for hidden layer activation. For output layer
activation, if neural network is performing
classification where the dependent variable to be
predicted has three or more values softmax activation
is the best choice. If neural network is performing
classification where the dependent variable has
exactly two possible values the logistic sigmoid
activation function is the best choice for output layer
activation.
IV. TRAINING
The ultimate goal of a neural network is to make a
prediction. In order to make a prediction, a neural
network must first be trained. Training a neural
network means finding a set of good weights and bias
values so that the known outputs of some training data
match the outputs computed using the weights and

3. 3
bias values. The resulting weights and bias values for
a particular problem are often collectively called a
model and it’s used to predict the output for unseen
inputs that is computed for known output values.
The data set consists of a total of 5 items data set and
splits it randomly into a 4-item subset (80%) to be used
for training and a 1-item subset (20%) to be used for
testing, that is, to be used to estimate the probability of
a correct classification on data that has not been seen
before. Training a 6-7-1 fully connected feed-forward
NN using the back-propagation algorithm in
conjunction with a technique that is called incremental
training.
After training has completed, the accuracy of the
resulting model is computed on the training set and on
the test set.
A. Incremental Training
There are two main approaches for training a neural
network. As usual, the two approaches have several
different names but two of the most common terms are
Incremental and Batch. In high-level pseudo-code,
Batch Training is:
loop until done
o for each training item
 compute error and
accumulate total error
o end for
o use total error to update weights and
bias values
end loop
The essence of batch training is that an error metric for
the entire training set is computed and then that single
error value is used to update all the neural network's
weights and biases.
In high-level Pseudo-code, incremental training is:
loop until done
o for each training item
 compute error for current
item
 use item's error to update
weights and bias values
o end for
end loop
In this training, weights and biases will be updated for
every training item. Based on some research evidence
suggests that incremental training given better results
than batch training.
In all training approaches there are two important
questions: “How do we determine when training is
finished?” and “How do we measure error?” will be
explained in detail main logic section.
B. Creating Training and Test Data
One of the strategies used in NN training is called
hold-out validation. It’s means to take the available
training data sets, and instead train the NN using all
the data, remove a portion of the data, typically
between 10–30%, to be used to estimate the
effectiveness of the trained model. The remaining part
of the available data is used for training.
Method Make-Train-Test accepts a matrix of data and
returns two matrices, the first return matrix holds a
random data 80% of the training data sets and the
another matrix is one that holds the other 20% of the
test data sets.
Algorithm for Make-Train-Test is:
 compute number of rows for train matrix: =
n1
 compute number of rows for test matrix: = n2
 make a copy of the source matrix
 scramble using the Fisher-Yates shuffle
algorithm the order of the rows of the copy
 now copy first n1 rows of source copy into
train
 copy remaining n2 rows of source copy into
test
 return train and test matrices
It’s accepts a data matrix to split, and a seed value to
initialize the randomization process. The method
returns the resulting train and test matrices as out
parameter.
After creating the training and test matrices, it sent to
neural network. The number of input and output nodes
are determined by the source data: six numeric x-
values(input) and one y-value(output). Specifying a
good number of hidden nodes is one of the important
when working with neural networks. Even though
there has been much research done in this area, picking
a good value for the number of hidden nodes is mostly
a matter of trial and error.
C. Main Logic
After data preprocess and a NN instantiated, and
trained by using Max Epochs, Learning Rate,
Momentum, Exit Error. The Max Epochs variable sets
a hard limit on the number of iterations that will be
performed in the training code and it will not
determine when to stop the main training. The
Learning Rate and momentum values control the
speed at which the backpropagation converges to a
final set of weight and bias values if it’s turns out,

4. 4
back-propagation is extremely sensitive to the values
used for the learning rate and momentum. Even small
changes in these value can have very large effects
difference between a NN that predicts well and one
that predicts poorly.
For simplicity, the neural network has a fourth
parameter, an early-exit error threshold value of 0.040
that is hard-coded into Training.
The heart of Training has two exit conditions. The first
is a hard limit based on the Max Epochs value. The
second exit condition is a small error threshold based
on the mean squared error of the current model over
the entire training set. If anyone exit conditions
satisfied the Training will stop.
One interesting approach to the loop-exit issue is to
divide all available data into three sets rather than just
training and test sets. The third set is called the
validation set. For example, using a 60-20-20 split the
data set.
D. Mean Square Error
Training the neural network using the training data,
but inside the main training we should compute the
MSE on the validation data set. After some time, the
MSE on the validation set would drop but at some
point, when over-fitting begins to occur, the MSE on
the validation set would begin to increase. We should
identify at what epoch the MSE on the validation set
began to steadily increase and then use the weights and
bias values at that epoch for the final neural network
model. And then we should compute the accuracy of
the model on the unused test data set to get an estimate
of the model's accuracy for new data.
Training computes and checks the MSE value on
every iteration through the training. Computing MSE
is a will more time for each iteration, so an alternative
is to compute MSE only every 10 or 100 or 1,000
iterations.
After training has completed, the internal weights and
bias values that were determined by the training
process.
V. COMPUTING ACCURACY
The Result of the NN on the training data and testing
data is computed only after training has completed.
The accuracy of the test data is the very important out
of two accuracy result. The accuracy of the test data
gives us an estimate of the accuracy of the model on
new data with unknown y-values(output). A high
accuracy on the training data does not necessarily
indicate a good training, but a low accuracy on the test
data almost always indicates a poor training.
Fig. 4: Computing Accuracy Output
Method Accuracy uses a helper Max Index to find the
location of the largest computed y-value is correct or
incorrect prediction.
VI. CONCLUSION
Artificial NNs to forecasting electricity problems has
been studied in recent times, and so many number of
papers are report successful experiments and has been
published. However, not all the authors and
researchers in forecasting has not implemented in C#
language. Which has more advantages, in training we
can detected any inputs and data set has given wrong
by using diagnostic tools in visual studio 2015 if data
sets are very large can be stored in database and we
can new data sets to database, which makes efficient
computing and training. If we take less number of data
sets, we can’t make train Artificial NN as we seen Fig.
4 accuracy training and accuracy test data result is very
less compared to one.
REFERENCES
[1] Arunesh Kumar Singh, Ibraheem, S.
Khatoon, Md. Muazzam, D. K. Chaturvedi
“An Overview of Electricity Demand
Forecasting Techniques,” Vol.3, No.3, 2013-
National Conference on Emerging Trends in

5. 5
Electrical, Instrumentation &
Communication Engineering.
[2] R. Xu, G. K. Venayamoorthy, and D. C.
Wunsch, “Modeling of gene regulatory
networks with hybrid differential evolution
and particle swarm optimization,” Neural
Networks, Vol. 20, pp. 917– 927, 2007.
[3] Azadeh, S. F. Ghaderi, S. Tarverdian, and M.
Saberi, “Integration of artificial neural
networks and genetic algorithm to predict
electrical energy consumption,” Applied
Mathematics and Computation, Vol. 186, pp.
1731–1741, 2007.
[4] F. Hobbs, S. Jitprapaikulsarn, and S. Konda,
“Analysis of the value for unit commitment
of improved load forecasts,” IEEE Trans on
Power Systems, Vol. 14, no. 4, pp. 1342–
1348, 1999.
[5] S. A. Billings, and X. Hong, “Dual-
orthogonal radial basis function networks for
nonlinear time series prediction,” Neural
Networks, Vol. 18, pp. 479–493, 1998.
[6] L. Yue, Y. Zhang, Q. Zhong, and Z. Wang,
“Mid-long term load forecasting based on
fuzzy logic clustering neural network
approach,” Journal of North China Electric
Power University, Vol. 35, no. 2, pp. 42–46,
2008.
[7] A. Kazemi et al., “A Multi-Level Artificial
Neural Network for Gasoline Demand
Forecasting of Iran”, In Second International
Conference on Computer and Electrical
Engineering, pp. 61-64, 2009.
[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 Engineering Energy and Electrical
Drives. pp. 675-679, 2009.
[9] Sackdara, et al., “Electricity Demand
Forecasting of Electricity DU Lao (EDL)
using Neural Networks”, TENCON 2010, pp.
640-644, 2010.