Short Term Load Forecasting Using Artificial Neural Network
Shivam A. Koli
Department of Electrical Engineering, walchand college of engineering, Vishrambag sangli, Maharashtra
416415, India
Abstract.
Short-term electricity demand forecasting is crucial for the effective control and scheduling of power systems.
Traditional econometric approaches, which often assume linear relationships between weather variables and
load demand, may not provide accurate results due to the nonlinear and non-stationary nature of these
relationships. This paper proposes an adaptable technique using Artificial Neural Network (ANN) for load
forecasting. ANNs, utilize supervised learning to map historical data and various inputs to forecast future
loads. The proposed model employs a single-layer neural network with one input layer, one hidden layer, and
one output layer. Key input variables include hour and day indicators and weather-related inputs. The ANN’s
ability to adapt and learn from data makes it a valuable tool for optimizing unit commitment, startup and
shutdown of power plants, control of spinning reserves, and power trading in interconnected systems.
Keywords: load forecasting; artificial neural network.
1.Introduction- commitment, start up and shut down of thermal
plants, control of spinning reserve and buying and
Short-term forecasts of electricity demand selling of power in interconnected systems.an
are needed for control and scheduling of power advantage of the Artificial neural network is that it
systems, and hence they are very important for does not require any assumption between load and
dispatchers. Short-term electric load demand is often weather variables in advance. Moreover, the
considered as a function of weather variables and literature suggests that the accuracy of forecasts is
human social activities. Traditional econometric better than traditional econometric forecasting.
approaches establish functional relationships In this literature we propose a method of
between weather variables and current load demand load forecasting based on a single-layer neural
for the forecasting function, often assuming a linear network and input variables to be used for such
relationship. However, as Park et al. (1991) indicate, neural network.
the econometric approach may not give sufficiently
accurate results because of nonlinear and non- 2.Artificial neural network
stationary relationships between the load data and
weather variables. Therefore, an adaptable Artificial neural network is an information
technique is needed. paradigm that is inspired by biological nervous
An artificial neural network (ANN) systems. The system structure is composed of a large
approach is such an adaptable technique. The number of highly interconnected processing
Artificial neural network uses a supervised learning elements (neurons) working together to solve
technique for training neural networks. The goal of specific problems. Each unit operates only on its
this type of network is to create a model that local data and on the inputs, they receive via the
correctly maps the input to the output using connections. The proposed neural network consists
historical data and various inputs so that the model of one input layer, one hidden layer and one output
can then be used to produce the forecast. These load layer. The neural network maps the input vector X
forecasts play a vital role in optimum unit to the output vector Ô (Fig. 1).
Figure 2: hourly Load curve in India
3.2. Day indicator D(n)
Figure 3:neural network structure.
Load also changes throughout the week it
The output is a sum of the weighted nonlinear terms is clear from the fig.3 that power consumption in the
given by where weekend (day 6,7) drops compared to power
consumption in the weekdays. To take this into
account while calculating load forecasting, variable
associated with day of week D(n) where n changes
from 1 to 7 representing days of week
Wᵢⱼ (i = l,..., n ; j = l ,..., nᵢ) is the weight factor
between the ith input node and the ith hidden node;
v ᵢ is the threshold value at the ith hidden node
defined by the algorithm; and xⱼ is the jth input
element.
The learning of a network is a recursive
process which adapts the connection strengths, i.e.
the weights among the nodes and the threshold
values at each iteration. The error function E is to
update the weights at each recursive minimum:
where K is the number of cases used to train the Figure 1: load changes during week
network O(m) is the actual output of the training
case m, and Ô(m) is the forecasted output of the 3.3. Weather indicating variables
network.
Weather variables primarily include
3.selection of input variables temperature, wind velocity cloud coverage, rain etc.
out of all the above variables temperature seems to
The most important work in building an have strongest correlation to power consumption as
Artificial neural network load forecasting model is displayed by Fig.4. Although wind velocity and
the selection of input variables. Selection of inputs cloud coverage have their own effects on load due to
is based on judgement of experience of the engineer its effect being very small no significant relation is
and it is carried out by trial and error since there is seen. therefore, only temperature as a weather
no rule of thumb that can be followed to determine variable is considered while calculating load
which inputs to use. However, some statistical forecasting
analysis can be done to determine the variables with
most influence on power consumption. These inputs
are (a) hour and day, (b) weather associated inputs
and (c) historical loads
3.1. Hour indicator H(i)
Load changes throughout the day from one hour to
another. On a typical day in India peak load occurs Figure 3: load changes during week
generally between 06:00PM to 08:00PM. And
minimum load occurs at 03:00AM(Fig.2). We can
conclude that the magnitude differs between various
hours by a large margin there for an hour indicator
H(i) (where i changes between 1-24) is very helpful
while calculating short term load Figure 4: monthly power consumption and temperature
graph
3.5. Normalized data 6.limitations of ANN
Normalization or scaling is not really a A major problem in modelling load
functional requirement for the neural networks to forecasting is that there are usually many trends in
learn, but it significantly helps as it transposes the the data. For example, there may be a seasonal trend
input variables into the data range that the sigmoid reflecting the seasons of the year. As well as there
activation functions lie in (i.e. for logistic [0, 1] and may be monthly and/or weekly trends in the data as
tanh [-1, 1]. And it makes it so that all inputs are in well. These trends make it difficult to correctly
comparable range model load forecasting data.
Each input data is multiplied by an Artificial neural networks can become too
amplitude and shifted by an offset. The amplitude complex because of their architecture and the huge
and offset are often referred to as normalization datasets used to train them. They can also memorize
coefficients. The normalization coefficients are the training data. These can lead to poor
computed based on the minimum and maximum generalization of new data.
values found across all of the data sets. The formulae
used to compute these coefficients are They can also be vulnerable to adversarial
examples or minor changes in the input data. These
changes can cause the particular artificial neural
network to make incorrect decisions and irrelevant
where Max(i) and Min(i) are the maximum and outcomes.
minimum values The input components normalize
7. Conclusion
the data using the following formula:
The application of Artificial Neural
Networks (ANNs) for short-term electricity load
4.Training of ANN forecasting presents a significant advancement over
traditional econometric methods. By leveraging the
Training is a process to find the input–
adaptability and learning capabilities of ANNs, this
output map by repeatedly analyzing the training set
approach addresses the nonlinear and non-stationary
as the network learns. The parameters of the network
relationships between load demand and influencing
obtain optimal values during this process and the
variables such as weather and human activities. The
performance checked with the test set. The
proposed single-layer neural network model, which
ingredients for supervised learning are the input, the
incorporates key input variables like hour and day
desired response, the definition of error and a
indicators and weather-related inputs demonstrates
learning law.
improved forecasting accuracy.
Training can be implemented in two ways.
This enhanced accuracy is critical for
Either we present a pattern and adapt the weights
optimizing power system operations, including unit
(on-line training), or we present all the patterns in
commitment, thermal plant management, spinning
the input file (an epoch), accumulate the weight
reserve control, and power trading. The findings
updates, then update the weights with the average
underscore the potential of artificial neural networks
weight update, called batch learning.in proposed
to enhance load forecasting, offering a efficient and
ANN backpropagation can be implemented to
flexible solution that can adapt to the dynamic
effectively train the artificial neural network.
nature of electricity demand.
Backpropagation is an iterative algorithm, that helps
to minimize the cost function by determining which Bibliography
weights and biases should be adjusted. During every
epoch, the model learns by adapting the weights and (1) Ongsakul, Weerakorn_ Vo, Dieu Ngoc - Artificial
biases to minimize the loss by moving down toward intelligence in power system optimization (2013,
the gradient of the error. Thus, it involves the two CRC Press, Taylor & Francis)
most popular optimization algorithms, such
as gradient descent or stochastic gradient descent. (2) P.K. Dash et al./Electric Power Systems
Research 32 (1995) 19-23
5. Performance measures of the network
(3) O.A.S. Carpinteiro et al. / Applied Soft
In order to evaluate the performance of the Computing 4 (2004) 405–412
network model, the following common performance
measures are used automatically by the software: 4) N. Kandil et al. / Electrical Power and Energy
mean square error (MSE), normalized mean square Systems 28 (2006) 525–530
error (NMSE), correlation coefficient, % error, and (5) S. C. Bhattacharya and L. T. Tha Short - Term
the minimum description length (MDL). The most ELECTRIC LOAD forecasting
common measure is to minimize MSE in the
forecasting but this measure has some drawbacks. (6) S.M. Islam et al./Electric Power Systems
Alternative measures such as percentage error, mean Research 34 (1995) 1-9
percentage error (MPE) and mean absolute
percentage error (MAPE) are also presented in this (7) M. Ghiassi et al. / Electric Power Systems
study Research 76 (2006) 302–316
