Sunteți pe pagina 1din 2

Artificial Neural Network Method in Load

Forecasting

1st Ian Benedict Pongcol 2nd Sean Clark Dominguez


Department of Electrical Engineering and Technology Department of Electrical Engineering and Technology
College of Engineering and Technology College of Engineering and Technology
Mindanao State University - Iligan Institute of Technology Mindanao State University - Iligan Institute of Technology
Iligan City, Philippines Iligan City, Philippines
ianbenedict.pongcol@g.msuiit.edu.ph seanclark.dominguez@g.msuiit.edu.ph

rd
Neilmer Lahoylahoy
Department of Electrical Engineering and Technology
College of Engineering and Technology
Mindanao State University - Iligan Institute of Technology
Iligan City, Philippines
neilmer.lahoylahoy@g.msuiit.edu.ph

Abstract—This document presents the capability of the Arti- provide data that can be used with other utilities for possible
ficial Neural Network (ANN)) in load forecasting. In machine energy interchange[1].
learning, ANN is one of the most widely used tools. The neural In detail, ANN basically acts as a human brain.A neural
network learns from its recieved data, implying that its reliability
increases as the amount of data it processes. Having high network is a massively parallel processor consisting of easy
accuracy rate in load forecasting is paramount in power system processing units, known as neurons. This network comprises
operation. Due to the absence of actual data, synthetic data of a number of layers comprising neurons and weights
is used in training the neural network instead. The neural connected with the neuron association where the data is
network was trained repeatitively to have a good fitting condition carried in feed-forward way.In Fig 1, an artificial neuron
in training, validation and testing. Simulation have shown the
comparison between target and forecasted load data. model is shown. ANN neurons comprises of three primary
Index Terms—Artificial Neural Network(ANN), forecasting, components:-weights linking the nodes, function summation
Power System Operation, synthetic data within the node, and function transfer. I input layer, I con-
cealed layer and I output layer are present in this model,
I. I NTRODUCTION as this MLP structure is demonstrated to tackle nearly any
kind of non-linear connection.The neural network comprises
The Artificial Neural Network (ANN) is a computational of a two-layer feed-forward network with sigmoid activation
method that mimics the thought process of the brain. It is function in concealed neurons and linear in output neurons.
composed of the input, output, and hidden layers. The hidden First, feed-forward data is passed through input to output,
layers utilize nodes, connected through weights, to calculate then weight is adjusted through back propagation. Weight and
the output. Its structure is similar to how neurons are attached bias adjustment is called learning by the Levenberg-Marquardt
to other neurons. The ANN is also capable of self-correction back propagation algorithm. [2]
by manipulating the weights between its nodes.
This computational model is prevalent in the areas of image II. F ORMULATION
processing and load forecasting. Though it might not be A. Node Calculations
reliable on fields that has minimal to no avalable data. This is
In neural network node calculations, the general formula
because ANN relies on the amount of available data, which
below is used to calculate for the ouput data of each node.
the neural network trains upon repeatitively.
n
When it comes to load forecasting, different energy sys- X
tem load forecasting methods have been suggested in recent Xo = Xi Wi − b (1)
1
decades. Load forecasting with lead times enables the system
operator to effectively plan spinning reserve distribution from a where, = bias
few minutes to several days. Furthermore, load forecasting can ¯ to node
Xi = input
Xo = output of node IV. R ESULTS AND D ISCUSSIONS
w = weights The trained network had a good fitting and is now tested
n = the number of inputs from the incoming layer with the testing data that was synthetically generated. A plot
i = a counter from 0 to n is generated to show the comparison of forecasted load and
The figure above shows how input data is calculated as it the target load. In Fig 2 shows the regression fitting of the
enters a node j. An activation function is used to map the neural network with a Training Rate of 0.85017, Validation
output of the node to a desired lower and upper limits. The Rate of 0.99729, Test Rate of 0.99782 and an Overall Rate of
activation function shown in the figure is the sigmoid function, 0.89988. c
represented as: In Fig.3 shows the result of the forecasted load vs the target
1 load. It is clear that the forecasted is much closer to the target
ϕ= (2)
1 + e−x load.
B. Supervised Network Training
The process of network training used in this paper is Fig. 2. Shows the Forecasted vs the Target outcome.
known as the supervised training. It uses predetermined input
and output values to repeatitively recalibrate the network
weights. The recalibration continues until the error between the V. C ONCLUSION
predetermined and the network ouputs becomes insignificant. The simulation proved that ANN is a good tool for load
To train the network, we use the generalized learning rule, forecasting which explains its popularity as a forecasting
shown below as: method. Matlab’s NNtool was also able to create a neural
network given a handful of predetermined input and output
Wi j = Wi j + αδi Xj (3) data values, though it took a number of iterations in training
before achieving a good regression fitting.
R EFERENCES
δi = ϕ0 vi ei (4)
[1] D.C. Park, M.A. El-Sharkawi, R.J. Marks 11, L.E. Atlas and M.J.
Damborg, Electric Load Forecasting Using An Artificial Neural Net-
where, work, IEEE Transactions on Power Systems, Vo1.6, No. 2, May 1991,
Wi j = is the weight between nodes i and j p. 442
α = is the learning rate, determines magnitude of weight [2] S. Singh, S.Hussain, and M.A. Bazaz, Short Term Load Forecasting
Using Artificial Neural Network, 2017 Fourth International Conference
change every iteration on Image Information Processing (ICIIP), 2017, pp 159-163.
Xj = ouput of node j [3]
ϕ0 = derivative of the activation function
vi = the weighted sum of the output node i
ei = error of the ouput of node i

The Wi j on the left hand side is the updated weight of


Wi j. The second term of the right hand side αδi Xj is used
to recalibrate Wi j.

Fig. 1.

III. M ETHODOLOGY
Synthetic data was used for ANN training. The data consists
of 5 input variables and a target output variable.

A. Modelling
The neural network training toolbox, found on matlab,
has been used to create the neural network consisting of 2
layers and 10 nodes, training and learning algorithm has been
selected. The neural network has been retrained multiple times
to achieved a favorable fitting between training, validation and
testing for higher accuracy. From the toolbox, code can be
generated. To forecast, the the synthetic input data was used
in the neural network to get the expected output.

S-ar putea să vă placă și