PSO optimized radial basis function neural network based electric load forecasting model

Accurate and robust load forecasting models play an important role in power system planning. Due to smaller size and inherent property of good classification, Radial Basis Function Neural Network (RBFNN) is always preferred over other neural network structures. It is used by researchers as an effect...

Full description

Saved in:
Bibliographic Details
Published in2014 Australasian Universities Power Engineering Conference (AUPEC) pp. 1 - 6
Main Authors Kumar Singh, Navneet, Kumar Singh, Asheesh, Kumar, Pradeep
Format Conference Proceeding
LanguageEnglish
Published ACPE 01.09.2014
Subjects
Artificial neural network
Artificial neural networks
Forecasting
Load forecasting
Load modeling
Neurons
particle swarm optimization
Predictive models
radial basis function neural network
short-term load forecasting
Online AccessGet full text

Cover

More Information
Summary:Accurate and robust load forecasting models play an important role in power system planning. Due to smaller size and inherent property of good classification, Radial Basis Function Neural Network (RBFNN) is always preferred over other neural network structures. It is used by researchers as an effective tool for Short-Term Load Forecasting (STLF). The smaller size of this network may lead its output to be a local solution. To train RBFNN, fixing centre widths of hidden layer activation functions and the output layer weights are important. To solve this problem of trapping in local optima, a hybrid forecasting model, i.e., Particle Swarm Optimization (PSO) based RBFNN (PRBFNN) is proposed in this paper. In the proposed model centre widths and output layer weights are optimized by PSO. Therefore, the proposed model keeps the advantages of PSO, as well as RBFNN. The proposed model is tested on the hourly load data for New South Wales, Australia. The results obtained show that the accuracy of the proposed model, in terms of Mean of Mean Absolute Percentage Error (MMAPE) is better than existing artificial neural network based approaches, i.e., Feed Forward Neural Network, RBFNN and Elman Neural Network. The forecasting performance of proposed model, and classical models, i.e., Auto-regressive (AR) and Moving Average (MA), presented in a past research work, is also compared. Again, the performance of proposed model is found better.
DOI:10.1109/AUPEC.2014.6966631