IBRNN-EDFM: An Improved Bayesian Regularized Neural Network based Electricity Demand Forecasting Model

Abstract

Electricity demand forecasting is one of the main challenging work in the field of data mining. Bayesian regularized neural networks are more strong than the standard back propagation algorithms. Bayes's theorem is the statistical model that is used to compute the statistical values in a powerful manner. It requires high computational cost, specifically with a large number of parameters. Moreover, the simulations provide somewhat dissimilar answers unless the same random seed is used. In order to overcome these issues, an Improved Bayesian Regularized Neural Network (IBRNN) is proposed as a novel method for electricity demand forecasting. Here, the Department of Economics and Statistics and Tamil Nadu Electricity Board (TNEB) data information are taken for forecasting the demand. It is used for both testing and training the proposed Improved Bayesian Regularized Neural Network based Electricity Demand Forecasting Model (IBRNN-EDFM) system. Initially, the data information are preprocessed to eliminate the error and unrelated attributes in the dataset. Then, the mean, deviation and initial objective functions (α and β) are calculated. Here, the conditional entropy and Jacobian matrix are also computed for improving the regularization process. The objective function is recomputed based on the mean and deviation values. This process will be repeated until convergence. Finally, the demand is forecasted with the help of Bayesian regularization. The experimental results evaluate the performance of the proposed system in terms of precision, recall, f-measure and accuracy. The proposed IBRNN-EDFM provides the best results, when compared to the other algorithms such as, Naivy bayes, random forest and decision rule.