Bayesian Neural networks for short term load forecasting

Abstract

Is this paper, Bayesian approach was used to learn the artificial neural network. In Bayesian ANN, the error function consists of two terms: first term is the error term of entire data, second term is the extra regularizing term(also called weight decay term ) which can penalize large weight. Each weight and the error were considered as random variables, their prior probability distributions are normal with zero mean, and their variances constant called the hyper-parameters. The main work of Bayesian approach is obtain the most probable values of hyper-parameters, such that Margin likelihood get maximum values. We used Bayesian Neural network and ordinary ANN as base models to forecast the hour power load. The forecasting results show that the MAPE and RMSE of the Bayesian ANN are all less than that of other Classical ANN. Bayesian ANN has better performance, it can be applied to real forecasting work.