Short-term load forecasting based on Bayesian neural networks learned by Hybrid Monte Carlo method

This paper reports on Bayesian technique to design an optimal neural networks based model for short term load forecasting. Usually, The weight vector has Normal prior distribution, and the posterior distribution is approximated by Gaussian approximation. In order to avoid Gaussian approximation, we used Hybrid Monte Carlo algorithm to learn the weight vector, such that the Hamilton energy function has minimal value. In Hybrid Monte Carlo algorithm, the Hamilton function is the regularized error function, and the position variables is weights. In simulation experiments, two types Bayesian neural networks with Normal weight and Cauchy weight are used to hour load forecasting. Experiment result show that these two approaches have good performance (evaluated by MAPE and RMSE).