Energy Demand Forecasting in China Based on Dynamic RBF Neural Network

A dynamic radial basis function (RBF) network model is proposed for energy demand forecasting in this paper. Firstly, we present a time series forecasting framework based on variable structure RBF network. In this framework, both the number of basis function and the input orders are variable. Secondly, an on-line prediction algorithm using sequential Monte Carlo (SMC) method is developed. Due to the high dimensional state-spaces, the Rao-Blackwellised particle filter is adopted to compute the posterior probability density function of state variables. In this SMC algorithm, the sub-space sampling, state prediction, weight updating, exact computation with Kalman filter and the change of RBF structure have been discussed in detail. At last, the data of total energy demand in China are analyzed and experimental results indicate that the proposed model and prediction algorithm are effective.