Sparse Bayesian learning using combined kernels for medium term load forecasting

A pattern recognition method based on probabilistic forecasting, Sparse Bayesian Learning (SBL) model is applied for regression in medium term load forecasting. And for the kernel functions chosen, the paper utilizes linear combination principle to construct multiple combined kernel functions, the Gaussian kernel with polynomial kernel and tensor product spline kernel are collected. The parameters of these combined kernels are optimized by Particle Swarm Optimization (PSO). With the training and testing sample data from “2001 world-wide competition of electricity load forecasting”, the results show that all combined kernel models exhibit better accuracy than single kernel models. Besides, probabilistic forecasting results are also given based on the exclusive probability property of Sparse Bayesian Learning.