Wind power grouping forecasts and its uncertainty analysis using optimized relevance vector machine

Relevance vector machine, a sparse probabilistic learning machine based on the kernel function, has excellent ability of prediction and generalization. It is proposed by this paper that the optimized relevance vector machine (ORVM) is a wind power interval forecasting model which is able to provide a certain prediction value and its possible fluctuation range at a given confidence level. The proposed model characterizes in insufficient sample training and uncertainty analysis and is greatly suitable to most of wind farms in China (newly built or large scale wind farms). First, a grouping mechanism has been used to divide wind turbines into several groups to establish forecasting model separately. Second, a selection method properly taking the characteristics of NWP error distribution into consideration was presented to improve forecasting accuracy of each group. Third, the parameters of the kernel function and initial value of iteration are determined by particle swarm optimization to further enhance forecasting accuracy. Two wind farms in China are involved in the process of primary data collection. The performance data obtained from ORVM models are tested against the predicted data generated by GA-ANN and SVM. Results show that the proposed model has better prediction accuracy, wider application scope and more efficient calculation.