Wavelet Transform and PSO Support Vector Machine Based Approach for Time Series Forecasting

Abstract

To accurately predict the non-stationary time series, an approach based on integration of wavelet transform, PSO (Particle Swarm Optimization) and SVM (Support Vector Machine) is proposed. Wavelet decomposition is used to reduce the complexity of time series. Different components are predicted by their corresponding SVM forecasters, respectively, after wavelet transform. The final forecasting result is obtained by combining all predicted results. Taking prediction residual as the fitness value, the parameters of SVM are optimized by a PSO based process. The proposed approach is applied into a coal working face gas concentration forecasting. The results show that simply implanted ANN or SVM based prediction method is not effective when sudden change occurs. The prediction method based on wavelet transform and SVM has better tracking ability and dynamic behavior for suddenly changed data. The performance of the forecaster is remarkably improved to obtain the averaged biases within 3% using the best parameter determined by PSO, which indicates that the suggested approach is feasible and effective.