Online multi-step-ahead time series prediction based on LSSVR using UKF with sliding-windows

Accurate multi-step-ahead prediction over long future horizons posts great challenges for the application of time series prediction. A novel online multi-step-ahead prediction method based on least squares support vector regression (LSSVR) is proposed in this paper. Taken the superiorities of using sliding-windows to reduce largely computation burden and implementing LSSVR model updating by Unscented Kalman Filter (UKF) into consideration, the proposed method not only can construct online predicted model in much fewer training data (such as the size of original training data set required is only the sum of embedding dimension corresponding to phase-space-reconstruction and the length of sliding-windows), but also has the better accuracy over multi-step-ahead prediction. When the prediction horizon reached the predefined step p in the process of predicting, model parameters consisted of kernel width σ, support values {αk}k=1L and bias term b are updated by new arrived measurements and UKF. Finally, several simulations are provided to show the validity and applicability of the proposed method.