Improved multi-kernel LS-SVR for time series online prediction with incremental learning

Abstract

Since it is difficult to establish precise physical model of complex systems, time series prediction is often used to predict their health trend and running state. Aiming at online prediction, we proposed a new scheme to fix the problems of time series online prediction, which is based on LS-SVR model and incremental learning algorithm. The scheme includes two aspects. Firstly, by replacing single kernel with new fixed kernel consisting of several basis kernels, a better information mapping in high dimension is obtained; secondly, by establishing new LS-SVR model without bias term b, the calculation process with incremental learning is simplified. Prediction experiment is performed via certain avionics application. The results indicate preliminarily that the proposed scheme is an effective prediction approach for its good prediction precision and less computing time. The method will be useful in actual application.