Chaotic time series prediction based on multi-kernel learning support vector regression

Abstract

Multi-kernel learning support vector regression (MKL-SVR) are proposed for chaotic time series prediction to solve the problems of kernel selection and hyper-parameter determination when using the standard SVR. The algorithm is realized through quadratic constrained quadratic programming (QCQP) in the hybrid kernel space, which not only reduces the number of support vectors, but also improves the prediction performance. Finally, it is applied to Mackey-Glass, Lorenz and Henon chaotic time series prediction. The results indicate that the proposed method can effectively increase the prediction precision, accelerate the convergency of cascade learning and enhance the generalization of prediction model.