Wavelet neural network based on MSUKF and its applications in chaotic time series prediction

Abstract

Wavelet neural network (WNN) trained by unscented Kalman filter (UKF) has many merits of fast convergent rate and small prediction error without computing the Jacobian matrix. Based on this, an improved UKF is introduced into the parameters estimation for WNN. The algorithm uses an unscented transform (UT) based on minimal skew simplex Sigma point sampling strategy in the frame of Kalman filter, which not only inherits all the merits of UKF, but also increases the computational efficiency. The experimental results for chaotic time series prediction show that WNN of the improved UKF has the faster training speed and higher prediction precision than that of EKF, and has a similar precision with that of UKF but high computational efficiency. In addition, it has also a good applicability to the chaotic time series prediction.