Neural Network Forecast Algorithm Based on Iterated Unscented Kalman Filter

Abstract

Abstrac t—A novel algorithm based on the iterated unscented Kalman filter (IUKF) is proposed in this paper to train the weights and bias of the neural network. In the proposed algorithm, the weights and bias are considered as the states, and the outputs of the network are used as the measurements for the IUKF. In IUKF, the iteration concept is introduced into the unscented Kalman filter (UKF). By substituting the updated mean and covariance into the unscented transformation (UT), the total forecast precision is improved. Taking the Mackey-Glass chaos time sequences as an input of the net, the neural network is simulated with the IUKF, UKF and back propagation (BP) algorithms. The simulation results indicate that the IUKF algorithm has a faster training speed and higher forecast precision than the BP algorithm. Moreover, the IUKF algorithm avoids the network’s convergence getting into the local minimum points. Compared with UKF algorithm, the proposed algorithm has a higher forecast precision.