Online Chaotic Time Series Prediction Based on Square Root Kalman Filter Extreme Learning Machine

In this paper, we proposed a novel neural network prediction model based on extreme learning machine for online chaotic time series prediction problems. The model is characterized by robustness and generalization. The initial weights are initialized by orthogonal matrix to improve the generalization performance and the output weights are updated by square root Kalman filter. The convergence of the algorithm is proved by Lyapunov stability theorem. Simulations based on artificial and real-life data sets demonstrate the effectiveness of the proposed model.