Chaotic Time Series Prediction Using Data Fusion

One of the main problems in chaotic time series prediction is that the underlying nonlinear dynamics is usually unknown. Using a nonlinear predictor to predict a chaotic time series usually puts a limit on the accuracy since the nonlinear predictor is basically an approximation of the unknown nonlinear mapping. In this paper, we propose using fusion of predictors as a method to improve the performance of chaotic time series prediction. Different nonlinear predictors with distinct characteristics including the multi-layer perceptron neural network, radial basis function (RBF) neural network, fuzzy inference system, recurrent neural network, Volterra filter, and local linear predictor are used to predict a chaotic time series. Their predictions are then combined to produce a more accurate prediction by using the linearly constrained least square (LCLS) fusion method. The proposed prediction fusion method is evaluated using simulated chaotic time series based on the Mackey-Glass equation and Ikeda system. Results show that the fused predictor consistently outperforms all the individual predictors.