Data fusion combined with echo state network for multivariate time series prediction in complex electromechanical system

Multivariate time series is widely used to evaluate and predict the health state of system. To focus on the importance of prediction accuracy for multivariate time series, many a research have been done on improving the performance of predictive model. However, for multivariate time series, the prediction accuracy depends not only on predictive model but also on the input information. In this paper, we combine data fusion with ESN to improve the prediction accuracy of multivariate time series, which focuses on both the processing of the input information of predictive model and the optimizing of predictive model. First, multi-sources data fusion is presented to obtain the new input for prediction model before predicting; phase space reconstruction and self-adaptive weighted fusion algorithm are adopted to fuse multivariate time series and obtain more complete information. Then, leaky rectifier liner units are used to replace the original activation function, tanh, and locality preserving projection is employed to optimize the state matrix of the reservoir. Finally, the effectiveness of the proposed method is verified by an analysis of one case study of real compressor groups data sets in chemical production system. The results and a comparison with the traditional method show that the proposed method can greatly enhance the prediction accuracy of multivariate time series and the one-step ahead prediction accuracy is improved by three orders of magnitude as well as a better generalization ability is obtained in the multi-steps ahead prediction.