An Interval Prediction Model based on Improved Double Echo State Network and Bootstrap

According to the characteristics of non-linearity, high noise and strong coupling of time series in industrial process, an interval prediction method based on improved double reserve echo state network (DRESN) and bootstrap is proposed. Firstly, as traditional echo state network (ESN) cannot meet the requirement of the prediction accuracy, a new reservoir is added to construct DRESN, which the input variables can map to the different feature space and make the prediction performance parameters obtain the abundant decision information, so as to reduce the influence of other effects on models. Secondly, in order to find more suitable parameters of the reservoir, an improved particle swarm optimization (IPSO) is proposed to optimize the parameters in the double-reservoir. Thirdly, the prediction interval is constructed by introducing bootstrap and evaluated by comprehensive evaluation function. Based on the above, an interval prediction model DRESN-IPSO-Bootstrap is formed. Finally, the proposed interval prediction model is simulated by the standard function data set and purified terephthalic acid (PTA) solvent system. The results of simulation experiments show that the proposed method can obtain better interval estimation quality and prediction accuracy, which will provide guidance for the prediction and trend analysis of the key variables in the actual production process.