Application Research Of Interval Prediction Method Based On Double Reservoir Echo State Network And Improved Parallel Bootstrap

With the gradual development of modern industrial production scale,the industrial process is becoming more and more complex,the number of control points is also increasing,and higher requirement is put forward to the prediction of key points.Generally,the traditional point prediction method is used to estimate,but the model based on point prediction method can not meet the demand of complex industrial due to the variability of monitoring variables.Interval prediction is taken as a trend analysis method,it can not only achieve high prediction precision,but also visually view the change trend,which provides a certain reliability analysis for industrial process monitoring.Therefore,aiming at the problem of high noise,non-linear time series and multi factor inputs in industrial process,the following research work has been carried out.First,a point prediction method based on echo state network with double reserve pools is proposed.DRESN network,as a point prediction model,can show good non-linear mapping ability.Based on this model,an improved particle swarm optimization(IPSO)algorithm is proposed to optimize the network parameters in the double reserve pool.It introduces the idea of adaptive adjustment of the optimization parameters,and divides the optimization particles into several sub-populations to complete the search process independently,which avoids the disadvantages of local optimization of traditional PSO algorithm.Thus,the IPSO-DRESN model is formed,which combines the improved PSO algorithm with echo state network of double reserve pool.Second,it is considered that the bootstrap sampling method contains many repeated operations,which affects the training time of the model to a certain extent.Thus,an improved parallel bootstrap algorithm is proposed to optimize it.The sampling data by bootstrap is put on multiple core processors to complete the training and generalization of IPSO-DRESN model in parallel,so as to improve the computer manipulation and speed up the overall training speed.In addition,five evaluation indexes(including prediction interval coverage,average interval width and cumulative deviation,etc)are introduced to evaluate the interval quality.Third,the standard function data set with Gaussian noise and Heteroscedasticity noise is used to verify the interval prediction model proposed in this paper.Also,the data set of actual industrial process Pure Terephthalic Acid(PTA)solvent system is used for simulation verification,The experimental results show that the proposed IPSO-DRESN model combined with improved parallel bootstrap method has high prediction accuracy and interval estimation quality.