Time Series Prediction Study Of Echo State Network Optimized By Evolutionary Algorithm

Time series prediction plays an important role in human society.Terefore,it is important to researchers how to improve the accuracy of time series prediction.In recent years,echo state network(ESN)has attracted much attention because of its excellent performance on time series prediction.However,there are some urgent problems that need to be solved when the ESN is used for time series prediction.First,the reservoir parameter settings of ESN have a large impact on its performance.However,these parameters are usually determined empirically by researchers,which are not only time-consuming but also cannot be adjusted according to different time series adaptively.Secondly,the input weights of ESN are randomly generated,which leads to its better prediction performance on some time series,while it performs poorly on some specific time series.To address these problems,this paper will optimize the ESN in two aspects,the reserve reservoir parameter setting and the input weight setting.The main work of this paper is reflected in the following two aspects.(1)For the problem that ESN cannot select effectively appropriate reservoir parameters according to different time series,which leads to insufficient prediction performance of ESN,an adaptive elite guided artificial bee colony(AEABC)algorithm is proposed to optimize ESN.In AEABC,an adaptive elite-guided solution search strategy is proposed.This strategy introduces the elite solutions that select adaptively from the elite group to guide the generation of new solutions.Moreover,to balance exploration and exploitation better,an evolutionary state-based operation is incorporated into the strategy,using a sine function to control the elite group size according to the evaluation number,so that the size of the elite group is adjusted in time with the evolutionary state.In traditional ABC,the scout bee discards the original solution directly,and then generates a new solution by random initialization,which tends to cause the scout bee to lose its search experience.For this problem,AEABC proposes a scheme to get a new solution by linearly combining the discarded solution and the global best solution.It enables the scout bee to keep the beneficial information both of the discarded solution and the global best solution,thus improving the search efficiency of AEABC.Test the performance of AEABC,ABC,and 13 Improved ABC on the CEC2013 benchmarks.The experimental results show that AEABC has better performance.Therefore,using AEABC to optimize the parameters of ESN reservoir can improve the prediction performance of ESN.As a demonstration of the validity of the AEABC-ESN model,trials are executed on the Mackey-Glass time series and the Chinese society-wide monthly electricity consumption dataset.Experimental feedback indicates that the AEABC-ESN model owns excellent predictive power and generalization capacity.(2)For the problem that the ESN performs poorly on some time series due to randomly generated input weights,an adaptive differential evolution algorithm based on dual experience combination(DECDE)is proposed to optimize the ESN.In the proposed DECDE,a parameter adaptive mechanism based on the combination of individual experience and collective experience is presentedc.In this mechanism,each individual has its own scaling factor(F)and crossover control parameter(CR).In order to improve the performance of the algorithm,the individual adaptively updates the parameter values using its own experience and the collective experience of several successful individuals,which not only makes good use of the individual’s own evolution information,but also combines the useful information of the collective.In addition,a new mutation strategy with external archiving is proposed.In this mutation strategy,a parameter to adjust the greediness of the variation strategy is designed,and it changes dynamically with the increase of the number of function evaluations in the process of evolution,adaptively adjusting the greediness of the mutation strategy at different evolutionary stages,which better balances the exploration and exploitation of the algorithm,thus improves the performance of the algorithm.The algorithm is experimented on the CEC2017 benchmark function,DECDE algorithm is compared with several improved DE algorithms.The experimental results show that the DECDE algorithm can achieve better solutions and faster convergence speed.Therefore,optimizing the input weights of the ESN using DECDE can improve the performance of the ESN.To verify the effectiveness of the DECDE-ESN model,simulation experiments are conducted using different time series.From the experimental results,it can be seen that the proposed DECDE-ESN model has a high prediction accuracy.