Multivariate Time Series Prediction Based On Echo State Network

In daily life and the practical application of big data,it is usually necessary to record the data in the observation system and actual production.These data recorded in time sequence are the basic time series.Multivariate time series is obtained when multiple variables are observed or multiple data are recorded simultaneously in the system.The purpose of studying time series is to analyze the structural characteristics of time series and excavate its internal information rules,so as to correctly predict the future trend of time series and realize the adjustment through certain human behaviors to make the system run in accordance with the expected trajectory.Compared with unary time series,multivariate time series prediction can provide more information.Therefore,it is particularly important to establish a suitable model applied to multivariate time series.For this reason,this thesis proposes a combinational prediction model to reduce dimensionality of multivariate time series by using kernel principal component analysis(KPCA)and optimize ESN by combining with improved gray Wolf algorithm.The following studies are mainly done in this thesis:There is always correlation and redundancy among many observed variables in multivariate time series.The principal component of data is extracted through KPCA,which not only removes the correlated variables in time series,but also stores most of the original data information in the remaining variables.ESN was selected as the modeling core.In view of the shortcomings of the traditional forward neural network,such as easy to fall into local optimum and slow convergence speed,this thesis chooses ESN as the research core,and the experimental simulation proves that the prediction accuracy of ESN is better than that of the traditional BP neural network.The improved grey Wolf algorithm was used to optimize ESN.Since reservoir parameters have a great influence on ESN accuracy,gray Wolf algorithm is adopted in this thesis to optimize reservoir parameters of ESN.In view of the disadvantages of the grey Wolf algorithm,such as poor population difference and slow convergence speed,the reverse learning strategy and nonlinear convergence factor were used to improve the grey Wolf algorithm,and the combined prediction model was finally established.The use of combination forecast model is set up on the air quality index AQI and stock market closing price of the two typical multivariate time series data for simulation,and compared with other prediction model,through the experimental results show that the proposed combination forecast method on error,evaluation of performance,reliability,and confidence interval is better than that of before improvement and a single forecasting model.