Time series exists widely in nature,such as meteorology,physics and so on.Because of the nonlinearity and non-stationarity of actual time series,there are still great problems in the accurate prediction of time series.But accurate forecasting time series is very important for normal production and life.Usually,for some complex systems,because the internal information of the system contains the characteristics of chaos,the traditional time series prediction model can not fully excavate the characteristics of the system,which leads to the accuracy of the forecasting model is often unsatisfactory.In order to improve the prediction effect of chaotic time series,this paper proposes a combined prediction model based on empirical mode decomposition(EMD)and neural network(NN).The main research contents include:Firstly,learned about neural networks and time series,reading literature and summarizing relevant methods.Based on statistical methods and artificial intelligence methods,the proposed prediction models,including Kalman filter,time series model and artificial neural network,the advantages and disadvantages of each model are analyzed and compared.Secondly,aimed at the limitation of single model,a chaotic time series prediction model based on neural network and its combination model is proposed.the model uses the EMD to decompose the original sequence.the fixed modal function IMFs and trend term obtained after decomposition are classified into "high,medium and low" frequencies according to the run-time criterion.The feedforward neural network trained by fuzzy first order conversion rule is used to predict the high frequency IMFs;and the ARIMA model is used to predict the intermediate frequency IMFs and trend term.The predicted results are superimposed to obtain the final prediction results.Lastly,the above model is used to Lorenz-63 the time series and Mackey-Glass time series,and the validity of the mixed model prediction is verified by various error criteria.The results show that the EMD-ARIMA-NN-FFOTR model proposed in this paper is superior to the traditional single model in the prediction effect,and the prediction data are very close to the real value. |