Research Of Time Series Prediction Model Based On CEEMDAN And Deep Learning

Time series prediction is applied to many aspects of the world,accurate prediction of future data in time series requires capturing representative series features.However,traditional time series prediction models have larger prediction error for non-stationary time series,machine learning related time series prediction models have lagging phenomenon in prediction.Aiming at above problems,propose a time series forecasting model combining CEEMDAN and deep learning.First,the time series is decomposed into eigenvalues series representing different time scales by CEEMDAN,and then use the Yo Y inspection strategy to determine the best sliding window parameters as the feature dimension for each decomposition result,use the multi-head attention and bidirectional LSTM to predict high frequency sequences,use the GRU to predict low frequency sequences,finally regard the prediction result of each scale sequence as the current moment feature of the time series,using the fully connected layer to perform non-linear fitting with the true value,and get the final prediction result.By comparing with 2kinds of traditional time series prediction models: single exponential smoothing and second exponential smoothing,the model in this paper can reduce the mean square error of prediction by1.88%,1.20%,0.23% and 3.78% on test set.By comparing with 4 kinds of neural network time series prediction models: BP neural network,RNN,LSTM and GRU on 4 groups of time series data,the model in this paper can reduce the mean square error of prediction by 0.75%,0.13%,0.14% and 1.29% on test set.It is proved that the model proposed in this paper can effectively reduce the prediction error and solve the prediction lagging phenomenon.Through ablation experiment,it is proved that bidirectional LSTM and multi-head attention can effectively improve the fitting ability of inflection point of the model in this paper.