Prediction model of a hybrid recurrent neural network based on sequence decomposition
Jia Zhao, Changxiang Li, Longzhe Han, Yannian Wu, Lieyang Wu · International Journal of Computing Science and Mathematics · 2023
A large number of random time series exhibit obvious nonlinear characteristics. In order to effectively learn the nonlinear characteristics of time series, this paper proposes a hybrid recurrent neural network (RNN) prediction model based on sequence decomposition. Firstly, this model uses the seasonal trend decomposition method based on local weight regression to decompose the original time series into trend series, seasonal series and residual series, and fuses these three sub-sequences with the original feature series to form a new feature series. Secondly, the input sequence is decomposed into three levels through the progressive decomposition network, and different neural networks are used to predict the decomposition sub-sequences of each level. Finally, the prediction results are spliced and put into the full connection layer for the final prediction. Simulation results show that the proposed model has the lowest prediction error and higher prediction accuracy.