D³CNet: Long-term time series prediction based on dual decomposition and dual-channel hybrid network

Recently, time series forecasting has been widely used in many fields. Time series feature extraction is a key research area, and the complexity and diversity of data make traditional feature decomposition methods difficult to adapt. Aiming at the problem that it is difficult to decompose the mixture of different scale information, and the single-channel model cannot capture the long-term correlation and local features in the multivariate time series at the same time, this paper proposes a long time series forecasting method based on dual decomposition and dual-channel hybrid network (D³CNet).The method utilizes S-EMD to decompose time series data at multiple scales, with each channel in the hybrid network used to process different aspects of the data. This method can capture both global and local information to make more accurate predictions. First, the EMD is utilized to decompose the time series data into components of different frequencies. Then, the data is seasonally decomposed to obtain smoother time series information and reconstruct the seasonal components. Finally, the data are fed into a D³CNet model to predict time series target values. In this paper, multivariate and univariate experiments are conducted on six commonly used datasets. The average MSE and MAE of D³CNet proposed in this paper are significantly improved on the datasets. In the ETT dataset, D³CNet reduces the average MSE by 29.4 % and MAE by 15.7 %. This further proves that D³CNet can reduce the error and improve the accuracy, thus also proving the effectiveness and superiority of the method.