Forecasting long-term sequences based on a seasonal and periodic-trend feature disentangled network

In this paper, we take into account different repetitive patterns of the seasonal, trendy, and complex periodic characteristics of time series data, and propose a long-term sequence forecasting model, which disentangles seasonal and periodic-trend feature representation. The proposed model consists of two stages. First stage concludes a decomposition component, which divides the temporal data into seasonal and periodic-trend terms. Second stage is a dual- branch prediction model. In one branch, we employ the Squeeze-and-Excitation (SE) attention mechanism to a 2D visual backbone block, which forms a two-level cascaded structure, to predict the multi-period attributes of seasonal term, and in the other, we use autoregression to forecast the periodic-trend term. Experiments on four benchmark datasets have shown that this proposed disentangled prediction network has more advanced prediction performance compared to the current advanced method TimesNet, , resulting in an average MSE reduction of 8.1% for long-term series prediction.