Time-frequency-Based Hybrid Channel Segmentation Network for Long-Term Time Series Forecasting

Many effective methods have been proposed for time series forecasting tasks. In order to obtain inter-channel correlation and long-term dependencies between time series data, we propose a time-frequency-based hybrid channel segmentation network (HCSNet) for long-term time series forecasting. We first structure a hybrid feature channel fusion structure to obtain inter-channel correlation by cross-channel summation and residual fusion of seasonal and trend features. We then propose a sequence segmentation mixed fusion structure to obtain long-term dependencies by processing segmented time series. Finally, we obtain the prediction results through fully connected and fused operations. Our proposed HCSNet shows surprising predictive performance when compared with other classic methods on GeForce RTX 4060Ti, according to the results of experiments on four commonly used time series datasets.