DFNet: Decomposition fusion model for long sequence time-series forecasting

The study of time series forecasting is of great significance, particularly as accurate forecasting under long time-series is critical to data-driven decision-making. Although existing deep learning models (such as recurrent neural networks, long short-term memory, time convolutional networks, and transformers) exhibit excellent forecasting performances, they do not consider the unique properties of time series, particularly the trend and seasonality. To address this issue, this study proposes a DFNet model with the following characteristics. (1) A three-branch decomposition method is designed to obtain the trend, seasonal, and irregular components, and the least squares method is adopted to correct the seasonal data and process different serial patterns separately to accurately obtain the internal regularity and periodicity, thereby improve model performance. (2) A segmented polynomial activation function is adopted to effectively handle problems associated with negative data loss and slow operation speed, and reduce the amount of model computation. (3) Based on the information fusion transfer mechanism, the extracted long-term dependencies are passed to each other and aggregated to enhance the coupling between sequences. Experiments on nine datasets demonstrated the effectiveness of the proposed model for long time-series prediction. The model showed a significantly superior prediction performance compared with existing prediction models.& COPY; 2023 Elsevier B.V. All rights reserved.

Automated summary

The DFNet model proposed in this study accurately obtains the trend, seasonal, and irregular components of time series, improves model performance, handles negative data loss and reduces model computation, and enhances the coupling between sequences. Experimental results demonstrate the significant superiority of the proposed model for long time-series prediction.