DFF-Net: Dynamic feature fusion network for time series prediction

Time series forecasting predicts future values based on historical observations within a sequential dataset. However, popular attention mechanisms exhibit high computational complexity when it comes to capturing channel correlations. Additionally, multi-scale feature fusion methods often generate information redundancy when processing diverse features, which can result in unstable model learning. In this paper, we propose a Dynamic Feature Fusion Network (DFF-Net) to address the aforementioned challenges. The network consists of two key modules: the Stochastic Feature Aggregator (SFA) and the Dimensional Mixer (DMix). First, the SFA module extracts core feature representations by utilizing random training sampling and weighted averaging during the inference process. These core features are then integrated with individual feature representations. Second, the DMix module utilizes scalable dimensional transformations to achieve feature compression and reconstruction. The compressed features and the reconstructed features are then concatenated to enhance data representations. Experimental results demonstrate that DFF-Net outperforms seven state-of-the-art methods in both prediction accuracy and computational efficiency across multiple benchmark time series datasets.