DFINet: Dual-Way Feature Interaction Network for Long-Term Series Forecasting

Intelligent long-term series forecasting is of great importance in the field of instrumentation and measurement (IM), as it provides advanced insights, optimizes decision-making processes, and facilitates proactive planning. However, existing methods encounter challenges in capturing dependencies within complex temporal patterns, integrating serial and spatial perspectives, and handling large-scale datasets efficiently. To tackle these challenges, we propose a novel approach specifically designed for IM applications, called the dual-way feature interaction network (DFINet). DFINet leverages the unique properties of time series data in the IM domain and incorporates two key units: the serial decomposition unit and the spatial reshuffling unit. These units enable the construction of a refined series that accurately captures long-term dependencies from both serial and spatial perspectives. Furthermore, we introduce a feature interaction module in DFINet, which extracts dependency information at a subsequence level. This enhancement significantly improves the model's ability to learn complex patterns within the time series data, making it highly effective in IM scenarios. To evaluate the performance of our proposed network, we conducted experiments on eight public benchmark datasets. The results demonstrate that DFINet outperforms state-of-the-art methods in terms of prediction accuracy and efficiency.