Time-Frequency Synergy for Power Load Forecasting: A Principal Frequency-Band Energy-Weighted Integrated Approach

The complexity of power load data greatly benefits from analyses conducted in both the time and frequency domains. Time-domain analysis excels at capturing local fluctuations within the data, while frequency-domain analysis is more effective at modeling long-term trends and global patterns. This advantage is particularly important for power load data, which often exhibits pronounced periodic behaviors. However, relying solely on frequency-domain modeling struggles to fully capture the inherent local fluctuations present in power load data. To leverage the strengths of both time-domain and frequency-domain analyses, we propose a novel time-frequency integrated model that effectively captures both local variations and periodic patterns in load data. Specifically, we propose a principal frequency-band energy-weighted fusion strategy that dynamically adjusts the weights based on the top k most important frequency bands of the input signal. In the frequency-domain module, we model different load variables to capture the periodic patterns in the data. In the time-domain module, we model individual load series to capture local fluctuations. Experimental results across multiple datasets demonstrate that our approach achieves state-of-the-art performance in power load forecasting.