Electricity Load Forecasting via Logarithmic Feature Fusion and Energy-Aware Multi-scale Aggregation

Accurate electric load forecasting is essential for smart grid dispatch and operation. The primary challenge lies in the scale disparities among features extracted from different domains and temporal resolutions, which result in training instability and fusion difficulties. A novel forecasting model named TFMixer is proposed to address these challenges. Two key mechanisms are incorporated: (1) the Log-MixGate mechanism is adopted to project time-domain and frequency-domain features into logarithmic space for stable geometric fusion, and (2) the Energy-aware Spectral Aggregation (EnSA) strategy is employed to dynamically balance multi-scale predictions based on energy consistency. Experiments on benchmark datasets show that TFMixer outperforms state-of-the-art methods while maintaining computational efficiency, providing a practical solution for real-time grid dispatch.