Short-Term Power Load Forecasting Based on Interpretable Time-Frequency Dual-Domain Fusion

Real-world power load series exhibit pronounced non-stationarity and complex spatiotemporal dependencies, while the black-box nature of deep learning limits its credibility in industrial dispatching. This paper proposes a probabilistic short-term load forecasting model integrating multi-scale temporal modeling and learnable frequency-domain filtering to improve forecasting accuracy, uncertainty quantification, and interpretability. The model employs parallel temporal and frequency branches. The temporal branch uses multi-scale patches to capture local fluctuations and long-range dependencies, while the frequency branch applies a learnable complex-valued filter to extract dominant global seasonal periodicities. A dynamic gating mechanism adaptively fuses features from both domains, and quantile regression generates prediction intervals. On the Quanzhou and Australian datasets, the proposed model achieves mean absolute percentage errors of 0.49% and 0.77%, respectively. The corresponding prediction interval coverage probabilities are 79.18% and 80.65%, both close to the nominal coverage of 80%. A SHapley Additive exPlanations (SHAP)-based post hoc attribution framework further identifies key factors influencing predictions through feature contributions. The results suggest that the proposed model effectively combines local temporal fluctuations with global frequency-domain periodic information and provides a degree of post hoc interpretability, supporting accurate probabilistic load forecasting and a clearer understanding of the model’s predictions.