An Interpretable Framework for Mid-Term Electricity Demand Forecasting Using Domain-Driven Spectral Decomposition

Abstract

As global energy systems transition toward carbon neutrality, accurate electricity demand forecasting has become essential for power planning and development. Existing forecasting methods primarily focus on temporal patterns or exogenous factors but often overlook interpretability, which is crucial for mid-term prediction decision support. This paper proposes the Electricity demand Prediction using dRIvers Spectral Method (E-PRISM), an interpretable mid-term electricity demand forecasting framework employing a ‘decompose-predict-combine’ architecture. E-PRISM decomposes complex demand time-series into five domain-specific interpretable components: trend, seasonal, stochastic, trading day effect, and holiday effect. To address key decomposition challenges, we develop two innovative methods: a temperature effect decoupling technique for holiday effect modeling and a ‘Temperature-matching and Proportional Scaling’ method to eliminate extreme temperature interference on trend identification. Then, Component-adaptive prediction strategies are designed based on each component's temporal characteristics, optimizing the trade-off between model complexity and prediction accuracy. Comprehensive validation using real-world data from 27 provincial subsidiaries of State Grid Corporation of China (2021–2024) demonstrates that E-PRISM consistently outperforms existing methods, reducing the mean absolute percentage error from 3.123% to 2.369% compared to state-of-the-art baselines, representing a 24% improvement in prediction accuracy. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.