A Novel Two-Stage Forecasting Framework With Difference-Based Learning for Power Load Forecasting

Abstract

Accurate short-term load forecasting is critical for the safe and efficient operation of power systems, in the context of increasing demand fluctuations and the growing penetration of renewable energy sources. However, the strong non-stationarity of load sequences poses significant challenges to existing forecasting models, making it difficult to effectively capture complex temporal dynamics and rapidly changing load patterns. To address these issues, a Two-Stage Decomposition Network (TSD-NET) is proposed as a forecasting framework, which tackles the above challenges through staged prediction, progressive refinement, and structured output fusion. The two-stage architecture first captures coarse-grained temporal trends to establish a stable forecasting baseline, followed by targeted adjustments for short-term fluctuations, thereby alleviating the adverse effects of non-stationarity. To further enhance the model’s capacity to learn fine-grained load features, the network incorporates a series of refinement blocks that iteratively extract previously under-learned patterns. A structured fusion module is employed to integrate outputs from all stages, improving predictive accuracy. Extensive experiments on real-world datasets from Australia, Shanghai, and Tetouan show that our method reduces MSE by up to 46.2% and MAPE by up to 23.7%, and improves $R^{2}$ by up to 4.5 percentage points compared with the best baseline, offering a robust and generalizable solution for dynamic load prediction in complex power-system environments.