A Shapley value-based dynamic ensemble framework for short-term load forecasting of industrial consumers

• A cooperative game-theoretic ensemble learning framework is proposed using dynamic Shapley value-based model weighting. • An autonomous framework with online calibration is developed for real-time Shapley contribution adjustment. • Case studies demonstrate that the proposed framework can coordinate complementary advantages of heterogeneous models. The large-scale integration of distributed energy resources requires accurate short-term load forecasting for modern power systems to maintain supply–demand balance and operational efficiency. This study focuses on the demand-side forecasting of industrial consumers, which requires precise forecasting to optimize demand response capabilities. To address the limitations of static prediction architectures in capturing multi-scale dynamic features and high-dimensional coupling characteristics of industrial loads, we propose a Shapley value-based dynamic ensemble learning framework that strategically integrates statistical models, machine learning models and deep learning models. By introducing Shapley values in the cooperative game theory to quantify individual model contributions and dynamically adjust the ensemble weights, the method achieves robust adaptation to load variations while maintaining high computational efficiency. The follow-up case study on high-resolution industrial load data demonstrates its superior performance over conventional static prediction architectures and static weighting schemes across multiple scenarios.