Enhanced industrial heat load forecasting in district networks via a multi-scale fusion ensemble deep learning

Abstract

Accurate heating load prediction is vital for optimizing the operation of thermal systems, improving energy utilization efficiency, reducing operational costs, enhancing user satisfaction, and promoting the use of renewable energy. To facilitate short-term prediction of heat consumption in industrial areas for practical applications, a multi-scale fusion ensemble model is proposed to address the issue of pressure balance in heating networks. Specifically, (1) Hierarchical Decomposition Approach: To overcome the limitation of relying solely on historical heat load data, a hierarchical decomposition mode is designed by combining Naïve Decomposition, Empirical Mode Decomposition, and Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise. This approach deeply explores the nonlinear characteristics of the heat load. (2) Integrated Heat Load Prediction Framework: An integrated prediction framework based on neural networks—including Back Propagation Networks, Recurrent Neural Networks, Long Short-Term Memory Networks, and Gated Recurrent Unit Networks is constructed. For each component, the optimal prediction model is adaptively selected, and the predicted results are fused using weighted averages. The proposed scheme was applied to 24-hour ahead heating load prediction for four regions of a thermal power company in Quzhou City, Zhejiang Province. The coefficients of determination R² achieved for the four regions were 0.8646, 0.8707, 0.8509, and 0.9422, respectively, with Mean Absolute Percentage Errors reaching 10.18%, 3.93%, 2.78%, and 2.31%. Compared with seven classical prediction models, as well as Transformer and its variants, the proposed model outperforms them across five performance indicators and demonstrates strong generalization ability.