Multi-energy load forecasting for integrated energy systems based on physical-gated and dual-stream decoupling
To address the issues of multi-timescale feature aliasing, insufficient physical interpretability, and difficulty in uncertainty quantification in multi-energy load forecasting for integrated energy systems, a joint point and probabilistic prediction method combining physics-gated dual-stream network with weighted conformal prediction is proposed. A causality-constrained seasonal-trend decomposition is adopted to decompose the load into trend and residual components, which are modeled distinctively with linear projection and a bidirectional gated recurrent unit, respectively. A gating mechanism based on cooling and heating degree days is introduced to adaptively fuse the dual-stream features, achieving physically interpretable multi-timescale representation. A multi-task learning framework is constructed, where the losses of electricity, heating, and cooling loads are automatically balanced through homoscedastic uncertainty weighting, and a Huber-relaxation-based coupling physical constraint across cooling, heating, and electricity is embedded. Furthermore, an exponentially decay-weighted conformal prediction is proposed to provide adaptive intervals with finite-sample coverage guarantee for non-stationary conditions. Experiments on a real-world dataset from the Tempe campus of Arizona State University demonstrate that the proposed model significantly outperforms the compared models in terms of root mean square error and interval score in multi-step prediction, achieving a favorable balance between coverage rate and interval width. © 2026 Elsevier Ltd
Библиографическая ссылка Chen, Qingbin, Yang, Dechang, Tomin N.V. Multi-energy load forecasting for integrated energy systems based on physical-gated and dual-stream decoupling // Applied Energy. Vol.426. No.128764. 2026. DOI: 10.1016/j.apenergy.2026.128764
Скопировать
Проиндексировано: WOS
SCOPUS
