A Three-Stage Hybrid Learning Framework for Sustainable Multi-Energy Load Forecasting in Park-Level Integrated Energy Systems

Accurate multi-energy load forecasting is essential for the low-carbon, efficient, and resilient operation of park-level Integrated Energy Systems (PIESs), where cooling, heating, and electricity networks interact closely and increasingly incorporate renewable energy resources. However, forecasting in such systems remains challenging due to complex cross-energy coupling, high-dimensional feature interactions, and pronounced nonlinearities under diverse meteorological and operational conditions. To address these challenges, this study develops a novel three-stage hybrid forecasting framework that integrates Recursive Feature Elimination with Cross-Validation (RFECV), a Multi-Task Long Short-Term Memory network (MTL-LSTM), and Random Forest (RF). In the first stage, RFECV performs adaptive and interpretable feature selection, ensuring robust model inputs and capturing meteorological drivers relevant to renewable energy dynamics. The second stage employs MTL-LSTM to jointly learn shared temporal dependencies and intrinsic coupling relationships among multiple energy loads. The final RF-based residual correction enhances local accuracy by capturing nonlinear residual patterns overlooked by deep learning. A real-world case study from an East China PIES verifies the superior predictive performance of the proposed framework, achieving mean absolute percentage errors of 4.65%, 2.79%, and 3.01% for cooling, heating, and electricity loads, respectively—substantially outperforming benchmark models. These results demonstrate that the proposed method offers a reliable, interpretable, and data-driven solution to support refined scheduling, renewable energy integration, and sustainable operational planning in modern multi-energy systems.