Multi-load forecasting for regional integrated energy systems using a hybrid ResNet-GRU-MTL model with dynamic grey wolf optimization

Regional integrated energy systems (RIES) leveraging multi-energy complementarity require accurate load forecasting for efficient operation, especially with increasing renewable energy integration. Current approaches face challenges including unsystematic preprocessing, inadequate coupling modeling, and inefficient hyperparameter optimization. Therefore, a novel multi-load forecasting framework is proposed to improve electrical, heating, and cooling load forecasting. Preprocessing combines local outlier factor, improved K-means++ clustering, and wavelet threshold denoising to address missing values, outliers, and noise. Temporal patterns across summer, winter, and transitional seasons are analyzed, with Spearman correlation identifying critical inputs. A novel model combining Residual Neural Networks (ResNet), Gated Recurrent Units (GRU), and Multi-Task Learning (MTL) achieves high forecasting accuracy, with mean absolute percentage error values for electrical, heating, and cooling loads as low as 1.67 %, 2.43 %, 0.79 % respectively, demonstrating robust performance across varying data conditions and geographic datasets. Validation across varying data durations, limited inputs, and geographic datasets confirms robust performance. An improved dynamic grey wolf optimization (DGWO) algorithm with nonlinear convergence and adaptive weights, optimizes hyperparameters, outperforming particle swarm optimization and grey wolf optimization by reducing mean absolute percentage error 8.02–20.62 % for electrical load, 14.06–15.57 % for heating load, and 11.03–14.42 % for cooling load, affirming enhanced accuracy and generalizability for engineering applications.