Interpretable probabilistic forecasting of building multi-energy loads using residual learning and global attention-enabled transformer: A case study for residential buildings in South Korea

• Interpretable probabilistic multi-energy load forecasting for building-level IES. • A global attention-enabled Transformer models inter-load and temporal dependencies. • Gaussian-based residual learning improves error calibration and prediction intervals. • Deep SHAP quantifies feature contributions and cross-validate the PMELF results. Accurate forecasting of building-level multi-energy loads is essential for artificial intelligence (AI)-supported design and resilient operation in residential u-IESs, enabling risk-aware planning under uncertainty. However, despite recent advances, existing multi-energy load forecasting (MELF) models still face two key challenges. First, existing models, which are primarily designed for regional-level integrated energy systems such as university campuses with relatively stable operating patterns, lack explicit mechanisms to capture complex coupling relationships among multi-energy loads and often overlook localized building-level load behaviors, thereby reducing sensitivity to subtle demand variations and degrading the quality of the resulting prediction intervals. Second, current MELF models offer limited interpretability, making it difficult to validate model behavior and elucidate the impact of multi-energy interactions on forecasting performance. To address these issues, this study proposes a new interpretable probabilistic MELF (IPMELF) framework that integrates global attention mechanisms with a residual learning-enabled multiple-decoder Transformer architecture for building-level u-IESs. First, a global attention mechanism that integrates feature-wise and time-wise attention with a fusion layer is employed within the Transformer network to explicitly capture the interrelationships among multiple loads as well as the temporal dependencies of multi-energy loads. Gaussian Process Regression-based residual learning is then integrated to enhance local sensitivity, calibrate forecasting results, and produce high-quality prediction intervals by modeling the discrepancy between deterministic predictions and the ground truth with a composite kernel. Finally, Deep Shapley Additive Explanations are applied as a cross-validator, offering clear rationale for PMELF predictions. Case studies conducted on real-world residential buildings in South Korea demonstrate the superior performance of the proposed IPMELF framework, achieving the highest average forecasting accuracy with an R 2 of 95.84%, the lowest pinball loss of 3.34, and improved coverage with narrower prediction intervals, while enhancing the interpretability and trustworthiness of the PMELF results.