Robust Probabilistic Load Forecasting for Multi-Energy Buildings Using Issue-Aware and Conformally Calibrated Gradient Boosting
High load variability and the low quality of building monitoring data pose substantial operational challenges for modern energy management systems. This study develops a robust and computationally efficient probabilistic forecasting framework by integrating data-issue handling with uncertainty calibration. Using an experimental design on a high-resolution multi-energy dataset (2018–2023), the study compares Gradient Boosting Decision Trees (GBDTs) with a linear baseline under a strict out-of-time validation protocol and Conformalized Quantile Regression (CQR). The results indicate the superiority of non-linear models: CatBoost delivers the best point-forecast accuracy, achieving a Mean Absolute Error (MAE) of 37,752.04 kW, corresponding to an 11–12% performance improvement over ElasticNet. Conformal calibration substantially improves the validity of prediction intervals, increasing the Prediction Interval Coverage Probability (PICP) from 82.22% to 87.28%, thereby approaching the nominal 90% confidence target without imposing strong distributional assumptions. Further ablation analyses reveal that rolling-window features contribute more to accuracy than external weather variables. Overall, these findings provide a practical contribution in the form of a forecasting method that is not only accurate but also statistically reliable in estimating operational risk, thereby bridging the gap between industry demands for robust systems and the constraints imposed by real-world data quality.