Evaluating the Reliability of Cross-Building Load Forecasting Under Distribution Shift and Sensor Missingness

Short-term building-electricity-load forecasting is important for energy management and smart-building operation. However, existing evaluations mainly focus on prediction accuracy and provide limited assessment of reliability under cross-building shifts and data quality variations. This study develops an empirical evaluation framework based on the HEEW dataset. A five-fold cross-building validation protocol was designed, including 55 main buildings and 27 additional buildings for extended sensitivity analysis. The prediction performance, sensor missingness robustness, prediction interval calibration, and model interpretation stability were evaluated. The results show that the LightGBM residual model outperforms the Lag-1 persistence baseline in cross-building forecasting. It reduces Normalized Mean Absolute Error (NMAE) by 11.3% in the unseen-building future-year scenario. Missingness experiments indicate that historical load features have a stronger influence on one-hour-ahead forecasting than weather features. A 50% missing rate in auxiliary load history increases NMAE by 96.71%, while weather feature missingness causes limited degradation. Conformal prediction maintains reasonable coverage for seen buildings but shows reduced coverage after transferring to unseen buildings. SHAP analysis further shows stable feature importance patterns across different scenarios. Overall, this study shows that the reliability of cross-building load forecasting should be evaluated from multiple angles, beyond point prediction accuracy alone.