Electricity Load Forecasting for 400 Public Buildings in Shenzhen: Model Comparison, Feature Selection, and TreeSHAP Interpretation

One-hour-ahead electricity-load forecasting offers a potential reference for near-term demand monitoring, while its practical implementation requires a balance between predictive performance and input-data requirements. This study develops one-hour-ahead forecasting models using 3,504,000 hourly observations from 400 public buildings in Shenzhen, China, over the full calendar year of 2021. One pooled model was fitted for each algorithm after chronological splitting within every building: the earliest 80% of records formed the training period and the latest 20% formed an isolated test period. Thirty-five predictors described building characteristics, target-hour calendar information, the most recent completed weather observations, socio-spatial context, and previous-day load profiles. Five-fold expanding-window cross-validation within the training period was used for preprocessing decisions, Bayesian hyperparameter optimization, model selection, feature ranking, and recursive feature elimination (RFE). After the candidate configurations were locked, they were refitted on the complete training period and evaluated once on the held-out test observations. CatBoost achieved the best 35-feature test performance (R2 = 0.945). Training-period RFE selected a 12-feature configuration, which achieved R2 = 0.948 on the held-out test set. A four-feature configuration comprising building area, target hour, previous-day peak load, and previous-day valley load achieved R2 = 0.901. TreeSHAP identified building scale, intraday schedule, and recent load-profile descriptors as the dominant predictive signals. These findings concern later-period forecasting for the same 400 Shenzhen buildings within 2021. Transfer to unseen buildings, other cities, or other years was not evaluated; SHAP attributions represent predictive associations rather than causal effects.