Objective The study aims to develop a feasible multi-horizon short-term load forecasting method for office buildings, where load is jointly affected by working-day behavior, holiday and adjusted working-day arrangements, and meteorological conditions, and where forecasting accuracy generally decreases as the forecast horizon increases. It also aims to clarify the role of meteorological factors under different forecast horizons and operating scenarios.

Method Hourly power load data of a typical office building from January 1, 2024 to May 24, 2025 were collected together with concurrent meteorological observations and forecast data. An input feature set was constructed using historical load, calendar information, holidays and adjusted working days, target-time enhanced features, and similar-day statistics. Extreme Gradient Boosting (XGBoost) was used to build forecasting models for 1 h, 2 h, 3 h, 6 h, 12 h and 24 h ahead, and an incremental model was further developed for the 24 h-ahead task. The forecasting performance was then evaluated from both overall and scenario-based perspectives.

Result The models show good forecasting performance across horizons from 1 h to 24 h. As the forecast horizon increases, the overall error increases and the correlation decreases. Historical load features remain the dominant contributors to forecasting accuracy. Target-time calendar features, holiday and adjusted working-day information, and similar-day features improve the performance for horizons of 6 h and above. Meteorological factors provide more evident gains for horizons of 3 h and above and in relatively stable scenarios. For 24 h-ahead forecasting, incremental modeling outperforms direct forecasting.

Conclusion A hierarchical strategy is suitable for short-term load forecasting of office buildings, including direct forecasting for short horizons, target-time enhancement for medium- and long-horizon forecasting, and incremental modeling for cross-day forecasting. Historical load is the core information source for short-horizon forecasting, while meteorological factors serve better as auxiliary correction factors with clear horizon and scenario dependence. Holiday, adjusted working-day, and similar-day features are key supplements for improving medium- and long-horizon forecasting. This framework provides a useful reference for building-side load forecasting and refined demand-side energy management.