Accurate cooling load forecasting is critical for optimizing heating, ventilation, and air conditioning (HVAC) system operations, reducing energy consumption, and advancing building sustainability objectives. This paper introduces the Hybrid Global and Sub-domain Approach (HGSA), a novel forecasting framework that synergistically combines advanced feature engineering, domain-specific data segmentation, and ensemble learning to deliver robust predictions across multiple time horizons. The methodology addresses the inherent complexity of cooling load dynamics by partitioning historical data into complementary domains—global, hourly, day of week, monthly, and temperature-based, each capturing distinct temporal and climatic patterns. HGSA incorporates four feature categories: temporal, weather-related, historical, and periodic factors, with the latter computed at both global and sub-domain levels to enhance pattern recognition. Multiple regression models are trained within each domain, and top-performing models are fused through weighted ensemble optimization to maximize predictive accuracy. Validated on 18 months of real-world data from a commercial building in Hong Kong, HGSA demonstrates substantial improvements over state-of-the-art methods including LSTM, LightGBM, Prophet, Autoregressive models, and Informer across four forecasting scenarios: 1-hour (CV-RMSE: 0.09), 24-hour (CV-RMSE: 0.161), 7-day (CV-RMSE: 0.157), and 1-month (CV-RMSE: 0.188) ahead predictions. The framework’s model-agnostic design ensures flexibility and practical deployability, while its consistently superior performance across short-, medium-, and long-term horizons establishes HGSA as a comprehensive solution for building energy management, HVAC optimization, and strategic energy planning applications.
Che et al. (Fri,) studied this question.