A dual-model stacking ensemble framework-based recommender system for intelligent adaptive HVAC scheduling in university buildings: predicting set-point reach time and optimal shut-off across variable room sizes and seasonal periods

Buildings account for approximately 40% of global energy consumption, with HVAC systems representing a dominant share. Conventional fixed-schedule operation, which disregards actual environmental conditions, often causes unnecessary pre- and post-conditioning runtime. This study proposes a dual-model HVAC intelligence approach, evaluated at Knowledge Hub Universities in Cairo, Egypt, assessing the predictive accuracy of adaptive HVAC start and shut-off timing and its potential for reducing unnecessary energy consumption.

The approach comprises two independently trained stacking ensemble models, evaluated across room sizes of 15, 30, 45, and 60 m² and observation periods of three, six, nine, and twelve months. Model A (Start Model) predicts the time required to reach the target setpoint temperature, enabling adaptive start-time estimation. Model B (Shut-Off Model) predicts the earliest deactivation time, keeping the room within the comfort band until occupancy ends, based on the last historical instant maintaining indoor temperature within that range. Both models use four daily inputs: outdoor temperature, indoor temperature, relative humidity, and room size. Each combines Decision Trees, Random Forests, XGBoost, and Neural Networks through a meta-learner, with hyperparameters tuned via grid search and k-fold cross-validation. Validation used temporally separated data and varied room sizes to test generalization.

The stacking ensemble consistently achieved the best performance. For three months, the Start Model achieved MAE 0.957 min, R² 0.928; the Shut-Off Model achieved MAE 1.391 min, R² 0.888. For six months, the Start Model achieved MAE 1.051 min, R² 0.926; the Shut-Off Model achieved MAE 1.347 min, R² 0.859. For nine and twelve months, the Start Model achieved MAE 1.053 and 1.221 min with R² 0.909 and 0.903, while the Shut-Off Model achieved MAE 1.347 and 1.328 min with R² 0.879 and 0.909. The resulting data-driven schedule achieved an estimated 19.6% to 21.6% reduction in energy consumption relative to fixed scheduling while maintaining comfort constraints.