HARM: Heterogeneous Feature Fusion and Hierarchical Adaptive Reconstruction Model for HVAC on/off Timing Prediction

Intermittent use of meeting rooms in public buildings accounts for a substantial share of avoidable heating, ventilation, and air-conditioning (HVAC) energy consumption, making accurate HVAC on/off lead-time prediction important for energy-efficient operation. However, existing methods face three limitations: a single interpolation strategy cannot simultaneously preserve local temperature gradients and enforce physical plausibility; hierarchical prior information cannot be used reliably to initialize efficiency features for newly installed devices under cold-start conditions; and differences in scale and sampling frequency across heterogeneous data sources obscure cross-level coupling. To address these challenges, this study proposes the Hierarchical Adaptive Reconstruction Model (HARM) for HVAC on/off lead-time prediction in meeting rooms in public buildings. HARM uses a structured heterogeneous feature fusion strategy to organize device-, system-, and environment-level states together with their physics-derived features into a unified model input. For historical indoor-unit temperature records that are irregularly sampled and incomplete, a four-layer adaptive time-series reconstruction module performs pattern-specific, shape-preserving reconstruction and produces physically constrained minute-level temperature trajectories for historical-state recovery and supervised response duration calculation. A hierarchical Bayesian cold-start estimation module further provides reliable estimates of room temperature change efficiency for newly installed devices by sharing priors across the city–park–room hierarchy. Independent predictors are trained for four sub-scenarios—cooling-on, heating-on, cooling-off, and heating-off—to estimate the response duration required for the indoor temperature to reach the comfort temperature boundary after the corresponding on/off operation under the current operating conditions. Because time-series reconstruction and hierarchical statistical parameter estimation are performed offline, online inference requires only structured feature construction and lightweight gradient boosting decision tree (GBDT) inference, without deep neural networks or GPU acceleration, thereby limiting online computational overhead. Experiments on real-world operational data from 15 cities in China show that HARM achieves the highest prediction accuracy in all four sub-scenarios, with an overall accuracy of approximately 94.5% and an accuracy of 96.20% in the cooling-on sub-scenario. The HARM significantly outperforms XGBoost, LightGBM, long short-term memory (LSTM), and support vector machine (SVM) baselines, demonstrating strong predictive performance and potential for lightweight operational implementation.