Time-series forecasting of indoor temperature: Toward sensor-free occupant-centric HVAC control

Indoor temperature is a critical variable for occupant-centric HVAC control in buildings, directly influencing thermal comfort and control performance. Many existing OCC deployments heavily rely on dense sensor networks, which face practical challenges related to cost, maintenance, and reliability in real buildings. To address this limitation, this study proposes a sensor-free time-series forecasting framework for indoor temperature prediction. The proposed approach relies only on limited and readily available information in buildings, including air conditioner (AC) operational settings, outdoor environmental conditions, temporal information, and leave- and arrival-related variables. An eXtreme Gradient Boosting (XGBoost) regression model is developed to predict indoor temperature in a time-series manner under realistic sensing constraints. The model is trained and evaluated using a time-series cross-validation scheme that strictly preserves chronological order and effectively prevents temporal leakage. The developed model achieves an R2 of 0.61, RMSE of 0.29 °C, MAE of 0.24 °C, MSE of 0.09, MAPE of 0.90%, and SR0.5 of 90.88%, demonstrating that indoor thermal dynamics can be reliably predicted without relying on indoor environmental sensors. In addition, SHapley Additive exPlanations (SHAP) are employed to analyze feature contributions and enhance model interpretability. The novelty of this study lies in demonstrating that reliable indoor temperature forecasting can be achieved under realistic sensing constraints by systematically leveraging only readily available information in buildings.