Integrating Physics-Informed and Machine Learning Models for Indoor Temperature Prediction

Nordic buildings face increasing demands for energy efficiency and consistent indoor thermal comfort, creating accurate indoor temperature predictions increasingly crucial for intelligent HVAC control. This research addresses this challenge by proposing a hybrid prediction framework that combines physics-informed learning with traditional data-driven models to improve long-term accuracy. This study uses a multiple-source dataset consisting of building sensors and satellite-derived solar radiation measurements. It includes important physical variables like the inside temperature, the temperature of the supply air, the temperature outside, and the amount of solar radiation. Using this multi-source dataset, several traditional machine learning models were developed for comparison, such as Extreme Gradient Boosting, Feedforward Neural Networks, Random Forest, and Long Short-Term Memory and the proposed Physics-Informed Neural Network model. Findings across the three forecasting horizons (short term: 1 hour, medium term: 6 hours, and long term: 12 hours) show that each model has the strengths at different periods- LSTM performs best in short-term prediction (MAE: 0.18), XGBoost in medium term prediction (MAE: 0.33), and the PINN achieves the most accurate (MAE: 0.29) in long-term prediction through its embedded physical constraints. Overall, the findings highlight the benefits of data-driven and physics-informed methods, and it emphasizes the potential of hybrid modelling approaches for supporting energy efficient, comfort oriented HVAC control in public buildings.