A Two-Phase Physics-informed Machine Learning Framework for Predicting Indoor Temperature and Cooling Energy in Present and Future Climates

Climate change is expected to substantially modify the thermal behaviour and cooling requirements of residential buildings, particularly in Mediterranean regions. Accurately capturing these effects requires modelling tools capable of providing reliable short-term predictions while remaining robust under future climatic conditions. In this context, Physics-Informed Neural Networks (PINNs) offer a promising solution by embedding physical conservation laws directly into the learning process, thereby improving generalisation and physical consistency. This study proposes a two-stage PINN-based framework to predict indoor operative temperature and cooling electricity power in a multi-residential building located in Bari, Southern Italy. Typical Meteorological Year (TMY) data and future climate projections for 2050 and 2100 are employed to generate dynamic simulation datasets using EnergyPlus for training and validation. In Phase I, a single-zone energy balance PINN is used to predict operative temperature, explicitly accounting for thermal inertia and external heat gains. In Phase II, the predicted operative temperature is coupled with environmental and thermal features to estimate the hourly cooling electricity power at the next time step. The results demonstrate stable predictive performance across all climate scenarios. Operative temperature prediction achieves RMSE values between 0.094 and 0.152 °C across scenarios, corresponding to only 0.96-1.52% of the total summer temperature range (9.36-10.02 °C). Recursive 24-hour forecasts demonstrate stable error propagation, with RMSE remaining below 0.22 °C at the end of the prediction horizon. Cooling electricity power is predicted with RMSE values between 0.089 and 0.144 kW, with limited bias across scenarios. Increased deviations are mainly observed during sharp HVAC activation peaks under future climate conditions, while overall temporal dynamics and daily load patterns are consistently reproduced. The proposed framework provides a physically consistent and scalable approach for integrating machine learning with building energy simulation, supporting future applications in building energy management, digital twins, and predictive control under climate change scenarios.