Physics-informed RC–MLP modelling for building thermal prediction

Reliable multi-horizon building thermal prediction is essential for energy-efficient HVAC operation, model predictive control, and demand response, yet existing models often face trade-offs among physical interpretability, data efficiency, and predictive accuracy. This study develops a physics-informed RC–MLP modelling framework for single-zone indoor temperature prediction using a publicly available high-resolution operational dataset from real office spaces. A 4R3C resistance–capacitance model was first selected as the grey-box thermal baseline by comparing candidate RC structures using both prediction accuracy and physical parameter plausibility. A residual multi-layer perceptron was then trained to correct systematic RC prediction errors using horizon-specific features derived from thermal history, solar forcing, time-of-day variables, and operational indicators. The framework was evaluated at prediction horizons of 5, 30, 60, 180, and 360 min under standard, seasonal shift, and limited-data conditions. Results show that prediction accuracy alone was insufficient for RC structure selection, as candidate models produced similar RMSE values while only the 4R3C structure satisfied all plausibility checks. Residual analysis revealed autocorrelated, diurnal, and solar-driven error structures. On the standard split, the RC–MLP model reduced RMSE by 4.4% at 30 min and 21.3% at 360 min relative to the RC baseline. Under seasonal shift, it retained a 16.8% improvement at 360 min, whereas limited-data experiments showed that mismatched residual calibration could degrade performance. These findings demonstrate that physically plausible RC models provide robust baseline predictions, while residual MLP correction is most effective when calibration data are representative of deployment conditions.