HB-PINNs: A Heat-Balance Physics-Informed Framework for Building Energy Loading Forecasting with Limited Data

Accurate prediction of building energy demand is essential for effective building control, demand response, and integrated energy management. Physics-informed neural networks (PINNs) offer an effective way to combine data-driven learning with physical knowledge, but most PINNs for building heating demand prediction still rely on simplified resistance-capacitance (RC) models as physical constraints. Since RC models cannot fully represent whole-building heat-balance processes, their simplified thermal representation limits the predictive performance of PINNs. This study introduces heat-balance physics-informed neural networks (HB-PINNs) as a lightweight intermediate modelling framework between purely data-driven learning and detailed physics-based simulation. By incorporating a more complete heat-balance constraint, HB-PINNs provide a physically informed but computationally practical representation of building thermal behaviour. This enables heating demand to be calculated more accurately from physically meaningful intermediate variables, particularly under limited-data conditions. Under identical input conditions, HB-PINNs are evaluated against conventional PINNs across multiple test sets, training data ratios, and prediction horizons. HB-PINNs achieve better predictive performance without significantly increasing computational cost, yielding an average ????2 of 94.99% across all scenarios and training ratios, compared with 88.69% for conventional PINNs. The results show that HB-PINNs provide higher accuracy and stronger adaptability under varying operating conditions, offering an extensible physics-informed framework for building energy modelling.