Robust Parameter Identification of a Reduced-Order Building Thermal Model Using an Inverse Physics-Informed Neural Network

Parameter estimation of reduced order building thermal models is the standard approach for obtaining representations that enable physically consistent temperature predictions. However, due to the constrained operating conditions observed in buildings, temperature dynamics are weakly excited and
key influences overlap in time, making it difficult to infer reliable causal relationships.
This poses a major challenge for parameter identification, making the identification process itself the limiting factor for robust indoor temperature predictions. Here we show that parameter identification using an inverse physics-informed neural network (PINN), which jointly estimates thermal parameters and continuous temperature trajectories under physical constraints, improves predictive accuracy across all prediction horizons and reduces variability in predictive accuracy across thermal models identified from
different training datasets. Moreover, we show that this approach achieves comparable data efficiency to a gradient-based parameter identification procedure. By decoupling trajectory approximation from parameter estimation, the inverse PINN becomes less sensitive to parameter initialization and more robust to variations in the available training data. Our findings suggest that inverse PINNs provide a promising alternative for parameter identification of reduced-order building thermal models in data-limited settings.