PINN with dynamic constraint optimization for complex air-based TABS thermal dynamics prediction

The growing complexity of HVAC systems poses challenges for thermal dynamics modelling. Physical modelling and data-driven approaches, despite their respective advantages, are difficult to meet the high accuracy requirements alone. Therefore, it is crucial to explore modelling strategies that incorporate the advantages of both. In this study, a novel physics-informed neural network (PINN) model with adaptive adjustment of physical constraint weights is proposed and designed for predicting room air temperature in air-based thermally activated building system (TABS). The model integrates the long short-term memory (LSTM), attention mechanism (AM), underlying physical model, and a dynamic physical constraint weight (λ[jls-end-space/]) adjustment strategy, which effectively integrates the physical laws and data information, significantly reduces the modelling error under different operating conditions such as ventilated and non-ventilated rooms, and flexibly balances the conflict between the physically-driven constraints and the data-fitting objectives. Compared to the optimal fixed λ PINN ((Formula presented) 0.6), the proposed dynamic λ strategy (exponential time-varying strategy) reduces the RMSE by 14.6%. The RMSE of exponential time-varying strategy-based model is reduced by 26.2%, 49.7%, 77.4% and 26.2% compared to neural network, LSTM model, 2R2C model and physical model respectively. The PINN framework based on optimization of dynamic physical constraint weights developed in this study provides an effective basis for realizing highly accurate temperature prediction of complex building thermal systems, especially air-based TABS.