Physics-guided framework of neural network for fast full-field temperature prediction of indoor environment

This study explored the fast full-field temperature prediction of indoor environment, which is valuable for improving energy efficiency and indoor thermal comfort. To this end, a physics-guided framework of neural networks was proposed to fast predict the full-field temperature by integrating the numerical simulation, physical laws and sparse measured data. The proposed framework comprised three basic components: (i) a surrogate model, (ii) a discrepancy model, (iii) a recovery model. First, a physics-informed neural network-based surrogate model approximating the behavior of high-fidelity simulation model was constructed to capture the trend of the temperature evolution. Thereafter, the transfer learning-based discrepancy model minimizing the discrepancy between the observation and direct numerical simulation was constructed with limited available observation data. Last, integrating the parameters of both surrogate and discrepancy model, the recovery model was built to give the best and fast full-filed temperature prediction. The proposed approach can bridge the gap between the numerical simulation and real world. The performance was validated and the results demonstrate that the proposed method provide a better full-field temperature prediction for the indoor environment with a small number of measured data.