PILSTM-Based synchronous prediction of internal temperature distribution in hydrogen fuel cells

Abstract

ABSTRACT Accurate prediction of the internal temperature distribution of hydrogen fuel cells is crucial for ensuring efficient, stable, and safe operation. However, existing methods still suffer from limited nonlinear representation of input variables, insufficient extraction of temporal dependencies in conventional physics-informed neural networks (PINNs), and difficulty in achieving synchronous prediction of internal temperature fields. To address these issues, a physics-informed long short-term memory network (PILSTM) is proposed for hydrogen fuel cell temperature prediction. Fourier feature mapping is introduced to enhance the nonlinear representation of the inputs, and an LSTM architecture is embedded into the PINN framework to strengthen temporal dependency learning. In addition, a joint loss function consisting of data loss, physics loss, initial-condition loss, and boundary-condition loss is constructed to enable collaborative optimization of temporal features and physical constraints. Experimental results show that, compared with conventional LSTM and PINN models, the proposed method reduces the root mean square error by 70.05% and 52.9%, respectively, and improves the coefficient of determination by 31.27% and 27.12%, respectively. These results demonstrate that the proposed PILSTM model can achieve accurate synchronous prediction of the internal temperature field of hydrogen fuel cells and has promising application potential.