Physics-informed neural networks based prediction of spatial hydrogen leakage concentration fields in hydrogen refueling stations

Accurate and timely prediction of hydrogen leakage dispersion is essential for safety management in hydrogen refueling stations (HRS). This study proposes a physics-informed neural networks (PINNs)-based model that reconstructs the spatial hydrogen concentration field in real-time from sparse monitoring data. The model integrates the continuity equation, momentum conservation, and convection-diffusion equations as physical constraints, and is validated under two representative environmental wind scenarios: downwind and upwind. Numerical experiments show that the PINNs model achieves superior performance, particularly under limited training data. For instance, under complex upwind conditions, it attains an R 2 of 0.932 using only 5 % of the data, outperforming a conventional neural network trained on 20 % (R 2 = 0.905). This work establishes a fast, robust, and physically consistent framework for hydrogen risk monitoring, providing technical support for safe operation in hydrogen infrastructure and demonstrating strong potential for real-world deployment. • Proposed a physics-informed prediction model for hydrogen concentration field. • Physical equations and boundary conditions improve physical consistency. • The model obtains better results than conventional models with small sample data. • Achieves R 2 = 0.966 with only 10 % training data in complex upwind scenarios.