Physics-Informed Neural Network for modeling and predicting temperature fluctuations in proton exchange membrane electrolysis
Islam Zerrougui, Zhongliang Li, Daniel Hissel
Abstract
Proton Exchange Membrane (PEM) electrolysis stands as a cornerstone technology in the clean energy sector, driving the production of hydrogen and oxygen from water. A critical aspect of ensuring the efficiency and safety of this process lies in the precise monitoring and control of temperature at the electrolysis outlet. However, accurately characterizing temperature changes within the PEM electrolysis system can be challenging due to the fluctuation of renewable energies. This study introduces an approach integrating data with fundamental physics principles known as Physics-Informed Neural Networks (PINNs). This method solves differential equations and estimates the unknown parameters governing the temperature dynamics within the PEM electrolysis system. We consider two distinct scenarios: a zero-dimensional model and a one-dimensional model. The results demonstrate the PINN’s proficiency in accurately identifying the parameters and solving for temperature fluctuations within the system with different input conditions. Furthermore, we compare the PINN with the Long Short-Term Memory (LSTM) method to predict the outlet temperature of the electrolysis. The PINN outperformed the LSTM method, highlighting its reliability and precision, achieving a Mean Squared Error (MSE) of 0.1596 compared to 1.2132 for LSTM models. The proposed method shows a high performance in dealing with sensor noises and avoids overfitting problems. This synergy of physics knowledge and data-driven learning opens new pathways towards real-time digital twins, enhanced predictive control, and improved reliability for PEM electrolysis and other complex, data-scarce energy systems. • PINNs are applied to identify parameters in 0D/1D PEM electrolyzer models. • High precision in predicting temperature distribution along gas/liquid channels. • PINNs models are compared to classical recurrent neural networks to highlight the model robustness. • PINNs are demonstrated to be robust against sensor noises and prevent overfitting efficiently. • PINNs demonstrate high adaptability to the time-varying parameters and applicability in data-scarce systems.