A novel digital twin modelling approach for forecasting PEMFC degradation and performance

Proton Exchange Membrane Fuel Cells (PEMFCs) are pivotal in automotive and stationary power generation due to their high efficiency and low carbon emissions. Despite extensive research integrating machine learning with physical models of computational fluid dynamics (CFD) for performance prediction, these methods face challenges during the operational phase due to inherent degradation phenomena. The primary barrier to PEMFC commercialisation is durability and reliability under dynamic conditions. This study introduces a novel digital twin modelling method using Bidirectional Long Short-Term Memory (BiLSTM) networks to predict degradation performance using real-world data. Innovative data pre-processing techniques, including a 'window sliding' method with overlapping time bins (e.g., 5-minute intervals), enhance pattern recognition. The model uses time-lagged inputs and a strategic time-based data splitting approach: 65% of the data for training, 15% for validation, and 20% for testing, totalling 1000 hours. This approach aims to refine forecasting accuracy, ensuring reliable and adaptable predictions under varying conditions. Validation on two PEMFC test sets, under constant and dynamic conditions, shows robust performance with R² values over 0.98 and RMSE below 0.11. The results present a robust framework for the development and advancement of predictive modelling, given limited data.