Fuel cell digital twin for remaining useful lifetime prediction and optimisation based on physics-guided neural network

Abstract

Fuel cells are a critical component in the transition from hydrogen to electricity, with significant potential applications in the transportation sector to facilitate sustainable energy transformation. However, issues related to fuel cell durability present major challenges that hinder their widespread adoption. Digital twin (DT) technology has gained significant attention in recent years for enhancing the lifespan of fuel cells through more effective predictive maintenance. However, developing an effective DT model typically requires extensive data. In this work, we propose a novel fuel cell digital twin (FCDT) framework based on the advanced physics-guided neural network (PGNN) method, which addresses the challenge of massive data dependence while also extending the fuel cell’s operational longevity. The PGNN-based approach combines the strengths of physical modeling and data-driven methods, enabling the model to be trained with limited data and providing accurate predictions of the remaining useful lifetime (RUL) using operational parameters as inputs. By interacting with the trained PGNN-based model, the Nelder–Mead algorithm automatically optimises real-time operational parameters, identifying the optimal solution to extend the fuel cell’s remaining lifespan. The experiments validate the PGNN-based model’s capability to capture complex degradation patterns and provide accurate RUL predictions, even under limited data conditions. Additionally, the optimization process confirms the effectiveness of proposed FCDT framework in refining operating parameters and significantly extending the fuel cell’s lifespan.