Preliminary development of a Digital Twin model for the assessment of a PEM electrolyser degradation using simulation-based data.

Effective green hydrogen production via electrolysis is a fundamental step in sustaining the establishment of such a zero-emission technology. Indeed, the overall loss of performances over time is strongly related to the aging of electrolyser components and it directly impacts the cost effectiveness of green hydrogen production. Therefore, being able to predict and identify degradation pathways can help limit the effects of degradation and optimize maintenance, that is, maximize stack lifespan, fine-tune operational efficiency, and enhance the overall competitiveness of electrolysis technologies. To address these challenges, the development of a Digital Twin (DT) models of physical systems has emerged as a promising approach. Indeed, DTs allow to carry out advanced analysis, such as predictive maintenance, or real-time performance assessment. With this work, it is presented the preliminary assessment of a neural-network based DT model of a 5 MW PEM electrolyser stack (with an annual production of 400 tons of hydrogen), developed and trained starting from a database composed by simulation-based synthetic data. In the first phase, a numerical model of a PEM electrolyser stack, designed to simulate the electrolyser operations under different loads and operating times is developed. The numerical model provides a comprehensive representation of operational dynamics, offering insight into system performances under variable conditions. The model replicates the physical and chemical processes involved in electrolysis operation, including, among others, stack efficiency and degradation rate. The resulting datasets, obtained through the simulation of the numerical model under different load and operation time conditions, are used to train a neural-network based DT model. After training and validation, the performances of the DT model are evaluated by using the metrics techniques as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Coefficient of Determination (R2). The DT model, therefore, can be used to manage the electrolyser operations, in terms of load utilization and uptime-downtime periods, and thus, to evaluate the resulting increase in stack lifetime. Moreover, by leveraging simulation-based data for a preliminary implementation of a DT model, the framework establishes a solid foundation for application in a real-world operational environment. Future developments of this work will focus on deploying the neural-network based DT model in hydrogen production facility, integrating the DT training set with real operating data. The deployment process will involve close collaboration with industrial stakeholders taking part in in the North Adriatic Hydrogen Valley project. The data will be collected and exchanged in real time from the physical system to the DT model, enabling the integration between the physical system and its virtual replication. This dynamic interaction will enhance system monitoring, supporting more effective decision-making and maintenance strategies for water electrolyser systems.