Data Driven System Identification for Solid Oxide Fuel Cell Systems

Energy generation is moving away from centralized fossil fuel based generators towards renewable energy to provide clean and reliable sources. Hydrogen-based generation such as solid oxide fuel cell is one of the promising solution. For efficient and optimized operations of the overall system, e.g. frequency or voltage support actions, accurate dynamic models of the generators can be highly beneficial. Those are often not provided by manufacturers in sufficient detail. Since the dynamics of fuel cells are non-linear and depend on a high number of hardto-measure parameters, white-box models are often hard or impossible to implement. The goal of this work is to develop and implement methods for data-driven physics-based model identification for partially unknown solid oxide fuel cells, that function with minimal measurement data. A mechanistic gray box model, a pre-trained feed forward neural network and long short-term memory neural network are implemented. They are evaluated by comparing their output to that of a simulated fuel cell stack in different scenarios. For large variations in operating conditions, the feed forward network shows the best performance. Close to the maximum power point, the long-short term memory based model performs best.