Abstract
A solid oxide fuel cell (SOFC) is a multiphysics system that involves heat transfer, mass transport, and electrochemical reactions to produce electrical power. Reduction and re-oxidation (Redox) cycling is a destructive reaction that can occur during SOFC operation. Redox induces various degradation mechanisms, such as electrode delamination, nickel agglomeration, and microstructural changes, which should be mitigated. The interplay of these mechanisms makes a post-Redox SOFC a nonlinear, time-varying, nonstationary dynamic system. Physics-based modeling of these complexities often leads to computationally expensive equations that are not suitable for the control and diagnostics of SOFCs. Here, a data-driven approach based on dilated convolutions and a self-attention mechanism is introduced to effectively capture the dynamics underlying SOFCs affected by Redox. Controlled Redox cycles are designed to collect appropriate experimental data for developing deep learning models, which are lacking in the current literature. The performance of the proposed model is validated on diverse unseen data sets gathered from different fuel cells and benchmarked against state-of-the-art models, in terms of prediction accuracy and computation complexity. The results indicate 31% accuracy improvement and 27% computation speed enhancement compared to the benchmarks.