A deep Conv-Atten-KAN architecture to identify solid oxide fuel cell dynamics under fuel starvation

Solid Oxide Fuel Cells (SOFCs) are coupled, nonlinear systems with interactions of heat transfer, mass transport, and electrochemical reactions that can produce clean, efficient electrical power. Fuel starvation is a frequent abnormal condition during SOFC operation, leading to microstructural changes and ultimately performance degradation. Identifying SOFC performance under fuel starvation is crucial for developing control and monitoring systems to enhance efficiency and longevity. Due to the complex interplay of multi-physics in SOFCs escalated by fuel starvation effects, physics-based modeling approaches are often computationally inefficient for real-world implementation. We introduce an experimental approach that induces fuel starvation in a single SOFC to gather diverse, representative data for data-driven modeling of its dynamics. Additionally, we propose a new deep neural network combining a convolutional Self-Attention mechanism and the Kolmogorov-Arnold approximation theory, called Conv-Atten-KAN, to effectively capture the nonlinear, time-varying dynamics underlying SOFCs affected by fuel starvation. The performance of the proposed model is evaluated using experimental data and benchmarked against state-of-the-art neural networks for system identification. The results indicate an average improvement of $28 \%$ in prediction accuracy compared to recurrent, convolutional, and fully-connected architectures.