A parallel Seq2Seq neural architecture for long-horizon performance forecasting and online condition monitoring of fuel cells

Solid Oxide Fuel Cells (SOFCs) are efficient and environmentally friendly power generation technologies that have seen increasing applications in recent years. However, their material instability at high operating temperatures makes them susceptible to degradation and unexpected failures. This demands developing robust monitoring algorithms to predict the onset of degradation, thereby facilitating maintenance strategies to extend SOFC longevity. SOFCs suffer from complex, non-stationary degradation processes such as nickel (Ni) reoxidation, where both externally and internally driven dynamics interact over long horizons. Accurately forecasting these dynamics is challenging because existing sequence-to-sequence models either overlook the coupling of internal and external factors or incur prohibitive computational costs that hinder real-time deployment. To address these challenges, this paper introduces an innovative deep neural forecasting network that integrates parallel computing layers, including one-dimensional dilated convolutions and multilayer perceptrons (MLPs), within an interpretable encoder–decoder framework. The design effectively captures redox-induced dynamics while offering efficient memory usage and parallelism. The effectiveness of the architecture is demonstrated through long-horizon forecasting of SOFC performance under Ni reoxidation degradation, benchmarked against transformer-, recurrent-, and MLP-based architectures. Extensive experiments on SOFC degradation datasets, collected from multiple lab-scale fuel cells, demonstrate that the proposed model consistently outperforms state-of-the-art models, achieving an 16 % improvement in Root Mean Squared Error, 20 % in symmetric Mean Absolute Percentage Error, and 23 % in Weighted Absolute Percentage Error metrics. More importantly, our model requires 21 % less Graphics Processing Units (GPU) resources than its counterparts while offering a 36 % faster latency during inference–key advantages for real-time deployment in early-stage degradation detection systems in fuel cell technology. • A novel sequence-to-sequence model that leverages dilated convolutions and MLPs, offering a competitive alternative to transformer-based models for long-horizon forecasting in fuel cells. • Comparison with state of the art neural forecasting models on diverse experimental data from fuel cells operated under nickel redox degradation. • Achieves approximately 19 % higher forecast accuracy with 21 % less GPU usage and 36 % faster runtime, enabling effective real-time monitoring of fuel cell degradation.