Deep learning with dual-stage attention mechanism for interpretable prediction of proton exchange membrane fuel cell performance degradation

Accurate and reliable estimation of performance degradation in proton exchange membrane fuel cells (PEMFCs) can contribute to the maintenance and risk management of fuel cell systems. However, current research overly emphasizes model ensemble strategies and data preprocessing, neglecting the improvement of internal mechanisms within the models. Few models can adequately capture the long-term dependencies of relevant features. In this study, a dual-stage attention mechanism (DA) network structure called DA-LSTM was developed based on the Long Short-Term Memory (LSTM) neural network to predict the performance degradation of PEMFCs. In the first stage, an input attention mechanism is introduced, utilizing an encoder to adaptively extract important information from each time step of the input features. In the second stage, a temporal attention mechanism is employed to obtain relevant temporal attention factors across all time steps. The proposed approach is tested on various datasets and exhibits favorable predictive performance for PEMFC performance degradation. When compared to different models, the DA-LSTM consistently outperforms other models, demonstrating superior stability and predictive capability. Additionally, the visualization of attention weights explains the relationship between input features and the performance degradation of PEMFC. This enables real-time monitoring of fuel cell systems, which in turn helps prolong the lifespan of PEMFC.