A performance degradation prediction model for PEMFC based on bi-directional long short-term memory and multi-head self-attention mechanism

Proton exchange membrane fuel cell (PEMFC) is a highly promising renewable energy conversion technology. However, durability issues have hindered their large-scale commercialization process. Performance degradation prediction is an essential component of PEMFC prognostics and health management and is critical for extending the service life of fuel cell. Given that, this paper proposes a novel data-driven prediction model that fuses multi-head self-attention (MHSA) mechanism and bi-directional long short-term memory (BiLSTM). This model can effectively capture different types of dependencies from large-scale high-dimensional data and achieve global information modeling. Specifically, the preprocessed historical voltage data and PEMFC system operating parameters are fed into the proposed prediction model. Where BiLSTM understands the contextual information and temporal dependencies in sequence data by calculating the hidden states in both forward and backward directions. MHSA captures the complex relationships and extracts key information in the input sequence by simultaneously learning multiple sets of attention weights between different locations. Finally, the proposed model is validated based on the health monitoring data under stationary and quasi-dynamic conditions. The validation results indicate that the proposed model can ensure absolute errors of less than 0.6 × 10−3 V for at least 71.9% of the prediction results under stationary and quasi-dynamic conditions (less than 1.2 × 10−3 V for at least 97.6% of prediction results).