AI-Driven Prognostics for PEM Fuel Cells Using Optimized Machine Learning and Deep Learning Models for Accurate Lifetime Prediction

Abstract Predicting how long proton exchange membrane (PEM) fuel cells will last is important for making them more reliable and reducing maintenance costs. However, most existing methods struggle to understand complex patterns in the data, leading to less accurate predictions. This paper introduces two new models to improve predictions. The first model, attention-based long short-term memory (ALSTM), is based on LSTM networks with an attention mechanism. This helps the model focus on the most important time steps in the data, making predictions more accurate. The second model, B-Ridge-JADE, selects the most useful features from the data using an adaptive differential evolution method (JADE). These selected features are then used to train a Bayesian Ridge (B-Ridge) regression model, which improves accuracy and reduces the computational costs. To test the models, two datasets (FC1 and FC2) from the IEEE 2014 Prognostics dataset were used. The results were compared with other deep learning (DL) and machine learning (ML) models using mean absolute percentage error (MAPE) and root mean squared error (RMSE). The results show that ALSTM performs better than all DL models, while B-Ridge-JADE outperforms both DL and ML models. For FC1, B-Ridge-JADE achieves RMSE of 0.001119 and MAPE of 0.000212. For FC2, it achieves RMSE of 0.001167 and MAPE of 0.000265. These findings show that the proposed models can help improve efficiency, reduce costs, and enhance maintenance planning for fuel cell systems in industrial applications.