Remaining useful life prediction of PEMFC based on matrix long short-term memory

Prognostics and health management play a crucial role in ensuring the safety and maintenance of fuel cell vehicles. To address the insufficient prediction accuracy in existing models when handling long-term sequences and high-dimensional data, this paper proposes a proton exchange membrane fuel cell (PEMFC) remaining useful life prediction method based on a matrix long short-term memory (M-LSTM) network. Specifically, the Spearman correlation coefficient analysis is first employed to quantitatively evaluate the multidimensional correlations within the experimental data, identify key variables highly associated with fuel cell degradation, optimize feature selection, and achieve dimensionality reduction. Subsequently, an M-LSTM life prediction model is developed, incorporating an exponential gating mechanism and a matrix storage unit to enhance the global modeling capability of data-driven methods in complex nonlinear feature learning and long-sequence data processing. The results demonstrate that, compared with the LSTM model, under single-step prediction, M-LSTM reduces MAE and MSE by 76.3% and 58.2%, respectively, while increasing R2 by 2.3%. Under multi-step prediction, M-LSTM reduces MAE and MSE by 79.4% and 60.4% respectively, while improving R2 by 5.2%.