Life prediction of fuel cells based on the LSTM-UPF hybrid method
Life prediction for fuel cells is crucial to fuel cell health management,offering their operation and maintenance guidance.Advantages of a long short-term memory neural network(LSTM)and an unscented particle filter(UPF)algorithm are combined to enhance condition adaptability in life prediction and ensure accuracy.The proposed LSTM-UPF hybrid method is designed to predict fuel cell life under steady-state and quasidynamic conditions.Initially,experimental data used for model training is optimized and decomposed into high-frequency and low-frequency components using the discrete wavelet transform technique.The LSTM algorithm predicts these components,forecasting the long-term aging trend of fuel cells.Drift correction is then applied to refine the trend prediction results.The fuel cell's remaining useful life(RUL)is estimated using the UPF algorithm based on the long-term aging trend.Evaluation indexes,including prediction life end,life prediction error,confidence interval width,and RUL prediction error,are adopted to assess different life prediction methods.Comparative results demonstrate that the LSTM-UPF hybrid life prediction method yields RUL prediction errors of 4.1％and 3.4％for steady-state and quasidynamic conditions,respectively.Moreover,it exhibits more accurate RUL predictions,high-quality confidence intervals,and strong adaptability in both scenarios.This study enhances the accuracy and confidence level of fuel cell life predictions.
万方数据
维普
