A novel long short-term memory networks-based data-driven prognostic strategy for proton exchange membrane fuel cells

Abstract

Proton exchange membrane fuel cell (PEMFC) long-term prognostic facilitates reducing the time/cost of the durability tests and is a critical starting point for control/maintenance suggestions. Long short-term memory (LSTM) recurrent neural networks have excellent time series processing capabilities and are proved to be useful for the short-term prognostic of PEMFC. However, LSTM prognostic models usually suffer from accumulated errors and model recognition uncertainties, which make it difficult to break the historical degradation data limitations, resulting in unsatisfactory long-term prediction performance. To tackle the problem, this paper proposes a novel model named navigation sequence driven LSTM (NSD-LSTM) for long-term prognostic. In the strategy, a navigation sequence is firstly generated by using an autoregressive integrated moving average model with exogenous variables. The sequence is then fed iteratively into LSTM in the implementation stage to achieve long-term perdition. The proposed strategy is evaluated using the aging experimental data of two types of PEMFC under different operating conditions. The long-term prognostic performance of the proposed model and other two state-of-the-art prognostic models, namely, nonlinear autoregressive exogenous and echo state network, are evaluated through comparison experiments. The simulation and experimental results show that the proposed prognostic strategy has better long-term degradation trend prediction consistency and remaining useful life estimation robustness. • Navigation sequence driven LSTM model is designed for long-term prognostic. • Aging test data of two fuel cells in constant/dynamic-load are used for evaluation. • Predicted degradation trends can break historical numerical interval limitations. • Long and short term degradation behaviors are considered in RUL estimation.