A hybrid CNN–BiLSTM–AT model optimized with enhanced whale optimization algorithm for remaining useful life forecasting of fuel cell

Abstract

To further improve the remaining useful life forecasting accuracy of fuel cells using classic deep learning models, a convolutional neural network combining bidirectional long and short-term memory networks (BiLSTM) and attention mechanism (AT) is optimized with the enhanced whale optimization algorithm (EWOA). Singular spectrum analysis preprocesses the attenuation data to eliminate noise and enhance its effective information; the CNN–BiLSTM model extracts spatiotemporal features and learns historical and future information; AT further explores the spatiotemporal correlation; and EWOA optimizes its hyperparameters to reduce human intervention error. Results demonstrate that, compared with long and short-term memory, CNN–LSTM, CNN–BiLSTM, CNN–BiLSTM–AT, and CNN–BiLSTM–AT optimized with other algorithms, the CNN–BiLSTM–AT model optimized with EWOA achieves lower root mean square error, mean absolute error, mean absolute percentage error, and relative errors of 0.1951%–0.2059%, 0.1267%–0.1538%, 0.0319%–0.0366%, and 0.026%–0.036%, respectively, with different training data. Importantly, the proposed model still maintains good prediction robustness with over 40% of the missing data.