CNN-DBLSTM: A long-term remaining life prediction framework for lithium-ion battery with small number of samples

Accurate prediction of lithium-ion batteries remaining useful life (RUL) is crucial for good energy management and performance enhancement of aerospace vehicles during operation. Currently, numerous intelligent prediction methods based on neural networks have been developed. However, these existing methods rely heavily on the amount of training data and can only achieve satisfactory performance when sufficient samples are available. Given that in practical applications, the historical data sample size is often insufficient, this article proposes a lithium-ion battery remaining useful life prediction method based on a combination of convolutional neural networks and deep bidirectional long short-term memory networks (CNN-DBLSTM), which solves the problem of poor prediction performance caused by insufficient historical data samples used for training. The experimental verification is carried out on the battery data set of the NASA Prognostics Center of Excellence (PCoE) and the battery data set of the University of Maryland Centre for Advanced Life Cycle Engineering (CALCE). The influence of the amount of training data on the prediction performance of CNN-DBLSTM was explored. Experimental results show that CNN-DBLSTM can achieve satisfactory prediction performance even with a small amount of training data. Specifically, under different NASA training data ratios, the root mean squared error (RMSE) can be maintained at around 0.011, and the mean absolute error (MAE) can be maintained at around 0.006. Under different CALCE training data ratios and long-term prediction scenarios, both RMSE and MAE metrics can be kept around 0.025. Overall, the proposed prediction method can achieve high prediction accuracy, but also has strong robustness and generalization.