Multivariate gated recurrent unit for battery remaining useful life prediction: A deep learning approach

This paper proposes the gated recurrent unit (GRU)-recurrent neural network (RNN), a deep learning approach to predict the remaining useful life (RUL) of lithium-ion batteries (LIBs), accurately. The GRU-RNN structure can self-learn the network parameters utilizing adaptive gradient descent algorithms, leading to a reduced computational cost. Unlike the long short-term memory (LSTM) model, GRU-RNN allows time-series dependencies to be tracked between degraded capacities without using any memory cell. This enables the method to predict non-linear capacity degradations and build an explicitly capacity-oriented RUL predictor. Additionally, feature selection based on the random forest technique was used to enhance the prediction precision. The analyses were conducted based on four separate cycling life testing datasets of a lithium-ion battery. The experimental results indicate that the average percentage of root mean square error for the proposed method is about 2% which respectively is 1.34 times and 8.32 times superior to the LSTM and support vector machine methods. The outcome of this work can be used for managing the Li-ion battery's improvement and optimization.

Automated summary

The paper presents a deep learning approach using GRU-RNN to accurately predict the remaining useful life of lithium-ion batteries. By self-learning network parameters and utilizing feature selection techniques, the method demonstrates superior prediction accuracy.