Gated recurrent unit least-squares generative adversarial network for battery cycle life prediction

Abstract

One of the main concerns of battery management systems is predicting the degradation of lithium-ion batteries, which remaining useful life prediction is an essential tool for prognostic and health management of batteries. In this study, we develop a novel prognostic architecture that is based on a least-squares generative adversarial network with the gated recurrent unit as the generator and multi-layer perceptron as the discriminator and use it to predict the Lithium-ion batteries’ remaining useful life. The proposed method aims to learn the probability distribution of future values in an adversarial training fashion. This generative adversarial network gives more penalties to large errors and addresses the vanishing gradient problem during training. As a result, the predicted values will get closer to the actual data. Furthermore, to obtain high prediction accuracy, time-domain features are evaluated using statistical formulas. The most important features are then selected using the random forest algorithm and fed to the network as a multivariate input set. The performances of the proposed method are tested using a battery degradation dataset from the data repository of Prognostics Center of Excellence at NASA. Furthermore, experimental data from lithium-ion cells at different current rates are conducted for evaluation and verification. The obtained outcomes demonstrate that the designed model achieves the low prediction error of 2.63% and maximum absolute error of 0.02.