Cross-Domain State-of-Charge Estimation of Li-Ion Batteries Based on Deep Transfer Neural Network With Multiscale Distribution Adaptation

Abstract

The success of deep learning in state-of-charge (SOC) estimation relies on the assumption that training and test data have the same distribution. However, this assumption is mostly invalid in real-world applications because the battery operating conditions are diverse and considerable battery data are difficult to obtain to train a specific deep estimator for each condition. To solve these problems, a deep transfer neural network (DTNN) with multiscale distribution adaptation (MDA), which generalizes the deep estimator for domain adaptation, is proposed for cross-domain SOC estimation. In this method, DTNN composed of convolutional and bidirectional recurrent neural networks is constructed to learn nonlinear dynamic features of battery measurements from the source and target domains. Then, MDA is developed to minimize the distribution discrepancy of the high-level transferable features between the source and target domains at multiple scales by simultaneously imposing constraint terms on the DTNN layers. These domain-shared features that obey to small discrepancy can enhance the generalizability and robustness of DTNN for target estimation tasks. Through extensive experiments on three battery data sets, the results show that compared with the state-of-the-art transfer learning methods, the adapted DTNN learned with limited battery data achieves the best performance under low-capacity discharge condition and charge-pause-discharge condition.