Online lithium battery SOC estimation based on adversarial domain adaptation under a small sample dilemma

In practical applications, a large amount of real labeling data are difficult to obtain because of the differences in the distribution of monitoring data from different batteries, thus resulting in the model's inability to obtain an accurate estimate of the state of charge. Therefore, we propose for the first time the use of an adversarial domain adaptation network (LSTM-DA, Long Short-Term Memory domain adaptation) to extract battery monitoring data and obtain the mapping relationship of the battery charge state. The model is first pretrained using the source domain data. Then, the feature extractor and domain discriminator are made to form an adversarial relationship through the gradient reversal layer (GRL). The simultaneous max-min operation allows the model to be trained for adversarial domain adaptation using source and target domain data. In this study, offline training and online prediction scenarios under incomplete discharge cycles are simulated using Panasonic and LG datasets, and the mean absolute error and root mean square error values of the estimation results on the target domain are 4.71% and 6.59%, respectively. The estimation accuracies under sparse sample conditions are higher than the existing benchmarks of deep learning methods and migration learning methods.