Cross Domain Transfer Network For State Of Charge Estimation Of Li-On Batteries

Deep learning methods have garnered considerable inter-est in State of Charge (SOC) estimation due to their capabilities to capture complex temporal patterns. However, the success of these methods is based on the labelled target data and the pre-sumption that both the training and testing data originate from identical distributions. In the real word, the operating conditions of batteries are diverse, which makes existing methods in-valid. To address this challenge, a cross domain transfer network (CDTN) with multiscale distribution adaptation (MDA) is proposed to reduce domain distances and align high-dimensional features. In this method, a feature extractor incor-porating convolutional neural network (CNN) and bidirectional long-short term memory (Bi-LSTM) is utilized. In the transfer learning stage, adversarial adaptation and MDA are combined to minimize the distribution discrepancy. Through extensive experiments, the results show the adapted CDTN achieves the best performance under practical discharge applications, which can prevent the batteries from over-discharging and prolong the lifespan effectively.