Hybrid Attention DNN Capacity Prediction for Fast Charging Batteries Based on Partial Charging Data

To ease range anxiety and shorten recharging time, today's electric vehicles (EVs) often require large-capacity battery systems with fast charging. It is important to monitor and evaluate the battery state of health (SOH) for EVs' safe and durable operation. In this work, an attention-based hybrid deep neural network (DNN) is developed to predict battery charge capacity at the end of each charging cycle. First, correlation analysis is implemented to reveal the correlations of charging variables to the charge capacity. Second, the charging curves of all battery cells are extracted, and 90 points of charging data collected within 5 minutes are used as input to predict the charge capacity. Third, by using the charging time (t), charging current (I), and cell voltage (V) data on the early constant current charging stage as inputs, the proposed method generates charge capacity prediction with an average RMSE of 0.0205 Ah (1.86% of the nominal capacity of 1.1 Ah), which outperforms mainstream deep learning methods including solo and hybrid models.