Equalizing data lengths across temperatures to enhance deep learning training for state of charge prediction

Abstract

State of charge (SOC) is an important value for electric vehicles as it provides information about how long they can be driven. Predicting SOC accurately has become an important research topic. Several systems and studies have been developed using various methods and algorithms such as deep learning, Kalman filter, and current counting to focus on SOC prediction at constant temperatures. However, it becomes more difficult to estimate SOC accurately as the temperature decreases. Since real temperatures vary between-30 degrees C and 50 degrees C depending on the season and region, temperature directly affects the performance of batteries. When the battery datasets publicly available on which many studies have been conducted are examined, it is found that the processing time varies according to temperature in the data of the same drive cycles. Therefore, when the data is intended to be utilized in artificial intelligence training, there is an imbalance in the data between temperatures, making it difficult for the trained model to generalize to different temperatures. This study proposes a method to solve the issue of imbalance in data between temperatures. The method generates multiple drive cycles from the same drive cycle by changing the sampling frequency of a drive cycle. By using these drive cycles in approximately equal amounts of data at all temperatures, the overall prediction error of the trained model is reduced. The proposed approach increases the accuracy of SOC estimation at low temperatures. Therefore, it has been shown that the proposed approach can be used in SOC estimation.