An Online Transfer Learning Framework for Cell SOC Online Estimation of Battery Pack in Complex Application Conditions

Abstract

Complex application conditions like different temperatures, different aging statuses, and cell inconsistency would cause a distribution difference between data domains and lead to a significant State of Charge (SOC) estimation error. Transfer learning’s Domain Adaptation (DA) offers an effective way to lessen the disparity in distribution across the data domains. However, the DA-based SOC online estimation method would lead to negative transfer learning due to differences in the target space. Besides, the DA method is challenging to apply SOC online estimation due to the complex feature data transformation and calculation. Based on SOC online estimation, this work first defines a time-varying partial target space-based domain adaptation problem. Then, an online transfer learning framework is designed to solve the above problem by learning the transfer transformation mechanism. Besides, a new Hoeffding-based extreme learning machine algorithm is proposed to learn the transformation mechanism better. As experiments confirmed, the proposed method is effective and can obtain accurate SOC estimation results under complex application conditions.