Battery state-of-charge estimation amid dynamic usage with physics-informed deep learning

Accurate estimation of state of charge (SOC) constitutes the basis to enable the reliable operations of lithium-ion batteries. The recent development in deep learning provides an emerging solution to SOC estimation. However, the limited training and testing profiles and the ignorance of battery working principles jeopardise the performance of deep learning-based methods. In this study, we propose to incorporate two kinds of domain knowledge into the deep learning-based methods. First, voltage and current sequences are decoupled into open circuit voltage (OCV), ohmic response and polarisation voltage to augment the input of deep neural networks (DNNs). Second, as conventional DNNs ignore the time-dependency in SOC estimation results, we propose a combination framework to adaptively fuse the SOC estimation results from the DNN and short-term Ampere-hour predictions. The proposed method is validated on a large dataset which is collected by conducting the tests on eight batteries at various real-world driving profiles and is compared with a basic long short-term memory DNN based on the input of only voltage and current. The results show that the proposed method can sharply reduce the SOC estimation root mean square error and maximum absolute error by 30.89% and 64.88%, respectively, with only slightly increased computational cost. Further validations under different temperatures and the applications of the proposed method to other DNNs also demonstrate its effectiveness. These results highlight the potential to boost the performance of DNNs by making effective use of battery domain knowledge.