Physics-Informed Machine Learning for State of Charge Estimation of Lithium-Ion Batteries

An accurate estimation of battery state of charge (SOC) is crucial to ensure the safety of Li-ion batteries (LIBs), and the recent boom in machine learning (ML) has illustrated a strong potential for SOC estimation. However, the ignorance of LIB internal dynamics greatly limits the performance of ML-based methods. Therefore, this work incorporates multiple battery internal dynamics into the ML-based method to ensure enhanced performance. First, using a simple mechanism based on the LIB model, the open circuit voltage, ohmic response, and temperature-corrected pattern are extracted from the measured data and provided as inputs to the neural networks (NNs). Then, to ensure time adaptability and accuracy, the predictions of the proposed design are adjusted using short-term Ampere-hour results. The proposed technique is validated using different real-world driving profiles. The comparative results with benchmark NNs-based methods verify a >60% reduction in root mean square error, demonstrating the effectiveness of incorporating battery internal dynamics into NNs.