Machine learning and feature engineering-based anode potential estimation method for lithium-ion batteries with application

Anode potential is a direct characterization parameter for lithium plating, and the real-time and accurate anode potential estimation is crucial for the safe and stable operation of lithium-ion batteries. An ideal method for estimating anode potential should possess characteristics, such as real-time capability, non-destructiveness, high accuracy, and strong adaptability to varying operating conditions. Motivated by this, this paper proposes a real-time anode potential estimation method based on machine learning and feature engineering. Firstly, by considering both the externally measurable signals and the internal mechanisms of the battery, we construct nine feature parameters that can be obtained online, forming the original feature library for anode potential. Subsequently, four highly robust features are selected to form the final feature set based on correlation and feature importance evaluations. Then, we train three machine learning models and evaluate their performance on the nine test datasets. The results show that the anode potential estimation method based on the backpropagation neural network (BPNN) performs the best with an average root mean square error (RMSE) of no more than 1.738 mV on different test datasets. Because of the superiority of the feature set used, the method proposed in this paper demonstrates enhanced accuracy and generalization capability. Compared to the existing literature, the RMSE is reduced up to 71.29 %. Finally, an online charging control system is established based on the trained BPNN and proportional-integral-derivative controller, and the optimized charging method based on the anode potential threshold is developed. Our proposed method can reliably estimate the anode potential in real-time, which provides a novel and effective solution for online charging control and safety warning of lithium-ion batteries.