Lithium battery state of health estimation using real-world vehicle data and an interpretable hybrid framework

Abstract

Accurate point estimation and uncertainty estimation of the state of health (SOH) of battery systems are crucial for alleviating user range anxiety and preventing battery safety incidents. On top of approximately 28 million real-world electric vehicle operation data samples, this paper embeds the categorical boosting algorithm as the base learner of the natural gradient boosting, proposing a novel interpretable N-CatBoost hybrid framework to achieve precise point estimation and uncertainty estimation of the battery SOH. Deep charging segments are selected to calculate the initial capacity, and the nonlinear decay trend of the capacity is derived through the Savitzky-Golay filter. Based on the influencing factors of capacity degradation, health features characterizing battery aging are extracted from electric vehicle data as model inputs. The model's hyperparameters are optimized using the particle swarm optimization algorithm, and it is compared with seven other popular machine learning algorithms. The results indicate that the proposed N-CatBoost model achieves the highest estimation accuracy, with mean absolute percentage error and root mean square error of 0.817 % and 1.156 Ah, respectively. In addition, the Shapley additive explanation method is utilized to make the model interpretable, providing full transparency of all predicted values. More importantly, the developed N-CatBoost model can achieve uncertainty estimation of its predictions, with 100 % of the actual capacity values successfully falling within the model's 99 % prediction interval. Therefore, the proposed N-CatBoost model is a reliable and effective method with great potential for deployment on the cloud side for the SOH estimation of batteries in large-scale vehicles.