A hybrid method with cascaded structure for early-stage remaining useful life prediction of lithium-ion battery

Lithium-ion batteries have been employed extensively in many important applications in the electronics industry. For safety and reliability, it is extremely critical to get an accurate and early-stage remaining useful life prognostic of lithium-ion batteries. However, battery lifetime predictions are challenging due to the nonlinear battery degradation and the operational diversity among batteries. To increase the prediction accuracy, this paper proposes a hybrid framework combining the model-based method and data-driven method. In this framework, after estimating the battery capacity using online operating data, battery lifetime is predicted by the model-based empirical model as well as the data-driven support vector regression model. For the empirical model, its adaptability is improved by updating the parameters dynamically with particle filters. For the support vector regression model, its performance is optimized by an artificial bee colony algorithm. Finally, a fusion method with cascaded structure is proposed to integrate predictions from these two models, which boosts the prediction accuracy by iteratively exerting two concatenated Kalman filters. The generality and effectiveness of the proposed method are verified on battery data sets provided by NASA and our testing bench, respectively. The experimental results illustrate that the proposed method can improve the prediction accuracy of battery remaining lifetime, especially at the early stage. RMSE and MAE of the proposed hybrid framework are within 4 and 3.5. Compared with two existed hybrid methods, RMSE of prediction can be reduced by at least 7.6%. A reduction of not less than 5.9% in MAE of prediction is achieved. (c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

This paper proposes a hybrid framework combining a model-based method and a data-driven method for accurately predicting the remaining useful life of lithium-ion batteries. The method improves prediction accuracy by dynamically updating parameters with particle filters and optimizing the performance of the support vector regression model using an artificial bee colony algorithm. Experimental results demonstrate the effectiveness of the proposed method, particularly in the early stage.