• Many existing models are trained primarily with first-life data and exhibit partial transferability performance in second-life conditions [23].
• Predictions contain considerable errors because of operational differences, such as varying temperature profiles, dynamic loading, and different degradation [24].
• Deep-learning models may guarantee precision, albeit with more volume, and often lack the relevant physics of the problem.
• The uncertainty quantification does not exist in many works [25].
根据已发现的不足和研究动机,本文的创新点如下:
• Utilization of first-life data from Li-ion batteries to estimate the SoH for their second life estimated by combining deterministic modeling (SVR) with a probabilistic interface (GPR) for improved accuracy and interpretability in second-life SoH estimation.
• Feature extraction from the first-life dataset, including time-, voltage-, cycle-, and current-based features. Selection of ten key features using the random forest (RF) out-of-bag (OOB) permuted error algorithm. Based on first-life data, the re-estimation of the second-life SoH by utilizing these ten features.
• Cross-battery validation approach and added degradation and noise strengths the acceptance of the proposed model for more generalized datasets and real-life applications. An uncertainty quantification work is conducted for the proposed method and compared with GPR to show the efficacy.
• Testing of hybrid method for different threshold levels of first life cycles to show its effectiveness under different first life cycle scenarios.
• Comparison of the proposed approach with ensemble methods based on error metrics and computational speed. Four different metrics i.e., root mean square error (RMSE), mean absolute error (MAE), R², and mean absolute percentage error (MAPE).