集成学习
计算机科学
人工智能
集合预报
机器学习
Boosting(机器学习)
极限学习机
高斯过程
多任务学习
超参数
电压
残余物
模式识别(心理学)
区间(图论)
高斯分布
人工神经网络
阿达布思
克里金
特征提取
钥匙(锁)
混合模型
学习迁移
算法
健康状况
特征(语言学)
过程(计算)
二次方程
支持向量机
相关系数
一般化
结构健康监测
趋同(经济学)
监督学习
作者
Bin Wang,Wenrui Cao,Mingtao Li,Jun Wang,Feng Zhao,Yizhe Yan,Chenyu Wang
出处
期刊:Energy & Fuels
[American Chemical Society]
日期:2026-04-08
卷期号:40 (16): 9073-9084
被引量:2
标识
DOI:10.1021/acs.energyfuels.5c06708
摘要
Data-driven methods are increasingly used for noninvasive state of health (SOH) research of lithium-ion batteries (LIBs). Conventional data-driven approaches for the battery SOH estimation usually extract aging features from partial charging curves. However, existing studies could not conclusively identify the specific charging curve interval with optimum aging characterization capability. To bridge this gap, this study introduces a two-stage correlation analysis method combining gray relational analysis (GRA) and Pearson correlation analysis to pinpoint the optimal charging voltage interval and refine the candidate features. On this basis, a novel ensemble learning framework is developed, featuring a residual-enhanced mixture of experts (MoE) network. Unlike conventional stacking methods, this framework innovatively integrates meta-feature encoding and residual fusion to mitigate information bottlenecks, effectively synergizing the complementary strengths of the Gaussian process regression (GPR), extreme gradient boosting (XGB), and elastic net (ENET). Furthermore, the hyperparameters of the framework are optimized with a multistrategy beluga whale optimization (BWO) algorithm, which can effectively guarantee convergence and escape local optima. The final results demonstrate two key findings: (1) the ensemble learning framework can improve the coefficient of determination by a maximum of 38.36% compared with individual machine learning models; (2) feature extraction approach from the optimal charging voltage interval can reduce the mean absolute error of the SOH estimation by up to 57.32%, compared with conventional partial charging curve approaches. In addition, a transfer learning approach based on the ensemble learning framework is successfully demonstrated for enhancing the model generalization capability across LIBs with different nominal capacities. Hence, the proposed ensemble learning framework based on the optimized interval of charging curves offers clear benefits and shows considerable promise in the field of high-precision SOH estimation for LIBs.
科研通智能强力驱动
Strongly Powered by AbleSci AI