人工神经网络
计算机科学
人工智能
残余物
点(几何)
模式识别(心理学)
机器学习
工程类
数据挖掘
特征(语言学)
噪音(视频)
作者
Jian Wu,Alexandre Ravey,Fei Gao
出处
期刊:Energy and AI
[Elsevier BV]
日期:2026-09-24
卷期号:26: 100902-100902
标识
DOI:10.1016/j.egyai.2026.100902
摘要
Accurate knee point prediction is essential for lithium-ion battery safety, lifetime management, and second-life applications. However, nonlinear degradation behavior and strong operating-condition dependence make early and robust knee prediction challenging. This work proposes an interpretable residual learning framework, termed Boruta–Attention–SHAP–Residual Network (BASR-Net), for early knee point prediction. Reference knees are identified using the Bacon–Watts model, while capacity–voltage difference features 𝛥 𝑄 ( 𝑉 ) , incremental capacity (IC) features, and usable-capacity-window (UCW) features are extracted from early-life data. BASR-Net integrates Boruta-based feature selection, SHAP-derived variable weighting, residual learning, and automated hyperparameter optimization to improve prediction accuracy and interpretability. Experiments on three battery aging datasets covering NMC and LFP chemistries demonstrate strong performance across multiple evaluation settings. On Dataset 1, BASR-Net achieves an MAE of 1.73 weeks and an 𝑅 2 of 0.926 under 10-fold cross-validation. Under the stricter entire-DoD-regime holdout test, it achieves an MAE of 1.43 weeks and an 𝑅 2 of 0.950. In CORAL-based transductive cross-dataset transfer, knee-capacity MAEs of 0.0192 Ah and 0.0228 Ah are obtained on two independent target datasets. Additional strict zero-shot experiments further confirm robust transfer without access to target-domain data during model development. Ablation, repeated-seed, and classical-regression comparisons further demonstrate the effectiveness and stability of the proposed framework.
科研通智能强力驱动
Strongly Powered by AbleSci AI