Scholarlink

<p>Effective and accurate estimation of the Remaining Useful Life of lithium-ion battery management systems is critical for its liability, safety and optimisation of cost ranging from electric vehicles to grid-scale energy storage. A variety of electrochemical and physics-based models offer valuable perspectives but these models very often fail to retain various attributes such as ability to scale, to compute and to modify to suit different operating and environmental conditions. In this work, we perform a comparative study of four supervised machine learning algorithms viz., Random Forest, Gradient Boosting, XGBoost and Multi-Layer Perceptron for estimation of remaining useful life on the basis of structured cycle-level sensor data. All the machine learning algorithms are implemented as regressor. Our results indicated that Random Forest Regressor provides the best result with mean squared error as 120.45 and R² score as 0.92. It also indicates that this model is the best suited to capture complex nonlinear degradation. In safety-critical areas, the need to enhance model interpretability led to the application of post-hoc Explainable AI (XAI) methods LIME and SHAP, which provided local and global feature attribution explainability, respectively. Results of our analysis show that discharge time and maximum discharge voltage are the most critical predictors and proportionally explain electrochemical theory is the main basis of battery degradation. The explainable machine learning approach ensures that model transparency complements predictive accuracy, therefore supporting responsible safety-critical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable Ensemble Learning to Predict Remaining Useful Life of Li- Ion Battery Using LIME and SHAP

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Siddhartha Roy,
  • Madhubrata Bhattacharya,
  • Ajanta Dasand,
  • Debabrata Datta

摘要

Effective and accurate estimation of the Remaining Useful Life of lithium-ion battery management systems is critical for its liability, safety and optimisation of cost ranging from electric vehicles to grid-scale energy storage. A variety of electrochemical and physics-based models offer valuable perspectives but these models very often fail to retain various attributes such as ability to scale, to compute and to modify to suit different operating and environmental conditions. In this work, we perform a comparative study of four supervised machine learning algorithms viz., Random Forest, Gradient Boosting, XGBoost and Multi-Layer Perceptron for estimation of remaining useful life on the basis of structured cycle-level sensor data. All the machine learning algorithms are implemented as regressor. Our results indicated that Random Forest Regressor provides the best result with mean squared error as 120.45 and R² score as 0.92. It also indicates that this model is the best suited to capture complex nonlinear degradation. In safety-critical areas, the need to enhance model interpretability led to the application of post-hoc Explainable AI (XAI) methods LIME and SHAP, which provided local and global feature attribution explainability, respectively. Results of our analysis show that discharge time and maximum discharge voltage are the most critical predictors and proportionally explain electrochemical theory is the main basis of battery degradation. The explainable machine learning approach ensures that model transparency complements predictive accuracy, therefore supporting responsible safety-critical applications.