Scholarlink

<p>Lithium-ion (Li-ion) batteries are valued for their high energy density and long life span, resulting from advanced chemical properties. Accurate estimation of battery state-of-charge (SoC) is essential for monitoring battery health, ensuring safe operation, and optimizing performance. Traditional SoC estimation methods often rely on complex chemical modeling, which may lack adaptability to diverse operating conditions. This study introduces a novel real-time, machine learning-based framework for SoC estimation, integrating hyperparameter optimization to achieve enhanced accuracy and generalizability across varied charging and discharging cycles. Using publicly available prognostic datasets from NASA, this work employs a data-driven approach with five key inputs—load voltage, load current, measured voltage, measured current, and battery temperature—to address limitations of conventional techniques. A systematic evaluation of nine state-of-the-art machine learning models was conducted, including linear regression, least absolute shrinkage and selection operator, K-nearest neighbor regressor, categorical boosting regressor (CatBoostR), extra trees regressor, random forest regressor, extreme gradient boosting regressor (XGBR), decision tree regressor (DTR), and gradient boosting regressor (GBR). Among these, XGBR, DTR, and GBR demonstrated superior performance, highlighting their suitability for this domain. A significant innovation lies in the use of GridSearchCV for hyperparameter tuning, optimizing parameters such as learning rate, tree depth, and number of estimators, which resulted in substantial improvement in predictive accuracy metrics (R<sup>2</sup>, MAE, RMSE). The deployment of the best performing model onto an ESP32 Node MCU microcontroller demonstrates the practical application of this research, enabling real-time SoC monitoring with resource-constrained embedded systems. This work not only advances the accuracy of SoC estimation but also provides a scalable solution for battery management systems, making it highly relevant for electric vehicles and renewable energy storage systems.</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

Hyperparameter-optimized machine learning model for real-time estimation of lithium-ion battery state-of-charge

  • 期刊论文

  • 发表日期：

    2025年2月17日

查看全文
  • Walid Touzout,
  • Mohamed Abdessamed Ait Chikh,
  • Fawzi Gougam,
  • Adel Afia,
  • Mohamed El-Amin Bouchikh,
  • Mohammed Ridha Tedjani

摘要

Lithium-ion (Li-ion) batteries are valued for their high energy density and long life span, resulting from advanced chemical properties. Accurate estimation of battery state-of-charge (SoC) is essential for monitoring battery health, ensuring safe operation, and optimizing performance. Traditional SoC estimation methods often rely on complex chemical modeling, which may lack adaptability to diverse operating conditions. This study introduces a novel real-time, machine learning-based framework for SoC estimation, integrating hyperparameter optimization to achieve enhanced accuracy and generalizability across varied charging and discharging cycles. Using publicly available prognostic datasets from NASA, this work employs a data-driven approach with five key inputs—load voltage, load current, measured voltage, measured current, and battery temperature—to address limitations of conventional techniques. A systematic evaluation of nine state-of-the-art machine learning models was conducted, including linear regression, least absolute shrinkage and selection operator, K-nearest neighbor regressor, categorical boosting regressor (CatBoostR), extra trees regressor, random forest regressor, extreme gradient boosting regressor (XGBR), decision tree regressor (DTR), and gradient boosting regressor (GBR). Among these, XGBR, DTR, and GBR demonstrated superior performance, highlighting their suitability for this domain. A significant innovation lies in the use of GridSearchCV for hyperparameter tuning, optimizing parameters such as learning rate, tree depth, and number of estimators, which resulted in substantial improvement in predictive accuracy metrics (R2, MAE, RMSE). The deployment of the best performing model onto an ESP32 Node MCU microcontroller demonstrates the practical application of this research, enabling real-time SoC monitoring with resource-constrained embedded systems. This work not only advances the accuracy of SoC estimation but also provides a scalable solution for battery management systems, making it highly relevant for electric vehicles and renewable energy storage systems.