Scholarlink

<p>Accurate estimation of State of Charge (SOC) is very important to optimize the performance, safety and lifetime battery management systems of electric vehicles, renewable energy storage and portable electronic in use. Conventional SOC estimation methods- including the measurement methods, the model based methods, machine learning methods-often have limitations in regards to the cumulative estimation errors, model dependencies, and physical interpretivity. Recent advances in deep learning-based SOC estimation have shown significant progress to better predict SOC estimation results, however, the existing model often cannot maintain the advantage of robustness showing different preload pressures (mechanical loading) and temperature variation (with and without external conditions, etc.). In order to overcome these limitations, a physics-guided time series transformer model is proposed in this paper to estimate the SOC with self attention mechanisms and physics informed constraints so that better estimation accuracy and reliability are achieved. A transformer -based architecture is used to tackle the challenging task of capturing long-range dependencies in battery time series data, while conservations of energy laws, SOC monotonicity, Voltage-SOC relationships, and range clipping are an integral part of the physical consistency of predictions. Moreover, the external factors including preload pressure and temperature are explicitly taken into consideration and it improves model generalization for diverse real world battery operating conditions. The proposed model is compared with the baseline and state-of-the-art models and are extensively experimentally validated with the lithium-ion battery datasets with varied temperature and preload pressure conditions. Results show that the proposed model has lower scores in mean absolute error and mean-squared error and higher R-squared score than traditional methods, which are deep-learning and model-based methods. An ablation study provides further evidence that each of the physics constraints contributes some performance improvement.</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

Physics-guided time-series transformer for accurate and robust battery SOC estimation

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • B. Subashini,
  • S. Ida Evangeline

摘要

Accurate estimation of State of Charge (SOC) is very important to optimize the performance, safety and lifetime battery management systems of electric vehicles, renewable energy storage and portable electronic in use. Conventional SOC estimation methods- including the measurement methods, the model based methods, machine learning methods-often have limitations in regards to the cumulative estimation errors, model dependencies, and physical interpretivity. Recent advances in deep learning-based SOC estimation have shown significant progress to better predict SOC estimation results, however, the existing model often cannot maintain the advantage of robustness showing different preload pressures (mechanical loading) and temperature variation (with and without external conditions, etc.). In order to overcome these limitations, a physics-guided time series transformer model is proposed in this paper to estimate the SOC with self attention mechanisms and physics informed constraints so that better estimation accuracy and reliability are achieved. A transformer -based architecture is used to tackle the challenging task of capturing long-range dependencies in battery time series data, while conservations of energy laws, SOC monotonicity, Voltage-SOC relationships, and range clipping are an integral part of the physical consistency of predictions. Moreover, the external factors including preload pressure and temperature are explicitly taken into consideration and it improves model generalization for diverse real world battery operating conditions. The proposed model is compared with the baseline and state-of-the-art models and are extensively experimentally validated with the lithium-ion battery datasets with varied temperature and preload pressure conditions. Results show that the proposed model has lower scores in mean absolute error and mean-squared error and higher R-squared score than traditional methods, which are deep-learning and model-based methods. An ablation study provides further evidence that each of the physics constraints contributes some performance improvement.