Scholarlink

<p>This paper investigates the impact of input variable selection on the accuracy of lithium-ion battery state-of-charge estimation (SOCE). Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) algorithms were used for SOCE. Traditional methods use voltage, current, and temperature as inputs. In contrast, this work considers additional signals from the vehicle’s CAN bus including speed, engine torque, and heater power to determine whether a richer set of inputs enhances prediction performance. The approach applies a wavelet transform to remove signal noise and employs correlation analysis followed by principal component analysis (PCA) to lower dimensionality and improve signal quality by removing collinear variables. Two modeling strategies are compared: one using features selected by correlation analysis, and the other using only three physically meaningful variables (voltage, current, and temperature). Experimental validation on real-world driving data of a 2014 BMW i3 (60 Ah) over 70 trips under varied environmental conditions confirms that more input data does not systematically lead to better accuracy. The results highlight the importance of variable selection and the benefits of simplicity when key variables are well chosen.</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

Impact of Input Data Augmentation on Lithium-Ion Battery SOC Estimation: Comparison Between LSTM and GRU

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Mohamed Bendaoud,
  • Mohammed Bouafia,
  • Ayda Boukhancha,
  • Najib Essounbouli,
  • Amine El Fathi

摘要

This paper investigates the impact of input variable selection on the accuracy of lithium-ion battery state-of-charge estimation (SOCE). Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) algorithms were used for SOCE. Traditional methods use voltage, current, and temperature as inputs. In contrast, this work considers additional signals from the vehicle’s CAN bus including speed, engine torque, and heater power to determine whether a richer set of inputs enhances prediction performance. The approach applies a wavelet transform to remove signal noise and employs correlation analysis followed by principal component analysis (PCA) to lower dimensionality and improve signal quality by removing collinear variables. Two modeling strategies are compared: one using features selected by correlation analysis, and the other using only three physically meaningful variables (voltage, current, and temperature). Experimental validation on real-world driving data of a 2014 BMW i3 (60 Ah) over 70 trips under varied environmental conditions confirms that more input data does not systematically lead to better accuracy. The results highlight the importance of variable selection and the benefits of simplicity when key variables are well chosen.