Scholarlink

<p>The present work provides a detailed comparison of three deep learning models and their application for SOC: Deep Feedforward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models under varying thermal conditions. This study aims to understand the practicality of the three advanced neural network architectures in order to determine the most suitable architecture for SOC estimation in EVs in real time. All models were trained using the Adam optimizer and evaluated in terms of training effectiveness, error convergence, and prediction accuracy under varying thermal conditions. The obtained results demonstrate that temperature has a significant impact on the dynamics of the SoC and battery voltage leading to significant nonlinearity and estimation deviations. Among the architectures that were evaluated, the FFNN showed the shortest training time (32&#xa0;s) highlighting its suitability for lightweight applications. However, it shows poor robustness and high sensitivity to temperature changes. The highest maximum error (26.02%) and a modest RMSE (0.024), confirming the FFNN limited ability to capture sequential SoC dynamics. The GRU provided a good balance between accuracy and computational cost by achieving better convergence stability, increased precision, and smoother estimation behavior with a maximum error and RMSE which does not exceed respectively 8.42% and 0.0212. Across all test conditions, the LSTM network achieved the highest overall performance, with the lowest MSE (1.7700e-05), MAE (0.0032), RMSE (0.0042), and maximum error (1.86%), offering the most precise and thermally consistent estimation. Overall, LSTM stands out as the most powerful and accurate solution, which making it the most suitable candidate for real-time battery management systems in electric vehicles.</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

A comparative study of deep learning architectures for Li-ion battery SoC estimation under varying thermal conditions: Electric vehicle application

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Radhia Jebahi,
  • Nadia Chaker,
  • Helmi Aloui

摘要

The present work provides a detailed comparison of three deep learning models and their application for SOC: Deep Feedforward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) models under varying thermal conditions. This study aims to understand the practicality of the three advanced neural network architectures in order to determine the most suitable architecture for SOC estimation in EVs in real time. All models were trained using the Adam optimizer and evaluated in terms of training effectiveness, error convergence, and prediction accuracy under varying thermal conditions. The obtained results demonstrate that temperature has a significant impact on the dynamics of the SoC and battery voltage leading to significant nonlinearity and estimation deviations. Among the architectures that were evaluated, the FFNN showed the shortest training time (32 s) highlighting its suitability for lightweight applications. However, it shows poor robustness and high sensitivity to temperature changes. The highest maximum error (26.02%) and a modest RMSE (0.024), confirming the FFNN limited ability to capture sequential SoC dynamics. The GRU provided a good balance between accuracy and computational cost by achieving better convergence stability, increased precision, and smoother estimation behavior with a maximum error and RMSE which does not exceed respectively 8.42% and 0.0212. Across all test conditions, the LSTM network achieved the highest overall performance, with the lowest MSE (1.7700e-05), MAE (0.0032), RMSE (0.0042), and maximum error (1.86%), offering the most precise and thermally consistent estimation. Overall, LSTM stands out as the most powerful and accurate solution, which making it the most suitable candidate for real-time battery management systems in electric vehicles.