Scholarlink

<p>Accurately predicting the State of Health (SoH) of Lithium-ion Batteries (LIBs) is critical for improving Electric Vehicle (EV) performance and enabling smarter charging strategies, yet existing methods often struggle with variable cycle data and the limitations of physical models in capturing the complex behaviour of LIBs. To overcome these challenges, the present work introduces a hybrid deep learning framework, the Optuna-Optimised Convolutional Neural Network and Stacked Long Short-Term Memory (OptiCNN-SLSTM) model, which combines a 1D CNN for temporal feature compression with a stateful LSTM to capture both inter- and intra-sample dependencies. A two-step hyperparameter optimisation process is applied, first exploring different parameter sets and then refining them through Optuna optimisation. The model was trained and tested on the NASA and Oxford datasets and validated on a real-world EV E-kart battery dataset, achieving strong accuracy with RMSE values of 2.15% (B0005) and 2.18% (B0007) on NASA, 0.67% (cell-1) and 1.02% (cell-2) on Oxford, and exceptionally low errors on the E-kart dataset (MSE = 0.001%, RMSE = 0.03%, MAE = 0.28%), demonstrating its robustness and generalisability for practical EV battery health estimation.</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

Explainable OptiCNN-SLSTM hybrid model for enhanced lithium-ion battery state of health prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月29日

查看全文
  • Ashima Singh,
  • Sanju Singh,
  • Rahul Kumar Kamboj,
  • Anju Bala

摘要

Accurately predicting the State of Health (SoH) of Lithium-ion Batteries (LIBs) is critical for improving Electric Vehicle (EV) performance and enabling smarter charging strategies, yet existing methods often struggle with variable cycle data and the limitations of physical models in capturing the complex behaviour of LIBs. To overcome these challenges, the present work introduces a hybrid deep learning framework, the Optuna-Optimised Convolutional Neural Network and Stacked Long Short-Term Memory (OptiCNN-SLSTM) model, which combines a 1D CNN for temporal feature compression with a stateful LSTM to capture both inter- and intra-sample dependencies. A two-step hyperparameter optimisation process is applied, first exploring different parameter sets and then refining them through Optuna optimisation. The model was trained and tested on the NASA and Oxford datasets and validated on a real-world EV E-kart battery dataset, achieving strong accuracy with RMSE values of 2.15% (B0005) and 2.18% (B0007) on NASA, 0.67% (cell-1) and 1.02% (cell-2) on Oxford, and exceptionally low errors on the E-kart dataset (MSE = 0.001%, RMSE = 0.03%, MAE = 0.28%), demonstrating its robustness and generalisability for practical EV battery health estimation.