Scholarlink

<p>Estimating the remaining useful life (RUL) of lithium-ion batteries is critical for early detection of battery failures in electric vehicles, optimization of maintenance planning, and improvement of system reliability. However, accurate RUL estimation is a challenging problem due to the nonlinear dynamics of battery aging processes, capacity regeneration phenomena, variable operating conditions, and limited measurable signals. Current deep learning methods have significant limitations, such as inadequate modeling of multi-scale temporal patterns, lack of phase-specific adaptive weighting, and insufficient search strategies in hyperparameter optimization. In this paper, a multi-branch convolutional neural network-attention (MB-CNN + Attention) architecture is proposed to simultaneously capture both short-term fluctuations and long-term trends in battery aging signals. The proposed architecture consists of three parallel convolution branches with different kernel sizes, a progressively dilation convolution strategy, a multi-head self-attention mechanism, a temporal convolutional network (TCN) block, and regression layers. The architecture’s 14 hyperparameters were optimized using Rüppell’s Fox Optimization (RFO) algorithm, which was introduced to the literature in 2025. RFO optimization increased the coefficient of determination (R²) by 1.42% and reduced the root mean square error (RMSE) by 15.58% compared to manual tuning. In a comprehensive evaluation performed on nine battery cells from the Hawaii Natural Energy Institute (HNEI) dataset, the proposed RFO-optimized MB-CNN + Attention method demonstrated RMSE improvements of 17.4%, 16.6%, and 13.7% compared to the RFO-optimized Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models, respectively.</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

RFO-Optimized multi-branch CNN-attention architecture for lithium-ion battery remaining useful life prediction

  • 期刊论文

  • 发表日期：

    2026年3月26日

查看全文
  • Timur Lale

摘要

Estimating the remaining useful life (RUL) of lithium-ion batteries is critical for early detection of battery failures in electric vehicles, optimization of maintenance planning, and improvement of system reliability. However, accurate RUL estimation is a challenging problem due to the nonlinear dynamics of battery aging processes, capacity regeneration phenomena, variable operating conditions, and limited measurable signals. Current deep learning methods have significant limitations, such as inadequate modeling of multi-scale temporal patterns, lack of phase-specific adaptive weighting, and insufficient search strategies in hyperparameter optimization. In this paper, a multi-branch convolutional neural network-attention (MB-CNN + Attention) architecture is proposed to simultaneously capture both short-term fluctuations and long-term trends in battery aging signals. The proposed architecture consists of three parallel convolution branches with different kernel sizes, a progressively dilation convolution strategy, a multi-head self-attention mechanism, a temporal convolutional network (TCN) block, and regression layers. The architecture’s 14 hyperparameters were optimized using Rüppell’s Fox Optimization (RFO) algorithm, which was introduced to the literature in 2025. RFO optimization increased the coefficient of determination (R²) by 1.42% and reduced the root mean square error (RMSE) by 15.58% compared to manual tuning. In a comprehensive evaluation performed on nine battery cells from the Hawaii Natural Energy Institute (HNEI) dataset, the proposed RFO-optimized MB-CNN + Attention method demonstrated RMSE improvements of 17.4%, 16.6%, and 13.7% compared to the RFO-optimized Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models, respectively.