Energy, Volume 360, 30 September 2026, DOI: https://doi.org/10.1016/j.energy.2026.141446

An Attentional Deep Learning Model Based on Improved Crested Porcupine Optimizer for State of Health Prediction of Lithium-Ion Batteries

Simin Peng¹, Daohan Zhang^1,2, Yuxia Jiang¹, Lin Wang¹, Yonggang Liu³, and Michael Pecht^4
1School of Electrical Engineering, Yancheng Institute of Technology, Yancheng, Yancheng, 224051, China
²Department of Control Science and Engineering, Jilin University, Changchun, 130012, China
³State Key Laboratory of Mechanical Transmission for Advanced Equipment, Chongqing University, Chongqing, 400000, China
⁴Center for Advanced Life Cycle Engineering (CALCE), University of Maryland, College Park, 20742, USA

For more information about this article and related research, please contact Prof. Michael Pecht.

Abstract:

This paper aims to accurately predict and effectively manage the state of health (SOH) of lithium-ion batteries to maximize their utilization. Traditional SOH prediction models based on deep learning frequently encounter issues like inadequate representation capabilities of health features (HFs) and improper hyper-parameter settings. An improved deep learning approach is developed to predict the SOH of lithium-ion batteries, leveraging a channel attention mechanism and improved Crested Porcupine Optimizer (ICPO). Firstly, a channel attention mechanism is incorporated into convolutional neural networks (CNNs) to emphasize crucial feature information, regardless of input conditions, enhancing the representation capabilities of HFs. Additionally, to overcome the shortcoming of recurrent neural network (RNN)-based single models, which ignore long-term dependencies between battery aging and HFs, a deep learning method based on the combination of gated recurrent unit (GRU) and Long Short-Term Memory (LSTM) is developed. Furthermore, an ICPO is integrated into the model to optimize hyper-parameters for SOH prediction. Two laboratory datasets and an on-road vehicle dataset are used to validate the developed method, with all SOH prediction errors of no more than 2.2% for laboratory datasets and root mean square errors of no more than 1.0% for on-road vehicle dataset, outperforming other methods such as LSTM and its improved variants in terms of SOH prediction accuracy. These results indicate that the developed method can not only address the phenomenon of battery capacity regeneration but also delivers highly accurate and robust SOH estimation in real-world scenarios.

This article is available for free online here until July 21.