A data-driven architecture fusing Kolmogorov-Arnold feature extraction and contextual-attention long short-term memory network for accurate state-of-charge estimation in lithium-ion batteries under dynamic operating conditions
Syed Abbas Ali Shah , Syed Maooz Ali Shah , Shunli Wang , Mahrukh , Shungang Ning , Ziqiang Xu , Mengqiang Wu
{"title":"A data-driven architecture fusing Kolmogorov-Arnold feature extraction and contextual-attention long short-term memory network for accurate state-of-charge estimation in lithium-ion batteries under dynamic operating conditions","authors":"Syed Abbas Ali Shah , Syed Maooz Ali Shah , Shunli Wang , Mahrukh , Shungang Ning , Ziqiang Xu , Mengqiang Wu","doi":"10.1016/j.est.2025.118819","DOIUrl":null,"url":null,"abstract":"<div><div>Accurate estimation of lithium-ion battery state of charge (SOC) under dynamic conditions remains challenging. This study introduces a hybrid architecture that combines a Kolmogorov-Arnold Network (KAN) for theoretical nonlinear decomposition of battery signals, a feature-wise contextual attention (FCA) mechanism for adaptive channel weighting, and stacked long short-term memory (LSTM) layers for temporal modeling, collectively termed KALSTM framework. The KAN front-end leverages the Kolmogorov–Arnold representation theorem to factorize current, voltage, and temperature into independent univariate mappings, recombining them through a single affine mixer to achieve universal approximation with far fewer parameters than a dense input layer. FCA replaces quadratic self-attention with a context-conditioned probability distribution over latent channels at each timestep, ensuring linear complexity in sequence length when feature dimension is fixed while selectively amplifying salient sensor information. The LSTM stack captures multiscale temporal dependencies within these refined embeddings to produce a single SOC estimate. Training was conducted on a subset of a publicly available battery dataset comprising urban-to-highway drive cycles collected at 0 °C, 10 °C, and 25 °C. Validation utilized the remaining records, including both fixed-temperature cycles and continuously varying thermal conditions, along with data from a different battery chemistry to evaluate generalization. The proposed KALSTM model achieved optimal accuracy, attaining an RMSE of 0.77 % and MAE of 0.63 % under fixed temperatures, and RMSE of 1.10 % and MAE of 0.89 % during dynamic thermal variations. It consistently outperformed a parameter-matched LSTM baseline and recent literature benchmarks. These results highlight its potential as a reliable and transferable tool for advanced battery state estimation.</div></div>","PeriodicalId":15942,"journal":{"name":"Journal of energy storage","volume":"139 ","pages":"Article 118819"},"PeriodicalIF":8.9000,"publicationDate":"2025-10-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of energy storage","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2352152X25035327","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENERGY & FUELS","Score":null,"Total":0}
引用次数: 0
Abstract
Accurate estimation of lithium-ion battery state of charge (SOC) under dynamic conditions remains challenging. This study introduces a hybrid architecture that combines a Kolmogorov-Arnold Network (KAN) for theoretical nonlinear decomposition of battery signals, a feature-wise contextual attention (FCA) mechanism for adaptive channel weighting, and stacked long short-term memory (LSTM) layers for temporal modeling, collectively termed KALSTM framework. The KAN front-end leverages the Kolmogorov–Arnold representation theorem to factorize current, voltage, and temperature into independent univariate mappings, recombining them through a single affine mixer to achieve universal approximation with far fewer parameters than a dense input layer. FCA replaces quadratic self-attention with a context-conditioned probability distribution over latent channels at each timestep, ensuring linear complexity in sequence length when feature dimension is fixed while selectively amplifying salient sensor information. The LSTM stack captures multiscale temporal dependencies within these refined embeddings to produce a single SOC estimate. Training was conducted on a subset of a publicly available battery dataset comprising urban-to-highway drive cycles collected at 0 °C, 10 °C, and 25 °C. Validation utilized the remaining records, including both fixed-temperature cycles and continuously varying thermal conditions, along with data from a different battery chemistry to evaluate generalization. The proposed KALSTM model achieved optimal accuracy, attaining an RMSE of 0.77 % and MAE of 0.63 % under fixed temperatures, and RMSE of 1.10 % and MAE of 0.89 % during dynamic thermal variations. It consistently outperformed a parameter-matched LSTM baseline and recent literature benchmarks. These results highlight its potential as a reliable and transferable tool for advanced battery state estimation.
期刊介绍:
Journal of energy storage focusses on all aspects of energy storage, in particular systems integration, electric grid integration, modelling and analysis, novel energy storage technologies, sizing and management strategies, business models for operation of storage systems and energy storage developments worldwide.