Early prediction of lithium-ion battery degradation with a generative pre-trained transformer

Sep 30, 2026

Upload by Ophelia

Early detection of degradation in lithium-ion batteries (LIBs) is essential for predictive maintenance and recycling, yet accurate forecasting remains difficult because early charging cycles show barely noticeable changes and degradation follows a long-term nonlinear trajectory. The study introduces BatteryGPT, a two-stage early-stage degradation prediction framework using a generative pre-trained transformer (GPT) to autoregressively predict full-lifecycle charging data, combined with a state-of-health (SOH) estimator to link predicted charging behavior with ageing features. Results show accurate future degradation prediction before capacity decay is evident, with strong performance even using only the first 30% or 5% of lifetime charging data.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Introduction

Early prediction motivation and challenges

Ageing-feature prediction tasks (SOH variation, knee point, EOL)

Method concept: BatteryGPT

Two-stage framework and GPT-based charging-data prediction

SOH estimation linking predicted charging data to ageing features

Article [https://doi.org/10.1038/s41467-025-66819-0](https://doi.org/10.1038/s41467-025-66819-0) Early prediction of lithium-ion battery degradation with a generative pre-trained transformer Received: 17 April 2024 | Accepted: 14 November 2025 | | --- | | | | Check for updates | Jincheng Hu 1,2,5, Pengyu Fu 1,2,5, Zhongbao Wei 3, Yanjun Huang 1 , Juliana Early4, Ashley Fly2 & Yuanjian Zhang 1,2 The early detection of degradation in lithium-ion batteries (LIBs) is crucial for effective predictive maintenance and recycling. However, accurately predicting the future degradation of LIBsin early stage is challenging duetothe barely noticeable performance changes at initial charging cycles and the long-term nonlinear degradation pattern. In this work, we propose a two-stage earlystage degradation prediction method, BatteryGPT, which employs a Generative Pre-trained Transformer (GPT) to autoregressively predict the charging data of entire lifecycle and a state-of-health (SOH) estimator to correlates the predicted charging data with ageing features in LIBs. The validation demonstrates that BatteryGPT can predict the future LIB degradation with high accuracy using early charging data, before any capacity degradation is evident. Predicting with the ﬁrst 30% of the battery lifetime, BatteryGPT signiﬁcantly outperforms baselines, achieving a root mean square error (RMSE) of 0.213% for SOH variation prediction, and mean absolute percent errors (MAPE) of 2.30% and 1.18% for knee point and EOL predictions. Even predicting with the ﬁrst 5% of lifetime charging data, BatteryGPT demonstrates strong early-stage prediction performance. The extensive application of lithium-ion batteries (LIBs) in storing power generated from sustainable energy sources, such as solar and wind, profoundly impacts the ﬁeld of energy research and methodologies1. This innovation has paved the way for envisioning an energy supply free from fossil fuels, representing a signiﬁcant stride towards net-zero emissions2. However, the performance of LIBs gradually declines over long-term usages due to various factors. For instance, the number of charge-discharge cycles, temperature variations, and efﬁciency of battery management system (BMS) all impact LIB performance. This negative shift is a long-term feedback delay, primarily manifested in the reduced capacity and increased internal resistance, inducing potential safety risks in emerging sustainable energy systems3. If LIB degradation could be detected in the early stages of the lifecycle, sustainable energy systems would enable proactive measures to mitigate these risks and enhance reliability and safety of power usages. Therefore, predicting future LIB degradation based on the initial charging cycles presents opportunities for the safe use and optimization of sustainable energy4. Accurately predicting performance degradation in advance introduces a set of ageing-feature prediction tasks that are crucial to fully unlock the potential of LIBs. These challenges include predicting the state-of-health (SOH) variations5, identifying the knee point6, and forecasting the end-of-life (EOL) time7. SOH describes the current performance of a LIB as a proportion of its initial performance, commonly deﬁned as the ratio of current capacity to initial capacity3. Predicting the SOH variations involves analyzing and forecasting 1School of Automotive Studies, Tongji University, Shanghai, China. 2Department of Aeronautical and Automotive Engineering, Loughborough University, Loughborough, United Kingdom. 3National Engineering Research Centre for Electric Vehicles, Beijing Institute of Technology, Beijing, China. 4School of Mechanical and Aerospace Engineering, Queen’s University Belfast, Northern Ireland, UK. 5These authors contributed equally: Jincheng Hu, Pengyu Fu. e-mail: [yanjun_huang@tongji.edu.cn](yanjun_huang@tongji.edu.cn); [Y.Y.Zhang@lboro.ac.uk](Y.Y.Zhang@lboro.ac.uk) characteristics changes in battery ageing

FAQs

Why is early prediction of LIB degradation challenging?

Early charging cycles contain performance changes that are barely noticeable, while degradation over time follows a long-term nonlinear pattern. These factors reduce the effectiveness of both model-based and data-driven approaches.

What is BatteryGPT and how does it work?

BatteryGPT is a two-stage method that uses a generative pre-trained transformer (GPT) to autoregressively predict charging data across the battery’s entire lifecycle. A state-of-health (SOH) estimator then correlates the predicted charging data with LIB ageing features.

How early can BatteryGPT predict degradation, and how accurate is it?

When predicting using the first 30% of battery lifetime, BatteryGPT outperforms baselines and achieves RMSE of 0.213% for SOH variation prediction, with MAPE of 2.30% for knee point and 1.18% for end-of-life (EOL) predictions. It also shows strong early-stage performance when using only the first 5% of lifetime charging data.

Sep 30, 2026

Upload by Ophelia