What question did this study set out to answer?

February 22, 2026Open Access

Enhanced Adaptive LSTM Framework for Robust Early-Life Prediction of Lithium-Ion Battery Degradation

Key Points

The study aims to improve early-life prediction of lithium-ion battery degradation using an enhanced LSTM framework.
Developed the Enhanced Adaptive LSTM (EA-LSTM) framework for prediction.
Utilized multicycle statistical features from the first 100 cycles for degradation behavior.
Trained using data from 118 cells and tested on 32 fast-charged cells.
Evaluated performance against baseline models like ANN and XGBoost.
Achieved 10% reduction in end-of-life mean absolute percentage error (MAPE) and 15% in root mean square error (RMSE).
Demonstrated <2% generalization gap across different datasets.
Maintained high accuracy while remaining lightweight with approximately 3 million floating point operations.

Abstract

• A multicycle feature pipeline captures early-cycle degradation behavior. • EA-LSTM uses adaptive gating and attention to enhance prognostic sensitivity. • KO, KP, and EoL are predicted jointly using only the first 100 cycles. • Strong cross-domain robustness with <2% generalization gap across datasets. • Lightweight design enables real-time deployment in embedded BMS systems. The early life prediction of lithium-ion battery degradation remains challenging because the first 100 charge–discharge cycles exhibit minimal observable aging while containing subtle precursors of long-term behavior. This study introduces a novel Enhanced Adaptive LSTM (EA-LSTM) framework that directly predicts the knee onset (KO), knee point (KP), and end-of-life (EoL) using only early cycle multicycle statistical features. The proposed model integrates adaptive gating with attention-based modulation to selectively amplify the prognostic signals embedded in sparse early life data. Trained on 118 cells from the MIT MaTR dataset and tested on 32 fast-charged cells from the CLO dataset, the EA-LSTM achieves consistently superior accuracy compared with the baseline ANN, XGBoost, Vanilla LSTM, and Transformer baselines, reducing the EoL MAPE by ∼10% and RMSE by ∼15%. External validation of 32 fast-charging cells from the CLO dataset confirmed strong domain transferability, with generalisation gaps below 2%. Additional evaluation on the Mendeley dataset, characterised by stochastic discharge profiles and ultra-long-life trajectories, shows that while zero-shot transfer degrades performance, a unified training strategy successfully recovers the predictive fidelity ( ≈ 13 % MAPE). The correlation-ranked Top-40 feature subset further enhances the accuracy while keeping the framework lightweight ( ∼ 3 × 10 6 FLOPs). These findings establish the EA-LSTM as a compact, data-efficient, and robust early life prognostic framework capable of capturing degradation trajectories long before measurable capacity loss occurs.

Read Full Paperexternally

Mark Helpful

Bookmark

Relay

View Full Paper