| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | 2 | 3 | 4 | 5 | 6 | |
| 7 | 8 | 9 | 10 | 11 | 12 | 13 |
| 14 | 15 | 16 | 17 | 18 | 19 | 20 |
| 21 | 22 | 23 | 24 | 25 | 26 | 27 |
| 28 | 29 | 30 |
- bms
- state of health
- 딥러닝 코드
- 코이딥
- 배터리 모델링
- Battery modeling
- 배터리 EIS
- 딥러닝
- 배터리 진단
- Azure
- AzureML
- 배터리 열화
- Battery Deep Learning
- 리튬배터리
- Machine Learning
- 머신러닝
- Deep learning
- 칼만필터
- 배터리 딥러닝
- eis
- 코드로 이해하는 딥러닝
- Battery SOH
- 배터리 연구
- Battery Management System
- 배터리 AI
- tensorflow
- Incremental Capacity Analysis
- 텐서플로우
- Battery AI
- 머신러닝 코드
- Today
- Total
Engineering insight
[Digital Discovery-2023] Uncertainty-aware and explainable machine learning for early prediction of battery degradation trajectory 본문
[Digital Discovery-2023] Uncertainty-aware and explainable machine learning for early prediction of battery degradation trajectory
Free-Nomad 2026. 4. 28. 00:35Uncertainty-aware and explainable machine learning for early prediction of battery degradation trajectory
DOI: https://doi.org/10.1039/d2dd00067a
HTML 본문: https://pubs.rsc.org/en/content/articlehtml/2023/dd/d2dd00067a
Uncertainty-aware and explainable machine learning for early prediction of battery degradation trajectory - Digital Discovery
Abstract Enhancing cell lifetime is a vital criterion in battery design and development. Because lifetime evaluation requires prolonged cycling experiments, early prediction of cell aging can significantly accelerate both the autonomous discovery of better
pubs.rsc.org
이 논문은 초기 cycle 데이터만으로 배터리의 EOL 숫자 하나가 아니라 전체 degradation trajectory를 uncertainty까지 포함해 예측하는 explainable LSTM 연구입니다.
1. 이 논문이 푼 문제
배터리 수명예측의 실전 문제는 단순히 언제 죽는지(EOL) 맞히는 것이 아니라, 초기에 얻을 수 있는 제한된 데이터만으로 앞으로 capacity fade가 어떤 경로를 따라 진행될지 읽어내는 것입니다. 실험 시간이 길고 비용이 크기 때문에 early-cycle signal만으로 long-term fate를 판단할 수 있다면 실험 설계, fast-charging screening, 개발 속도 모두 크게 개선됩니다.
초기 수십~수백 cycle에서 얻은 전기화학 signal만으로 future capacity trajectory 전체와 EOL를 uncertainty-aware하게 예측할 수 있는가?
2. 입력/출력: 이 battery-AI 논문의 실질적 핵심
입력
- 초기 capacity trajectory
- charging-rate statistics
- coulombic efficiency
- charge-discharge voltage gap
- differential-capacity 기반 covariate
출력
- 다음 cycle capacity의 mean / variance
- autoregressive rollout으로 생성한 full degradation trajectory
- 최종 EOL 추정
즉 이 논문은 입력도, 출력도 모두 실무적으로 중요합니다. 입력은 early-cycle electrochemical covariate이고, 출력은 점 하나가 아니라 trajectory 전체입니다.
Figure 1 — 문제 정의와 모델 프레임

무엇을 보여주나: 초기 cycle 데이터에서 long-term degradation trajectory를 예측하는 문제 설정과 모델 흐름을 요약합니다.
왜 중요한가: 이 논문이 EOL point prediction이 아니라 trajectory forecasting 문제를 푼다는 점을 가장 분명히 보여줍니다.
하지만 중요한 문제점이 있습니다.
LSTM의 입력으로 다양한 Covariate를 넣고, 이 출력으론 SOH(Mean)과 Uncertainty(Variance)입니다.
Variance( σ )는 LSTM을 학습할때 2σ2(Q−μ)^2+logσ 이런식으로 Loss Fucntion을 구성해서 도출하게됩니다.
그런 Covariate Input -> LSTM -> mean, variance가 출력이되는데요
Autorgressive는 결국 이 LSTM을 동일하게 재귀하는것으로 사용합니다.
그런데 보면 LSTM의 mean, variance만으로는 다시 입력으로 넣어서 Autoregressive를 할 수 없는구조입니다.
이게 이 논문의 한계입니다.
즉, 1~10Cycle의 Covariate데이터로 LSTM의 입력을했다면 11Cycle째의 SOH Mean, Variance를 출력해내고, 그다음 2~11Cycle로 12번째 Cycle의 SOH를 추정해야될텐데요
이때 11Cycle째의 데이터는 SOH밖에없고, Covariate데이터가 없습니다.
그렇기때문에 1~10Cycle째의 Covariate를 그대로 사용하고 SOH/Cycle수면 LSTM의 출력값으로 대응하여 Autoregressive합니다.
이말인 즉슨, 입력의 초기값 Covariate에만 의존하여 Autoregressive하며, 결국 뒷부분은 입력의 SOH밖에 바뀌지않는채로 계속 미래값을 RUL하게 되는것입니다.
3. 이 논문만의 novelty
- Point prediction이 아니라 full trajectory prediction
- mean+variance 출력으로 uncertainty-aware prediction
- LSTM ensemble 기반의 autoregressive rollout
- saliency 분석으로 왜 그런 예측이 나왔는지 해석 시도
핵심은 단순히 더 좋은 예측기가 아니라, early signal과 future degradation path를 연결하는 구조를 만들고, 그 예측의 신뢰도와 feature importance까지 같이 보려 했다는 점입니다.
Figure 2 — trajectory prediction 성능의 핵심 그림

무엇을 보여주나: 모델이 실제 capacity fade trajectory를 어떻게 따라가는지와 uncertainty band를 함께 보여줍니다.
왜 중요한가: 이 논문의 강점이 단일 EOL 숫자보다 trajectory 전체를 예측한다는 데 있음을 보여줍니다.
4. feature와 그 화학적 의미
사용자께서 중요하게 보시는 부분은 바로 이 대목입니다. 이 논문은 covariate engineering이 성능에 결정적임을 보여줍니다.
- Coulombic efficiency: lithium loss, side reaction, parasitic consumption과 연결
- Charge-discharge voltage gap: polarization, resistance growth, kinetic limitation과 연결
- Differential-capacity feature: phase-transition / redox feature 이동, active material utilization 변화와 연결
즉 이 논문은 단순 raw sequence만 넣지 않고, 열화를 더 직접적으로 반영하는 전기화학 feature를 모델에 먹인 것이 핵심입니다.
Figure 5 — covariate 설계가 왜 중요한가

무엇을 보여주나: covariate 제거/포함에 따른 성능 차이와 feature engineering의 중요성을 보여줍니다.
왜 중요한가: covariate를 빼면 EOL RMSE가 587 cycles까지 악화되어, 성능의 핵심이 모델 깊이만이 아니라 electrochemical representation이라는 점을 보여줍니다.
Figure 7 / 8 / 9 — saliency와 해석 가능성

무엇을 보여주나: 예측에 어떤 feature가 중요한지 saliency로 분석합니다.
왜 중요한가: short-lived cell에서 coulombic efficiency 중요도가 더 커지고, knee 전후에는 charging condition과 electrochemical covariate 중요도가 커지는 양상이 보입니다.
화학적 의미: 이는 lithium inventory loss, side reaction, kinetic deterioration이 서로 다른 수명 영역에서 다르게 작동한다는 해석과 정렬됩니다.
5. 성능 수치와 읽는 법
- 100 cycle 입력 기준 non-aged cell에서 EOL RMSE 106 cycles
- MAPE 10.6%
- covariate 제거 시 RMSE가 587 cycles까지 악화
이 수치가 의미하는 바는 분명합니다. 이 논문의 성능은 LSTM이라는 이름보다도 어떤 전기화학 covariate를 어떻게 설계했는가에 더 크게 의존합니다.
6. 실무적 의미
- early-cycle screening
- fast-charging protocol ranking
- 장기 cycling 시험 조기 종료
- active-learning형 배터리 개발 파이프라인
특히 uncertainty를 같이 출력한다는 점은 “지금 예측을 믿고 실험을 멈춰도 되는가?” 같은 실제 운영 판단에 중요합니다.
7. 한계
- 데이터셋이 작고 LFP 중심
- 실제 복잡한 운전조건 generalization은 아직 충분히 검증되지 않음
- uncertainty calibration도 완전하지 않음
- 다른 chemistry / protocol / temperature variation에 대한 범용성은 추가 검증 필요
Final Summary
Core
초기 cycle 데이터만으로 배터리의 전체 열화 trajectory와 EOL를 uncertainty까지 포함해 예측한 explainable LSTM 논문입니다.
Novelty
- EOL 숫자 하나가 아니라 full degradation trajectory를 autoregressive하게 생성
- mean+variance 출력과 ensemble로 uncertainty-aware prediction 구현
- saliency 분석으로 coulombic efficiency / voltage-gap / differential-capacity feature의 화학적 의미까지 연결
Practical takeaways
- early-cycle screening, fast-charging ranking, 실험 조기 종료에 유용
- 잘 설계된 electrochemical covariate가 black-box보다 더 실용적일 수 있음
Main limitations
작은 LFP 중심 데이터셋, 제한적 generalization, 불완전한 uncertainty calibration.