Data-driven lithium-ion battery capacity estimation from voltage relaxation

Jiangong Zhu, Yuan Huang, Michael Knapp, Xinhua Liu, Yixiu Wang, R. Bhushan Gopaluni, Yankai Cao, Michael Heere, Martin J. Mühlbauer, Liuda Mereacre, Haifeng Dai, Anatoliy Senyshyn, Xuezhe Wei, Helmut EhrenbergResearch 2021 · DOI

TL;DR

Researchers have developed a data-driven approach to estimate lithium-ion battery capacity by analyzing the "voltage relaxation" period—the brief rest phase after a full charge. By extracting just three statistical features from this curve, their machine learning models achieved estimation errors as low as 1.1%, even proving capable of generalizing across different battery chemistries (NCA to NCM) without retraining.

What Problem Does This Paper Address?

Knowing the exact capacity of a battery—its "State of Health" (SoH)—is vital for predicting driving range in electric vehicles (EVs) and ensuring safety. However, measuring capacity accurately usually requires a full discharge cycle, which rarely happens in real-world driving.

While researchers have tried using partial charge/discharge data or complex Equivalent Circuit Models (ECMs), these methods often struggle with high computational costs or sensitivity to varying environmental conditions. There is a significant need for a "non-invasive" method that can estimate health using easily obtainable data without requiring the battery's entire usage history.

Methods

The authors propose using the voltage relaxation curve—the way a battery's voltage settles after it reaches a full charge and the current stops.

  1. Feature Extraction: Instead of using the raw time-series data, they converted the relaxation curve into six statistical "fingerprints": Variance (VAR), Skewness (SKE), Maxima (MAX), Minima (MIN), Mean (MEAN), and Kurtosis (KUR).
  2. Feature Selection: Using XGBoost's F-score, they determined that VAR, SKE, and MAX were the most critical features, allowing them to simplify the model without losing accuracy.
  3. Machine Learning Models: They compared four distinct algorithms:
  4. ElasticNet: A regularized linear regression.
  5. XGBoost: A powerful gradient-boosted decision tree ensemble.
  6. Support Vector Regression (SVR): A kernel-based method excellent for non-linear relationships.
  7. Deep Neural Network (DNN): A multi-layer architecture designed to map complex inputs to outputs.

Datasets

The study utilized two massive, high-quality datasets generated by cycling commercial 18650-type cells in temperature-controlled chambers:

  • NCA Dataset: 27,330 data units from LiNi$_{0.86}$Co$_{0.11}$Al$_{0.03}$O$_2$ batteries cycled at 25°C, 35°C, and 45°C with varying charge rates.
  • NCM Dataset: 30,312 data units from LiNi$_{0.83}$Co$_{0.11}$Mn$_{0.07}$O$_2$ batteries, used primarily to test how well the models generalize to different chemistries.

Key Insights and Ideas

  • The "Cool-Down" Fingerprint: The core "aha" moment is that the thermodynamic and kinetic changes caused by battery aging are reflected in the relaxation curve. As a battery degrades, the way its voltage drops during rest changes in a predictable way.
  • Variance is King: Statistical analysis showed that Variance (VAR) has the highest priority for estimation. A decrease in VAR corresponds to a sharper distribution of relaxation voltages as the battery ages.
  • Chemistry Agnosticism: Perhaps the most surprising finding is that a model trained on NCA batteries can perform remarkably well on NCM batteries. This suggests that the underlying physics of voltage relaxation are consistent across different nickel-rich cathode chemistries.
  • Non-Linearity Wins: While linear models (ElasticNet) performed decently, non-linear models like XGBoost and DNN were significantly better at capturing the nuances of battery degradation.

Key Results

  • High Accuracy: On the NCA test set, both XGBoost and SVR achieved a Root-Mean-Square Error (RMSE) of just 1.1%.
  • Generalizability: When the model trained on NCA was applied to NCM batteries without changing any weights, the DNN performed best with an RMSE of 1.8%.
  • Efficiency: The researchers demonstrated that using only three statistical features (VAR, SKE, MAX) is sufficient to reach these high accuracy levels, making the approach computationally efficient enough for on-board vehicle systems.

Limitations and Future Work

While the results are promising, the authors and the context of the paper suggest a few areas for further exploration:

  • Rest Time Requirements: The study used a 30-minute rest period. In real-world EV scenarios, rest periods can vary wildly. Future work could investigate how shorter relaxation times (e.g., 5-10 minutes) impact accuracy.
  • State of Charge (SoC) Sensitivity: This study focused on relaxation after a full charge (100% SoC). Since many EV users don't always charge to 100%, adapting the model for partial charge states would be a logical next step.
  • Real-World Noise: The data was collected in controlled laboratory conditions. Testing the robustness of these statistical features against the "noisy" voltage measurements found in actual vehicle hardware remains a hurdle for deployment.

Summary updated

Abstract

Abstract Accurate capacity estimation is critical for reliable and safe operation of lithium-ion batteries. A proposed approach exploiting features from the relaxation voltage curve enables battery capacity estimation without requiring previous cycling information. Machine learning methods are used in the approach. A dataset including 27,330 data units are collected from batteries with LiNi 0.86 Co 0.11 Al 0.03 O 2 cathode (NCA battery) cycled at different temperatures and currents until reaching about 71% of their nominal capacity. One data unit comprises three statistical features (variance, skewness, and maxima) derived from the relaxation voltage curve after fully charging and the following discharge capacity for verification. Models adopting machine learning methods, i.e., ElasticNet, XGBoost, Support Vector Regression (SVR), and Deep Neural Network (DNN), are compared to estimate the battery capacity. Both XGBoost and SVR methods show good predictive ability with 1.1 % root-mean-square error (RMSE). The DNN method presents a 1.5% RMSE higher than that obtained using ElasticNet and SVR. 30,312 data units are extracted from batteries with LiNi 0.83 Co 0.11 Mn 0.07 O 2 cathode (NCM battery). The model trained by the NCA battery dataset is verified on the NCM battery dataset without changing model weights. The test RMSE is 3.1% for the XGBoost method and 1.8% RMSE for the DNN method, indicating the generalizability of the capacity estimation approach utilizing battery voltage relaxation.
BibTeX
@inproceedings{zhu2021data,
  author = {Jiangong Zhu and Yuan Huang and Michael Knapp and Xinhua Liu and Yixiu Wang and R. Bhushan Gopaluni and Yankai Cao and Michael Heere and Martin J. Mühlbauer and Liuda Mereacre and Haifeng Dai and Anatoliy Senyshyn and Xuezhe Wei and Helmut Ehrenberg},
  title = {{Data-driven lithium-ion battery capacity estimation from voltage relaxation}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-770709/v1},
}

Loading related content...

Figure