Researchers have developed a feature-enhanced ensemble learning method for rapidly estimating the capacity of lithium-ion batteries using only a short partial discharge segment from the initial stage of testing. The approach is designed to support battery capacity grading in industrial settings, where conventional full-discharge procedures can take several hours per cell and create a bottleneck for manufacturing, pack assembly, and second-life screening.
Accurate capacity estimation is a basic but important requirement for lithium-ion battery management. In production and reuse scenarios, cells need to be graded and matched so that battery packs can operate safely and reliably. However, traditional capacity testing typically relies on complete charge-discharge cycles. That provides useful information, but it also consumes time, energy, and testing resources. For high-throughput battery manufacturing or rapid second-life qualification, waiting hours for each full-cycle measurement can limit scalability and increase operational costs.
The new study addresses this problem by focusing on the early voltage response during the first discharge cycle. According to the article, the method uses voltage measurements from only the first 300-480 seconds of the initial discharge cycle to estimate nominal capacity. Although real production-line grading data are often difficult to access, the authors note that this first-cycle setup can serve as a controlled surrogate that captures key aspects of factory-based capacity labeling.
A central idea in the method is that the early discharge curve is not merely a short fragment of incomplete data. It can contain electrochemical signatures related to battery condition, including effects associated with lithium inventory loss, solid-electrolyte interphase evolution, charge-transfer resistance, and ion transport dynamics. To make those signatures more useful for prediction, the researchers extract physically interpretable health indicators from the short voltage window and emphasize information contained in second-order voltage derivatives.
The study then applies a nonlinear feature enhancement strategy to strengthen subtle capacity-related patterns while suppressing variability associated with manufacturing differences. This step is important because small differences in early voltage behavior may be difficult for a model to use directly, especially across cells with different chemistries or aging stages. By enhancing the features before prediction, the framework aims to make the useful signal more visible to the learning algorithm without requiring a complete cycle.
These engineered features are fed into a Multi-Decision Ensemble Learning architecture, or MDEL. Rather than relying on a single regression pathway, the ensemble design adaptively combines multiple decision routes to improve robustness. In practical terms, this is intended to help the model maintain accuracy across diverse capacity conditions, from fresh cells to end-of-life cells, while still using only a short initial discharge segment.
Experimental results reported in the paper suggest that the proposed approach can achieve capacity-estimation performance comparable to methods that use complete-cycle data. The method was evaluated on in-lab cells as well as the public CALCE and MIT datasets, which span different battery conditions and aging stages. The authors report a mean absolute error of no more than 0.0391 Ah, corresponding to no more than 1.63% of nominal capacity, while reducing testing time by more than 80%.
For battery manufacturing and reuse, that time reduction could be highly meaningful. A method that estimates capacity in minutes rather than hours may help improve high-throughput cell grading, support more precise pack matching, and accelerate the evaluation of used cells for second-life applications. It may also reduce the energy and equipment burden associated with repeated full-cycle testing, which is increasingly relevant as lithium-ion battery production and recycling ecosystems continue to expand.
Further validation will still be needed before such a method can be treated as a universal replacement for full-cycle testing in every production environment. Real factories may involve wider variations in cell format, chemistry, temperature, sensor quality, and operating procedures than benchmark datasets can fully represent. Even so, the study offers a strong indication that early-stage voltage information, when paired with physically informed feature extraction and ensemble learning, could make rapid lithium-ion battery capacity grading more practical and scalable.
Reference
Author:
Ziheng Zhou a, Chaolong Zhang a, Shi Chen a, Yan Zhang a, Lei Wang b
Title of original paper:
Feature-enhanced ensemble learning for accurate capacity estimation of lithium-ion batteries using partial discharging segments in initial stage based on second-order voltage derivatives
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725001380
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100388
Affiliations:
a College of Intelligent Science and Control Engineering, Jinling Institute of Technology, Nanjing 211169, China
b School of Electrical and Computer Engineering, Oklahoma State University, Stillwater, OK 74078, USA
Green Energy and Intelligent Transportation
Experimental study
Not applicable
Feature-enhanced ensemble learning for accurate capacity estimation of lithium-ion batteries using partial discharging segments in initial stage based on second-order voltage derivatives
28-Jan-2026