Predicting Battery Lifetime Under Varying Usage Conditions from Early Aging Data

TL;DR

This work tackles predicting Li-ion battery lifetime under varied charging, discharging, and DoD conditions using early-age data. It introduces new features derived from differential voltage and incremental capacity signals collected during weekly reference tests, and validates them on a large open dataset of 225 NMC/graphite cells. Two prediction approaches are shown: an elastic-net degradation-informed model and a hierarchical Bayesian model; the latter improves extrapolation to unseen cycling conditions and provides uncertainty quantification. The study emphasizes domain-informed feature engineering, reveals robust performance across diverse operating conditions, and provides a publicly available aging dataset to support further research.

Abstract

Accurate battery lifetime prediction is important for preventative maintenance, warranties, and improved cell design and manufacturing. However, manufacturing variability and usage-dependent degradation make life prediction challenging. Here, we investigate new features derived from capacity-voltage data in early life to predict the lifetime of cells cycled under widely varying charge rates, discharge rates, and depths of discharge. Features were extracted from regularly scheduled reference performance tests (i.e., low rate full cycles) during cycling. The early-life features capture a cell's state of health and the rate of change of component-level degradation modes, some of which correlate strongly with cell lifetime. Using a newly generated dataset from 225 nickel-manganese-cobalt/graphite Li-ion cells aged under a wide range of conditions, we demonstrate a lifetime prediction of in-distribution cells with 15.1% mean absolute percentage error using no more than the first 15% of data, for most cells. Further testing using a hierarchical Bayesian regression model shows improved performance on extrapolation, achieving 21.8% mean absolute percentage error for out-of-distribution cells. Our approach highlights the importance of using domain knowledge of lithium-ion battery degradation modes to inform feature engineering. Further, we provide the community with a new publicly available battery aging dataset with cells cycled beyond 80% of their rated capacity.

Figures (18)

• Figure 1: High-level overview of our approach. Unlike existing approaches for early prediction, we extract features from periodic reference performance tests instead of regular cycling data. In this example, we extract a feature from a partial voltage window of incremental capacity that is highly correlated with lifetime. From this and other features, we build a machine learning model to predict the lifetimes of new unseen cells.
• Figure 2: Overview of battery aging test conditions and capacity data. a, 3D scatter plot showing train-test split and cycling conditions used -- each point represents conditions for a group of four cells, and marker color indicates a data subset used to generate prediction results in Sec. \ref{['sec:results_and_discussion']}. b, Discharge capacity fade curves for all 225 NMC/graphite cells plotted past 80% their rated capacity (250mAh); color of each curve is scaled by cell lifetime. c, Histogram of the cell lifetimes at end-of-life (EOL) using 80% of rated capacity as threshold.
• Figure 3: Well-known early-life features do not explain the variance in our dataset, and a newly extracted feature from incremental capacity curves correlates better with lifetime. a, Cell lifetime for 225 NMC cells plotted as a function of $\mathrm{var}(\Delta Q_{\mathrm{w3}-\mathrm{w0}}(V))$; Pearson correlation coefficient -0.686. The two cells highlighted have similar values of $\mathrm{var}(\Delta Q_{\mathrm{w3}-\mathrm{w0}}(V))$ but very different lifetimes. b, Difference between discharge capacity curves as a function of voltage between week three and zero for the two cells highlighted in a. c, Cell lifetime plotted as a function of optimized feature $\text{mean}(\Delta dQ/dV_{\mathrm{w3}-\mathrm{w0}}^{3.60\mathrm{V}-3.90\mathrm{V}}(V))$, Pearson correlation coefficient $-0.848$. d, Incremental capacity curves from weeks three and zero for three representative cells; the change in these between the voltage limits over the first three weeks is shaded.
• Figure 4: a. Scatter plot of mean group lifetime vs. DoD; marker color indicates train/test subset. b. Histogram showing each subset's distribution of cell lifetimes. c. Mean and standard deviation of $\mathrm{RMSE_{\log(EOL)}}$ for five-fold repeated cross-validation on the ten candidate models.
• Figure 5: Overview of HBM results. a, Relationship between $\log (\left|\text{mean}(\Delta dQ/dV_{\mathrm{w3}-\mathrm{w0}}^{3.6V-3.9V}(V)\right|)$ and true lifetime across different clusters and train-test split ("Test" denotes samples from both high- and low-DoD sets). Fits, corresponding to mean parameter values, are plotted for each cluster. b, Predictions for each cluster with 2 standard deviations as the corresponding error bar for each sample. The embedded histograms show a summary of error bars