Source-linked AI summary
From Laboratory Aging Studies to Field Predictions: Quantifying Uncertainty in Battery Storage Lifetime Predictions
Melina Graner, Jan Figgener, Andreas Jossen, Holger Hesse
TL;DR
Battery lifetime forecasts are important for warranty and operational decisions, yet degradation models are mostly deterministic and rarely validated against field capacity measurements. This study applies an open-source probabilistic cell-aging framework to residential BESS using a cell-to-system approximation. Its predictions match the available field measurements, while the analysis identifies sources of end-of-life uncertainty and implications for aging-test design.
Problem
Most degradation models provide deterministic point predictions, while system-level field capacity validation of cell-level models is rare.
Source: §1 · p. 3, §1.1 · p. 5Method
The study applies an open-source probabilistic cell-aging framework to system-level residential BESS prediction using minimum- and maximum-stress scenarios.
Source: §2 · p. 6Results
0.4% mean absolute SOH error on independent cell-level validation profiles accompanied predictions containing all three available field capacity measurements; the combined predicted EOL range spans approximately three years.
Source: §5 · pp. 12–13Takeaways & Limitations
The EOL probability distributions support warranty terms, risk-aware maintenance scheduling, replacement planning, and second-life decisions.
Source: §5 · p. 13Training data covers degradation only down to approximately 80% SOH, so late-life knee-point behavior is outside the model scope.
Source: §2.1 · p. 7, §5 · p. 13
Abstract
from arXiv · showhide
Predicting how long a battery energy storage system will last is critical for warranty design, maintenance planning and investment decisions, yet degradation models are mostly deterministic and rarely validated against real field data. We apply an open-source probabilistic degradation framework, combined with a cell-to-system approximation, to bridge the gap between cell-level laboratory aging models and system-level field predictions for residential battery energy storage systems with quantified uncertainty. The framework predicts cell-level state-of-health to within 0.4 % mean absolute error, roughly half the error of prior models for this dataset. When applied to field operation data, the framework's predictions are consistent with all three available system-level capacity measurements - a benchmark rarely available for open probabilistic degradation models. Cell-level heterogeneity is approximated by two bounding stress scenarios differing only slightly (a 5 % spread in temperature and a 9 % spread in current). The mean degradation trajectories of the two scenarios reach end-of-life 10 months apart, while the full predicted end-of-life range across both scenarios spans approximately three years, about a third of the expected system lifetime. We link this uncertainty to two drivers. One is a systematic mismatch between laboratory test conditions and field-representative operating stress. The other is variability in the training data itself. These insights translate into concrete, resource-efficient recommendations for future aging study design, supporting more confident predictions of battery lifetime under real-world conditions.
Highlights
The paper presents an open-source probabilistic framework for battery degradation prediction, benchmarks its cell-to-system approximation against field measurements, and analyzes lifetime uncertainty and future test design.
Highlights
- The open-source framework predicts battery degradation with quantified uncertainty.It combines probabilistic degradation prediction with a cell-to-system approximation.Source: Highlights · p. 2
- The cell-to-system approximation is benchmarked against BESS field measurements.The paper highlights the benchmark as a contribution of the study.Source: Highlights · p. 2
- Predicted end-of-life uncertainty spans about one third of the expected system lifetime.The paper identifies this as a key result of the uncertainty analysis.Source: Highlights · p. 2
- Uncertainty-driver analysis identifies laboratory–field coverage gaps in test design.Source: Highlights · p. 2
- The analysis yields recommendations for resource-efficient, field-relevant aging test campaigns.Source: Highlights · p. 2
1. Introduction
The introduction motivates probabilistic battery lifetime prediction and system-level field validation, then presents a laboratory-trained approach and its implications for future aging studies.
1. Introduction
- Reliable degradation forecasts matter for warranty and operational planning, but most models provide deterministic estimates and cell-level models are rarely validated against system field measurements.The system-level gap matters because warranty compliance and operational value are ultimately determined at system level.Source: §1 · p. 3
- The study applies an open-source probabilistic framework to transfer LFP cell-aging models to residential BESS system-level SOH and RUL prediction.Its contributions include dynamic cell validation and benchmarking against field capacity measurements.Source: §1 · pp. 3–4
- Prediction intervals offer manufacturers and operators a quantified basis for warranty design, maintenance scheduling, and second-life planning.Source: §1 · p. 4
1.1. State of the Art
- Prior work leaves system-level capacity-based prognosis for real-world BESS operation insufficiently explored, with field-capacity benchmarking and quantified predictive uncertainty uncommon.Laboratory aging datasets also underrepresent LFP cells relative to their deployment in BESS.Source: §1.1 · p. 5
- Systematic cell-to-system effects and stochastic divergence in cell aging are distinct uncertainty sources, while test design must balance accelerated degradation with representative operating-condition coverage.The study compares laboratory and field stress distributions and identifies training-data variability as a driver of aleatoric uncertainty.Source: §1 · p. 4, §1.1 · p. 6
1.2. Previous Work
- The present work extends the open-source probabilistic framework [47] with validation accuracy, uncertainty-driver analysis, and aging-campaign recommendations, drawing on HESS field data with a matched cell chemistry.Source: §1.2 · p. 6
2. Methodology
The methodology applies probabilistic cell-aging models to dynamic operation, represents uncertainty through stochastic trajectories, and approximates system effects with two limiting stress scenarios.
2. Methodology
- The framework uses laboratory-trained probabilistic LFP cell models and applies them to cell- or system-level cycling data, using minimum- and maximum-stress scenarios for system predictions.Source: §2 · p. 6
2.1. The Prediction Framework
- Calendar and cyclic models predict relative capacity loss under static stress; dynamic profiles are processed in sequential 90-day blocks using stress features.Calendar inputs include mean SOC, temperature, and virtual time; cyclic features include SOC, temperature, charge and discharge rates, and depth-of-cycle.Source: §2.1 · pp. 6–7
- Training-data variability contributes aleatoric uncertainty, while bootstrap model ensembles represent epistemic uncertainty; observations only reach approximately 80% SOH, excluding late-life knee-point behavior.Source: §2.1 · p. 7
2.2. Uncertainty Quantification and Stochastic Trajectory Interpretation
- Each stochastic trajectory samples persistent deviations for calendar and cyclic aging, accumulates incremental losses, and clips negative capacity-loss outputs to zero.Source: §2.2 · pp. 7–8
2.3. Scaling from System to Cell-Level
- The cell-to-system approximation bounds plausible degradation using scenarios that combine the highest or lowest recorded module current and cell temperature.The maximum-stress scenario is particularly relevant because the weakest subunit in a series connection limits string capacity and system SOH.Source: §2.3 · p. 8
3. Results
The framework is validated against cell-level aging data and applied to field operation data for a residential battery system. Its predicted trajectories encompass the available field capacity measurements, while uncertainty in end-of-life timing remains substantial.
3.1. Validation on Cell-Level
- Approximately 0.4% mean absolute SOH prediction error was achieved on both cell-level profiles, versus approximately 0.9% for the prior bootstrapped semi-empirical model.The authors suggest that training on full time-resolved data, rather than per-condition scalar reductions, may explain the accuracy gap.Source: §3.1 · p. 8
3.2.1. Data
- Field data for system 14 combine multi-year HESS measurements with granular module-current and cell-temperature extrema; operation averages 152 FECs per year.The case study selects system 14 because it uses the same commercial LiFePO4/graphite cells as the laboratory-trained models.Source: §3.2.1 · pp. 8–9
3.2.2. Obtaining Cell-Level Data
- Cell-level currents are estimated by dividing module currents among parallel-connected cells under an equal-current-split assumption, and minimum/maximum current and temperature readings define the stress scenarios.The modules comprise 16 series-connected blocks of 14 parallel cells; the recorded minimum and maximum cell temperatures differ by 1.56 °C on average.Source: §3.2.2 · p. 9
3.2.3. Prediction Results and Evaluation
- All three converted field SOH measurements lie within the predicted intervals, plausibly bracketed by the Minimum- and Maximum-Stress Scenarios.The measurements fall near the Maximum-Stress mean, near the Minimum-Stress mean, and between the two means, respectively.Source: §3.2.3 · pp. 9–10
- The predicted EOL range spans approximately 3 years—roughly one third of mean system lifetime—while scenario mean trajectories reach EOL 10 months apart.The scenarios differ by roughly 5% in temperature and 9% in current.Source: §3.2.3 · p. 10
4. Uncertainty Driver Analysis
The analysis links wider field prediction intervals to weak laboratory support for field operating conditions and greater variability in degradation observations. It examines these uncertainty drivers and derives targeted recommendations for laboratory test design.
4.1. Epistemic Uncertainty: Laboratory-Field Alignment
- Coverage measures whether field samples fall within the laboratory range, while Focus Match and Nearest Overlap indicate local support that range coverage alone can miss.Low Nearest Overlap indicates weak local support; high Coverage therefore does not necessarily imply strong support.Source: §4.1 · p. 10
- Calendar SOC is well aligned, but field temperatures, though within the laboratory range, are less locally supported than validation temperatures, plausibly contributing to wider field intervals.Validation stress conditions generally match laboratory test points, whereas field usage more often occupies weakly supported or unsupported states; validation predictions show less spread.Source: §4.1 · pp. 10–11
- Cyclic aging shows its strongest laboratory-field mismatch in charge and discharge rates, with weak field support and especially poor discharge-rate support even for validation profiles.DOC remains comparatively well represented, while cyclic SOC and field temperature are less supported than in the validation profiles.Source: §4.1 · p. 11
- Field prediction intervals may also reflect longer-horizon extrapolation and dynamic usage, whose idle phases may reduce degradation severity or permit partial recovery.These possible effects are not isolated by the alignment analysis and remain plausible contributors to epistemic uncertainty.Source: §4.1 · p. 11
4.2. Aleatoric Uncertainty: Capacity Measurement Spreads
- Field-dominant calendar conditions include a 25°C, 100% SOC test point with a 5.9% observable relative capacity-loss spread, exceeding the other considered calendar points.The spread is driven by one cell’s drastic capacity loss; the corresponding validation profiles are dominated by moderate SOC states and generally show spreads below 1%.Source: §4.2 · p. 11
4.3. Uncertainty Sensitivity Analysis
- 3 years of EOL uncertainty narrowed to a little more than 1.5 years when one cell’s unusually wide spread was reduced to about 0.8%.The earliest threshold crossing shifted from 9 to almost 10 years, and the latest from 12 to 11.5 years; with three cells per test point, the study cannot distinguish an outlier from genuine variability.Source: §4.3 · p. 12
4.4. Implications for Laboratory Test Design
- Field-relevant test points can run alongside accelerated tests for the same campaign duration, improving support without adding calendar time, though they require parallel test channels.The study recommends prioritizing additional replicates at field-dominant test points; doubling from three to six cells lowers the relative uncertainty of the spread estimate from about 50% to 32%.Source: §4.4 · p. 12
5. Conclusion and Outlook
The study transfers a laboratory-trained cell-level probabilistic aging model to field-operated BESS and finds its predicted intervals bracket all three available system capacity measurements. It also identifies uncertainty drivers, practical applications, and limits of the current model.
Conclusion and Outlook
- The model achieves 0.4% mean absolute SOH error on independent cell-level validation profiles.The cell-level model is laboratory-trained and transferred to system-level SOH and EOL prediction for field-operated BESS.Source: §5 · p. 12
- All three available field capacity measurements over approximately four years fall within predicted intervals, while the two stress scenarios’ mean EOL predictions differ by 10 months and their combined EOL range spans about three years.The combined range is about a third of predicted system lifetime; the scenarios differ by roughly 5% in temperature and 9% in current.Source: §5 · p. 13
- Field-operation prediction intervals are wider than validation-profile intervals, with laboratory–field stress coverage gaps indicating an increased epistemic contribution and laboratory capacity-measurement spread indicating an increased aleatoric contribution.Replacing elevated-degradation measurements from one cell at a frequently visited test condition narrows predicted EOL range by roughly 50%.Source: §5 · p. 13
- EOL probability distributions benchmarked against field capacity measurements offer manufacturers a basis for SOH-based warranty terms and risk exposure assessment, and can support operators’ maintenance, replacement, and second-life planning.The framework also offers laboratories alignment metrics and sensitivity analysis as a basis for prioritizing test resources.Source: §5 · p. 13
- The training data extend only to approximately 80% SOH, leaving late-life knee-point behavior outside scope; the approximation also omits pack current spreads and manufacturing-related variation in initial cell capacity.Source: §5 · p. 13
- Future work could extend training into slowly degrading field-relevant regimes, use physics-informed models and automated outlier detection, and benchmark additional field cases and system topologies.Synthetic unaccelerated test points using Arrhenius-based temperature scaling are one proposed approach to coverage gaps.Source: §5 · p. 13
- The framework, code, and training data are openly available, and the authors present the approach as a practically usable tool for warranty design, maintenance planning, and future test-campaign design.Source: §5 · p. 13
Funding and acknowledgment
The study acknowledges project funding from the German Federal Ministry for Economic Affairs and Energy.
Funding and acknowledgment
- The research was funded by the German Federal Ministry for Economic Affairs and Energy via the BattLifeBoost project (grant 03EI4068C), overseen by Project Management Juelich.Source: Funding and acknowledgment · p. 13
CRediT authorship contribution statement
The statement assigns the paper’s research and writing contributions across four authors.
CRediT authorship contribution statement
- Melina Graner contributed to conceptualization, methodology, software, validation, analysis, investigation, data curation, drafting, and visualization; Jan Figgener to resources and review; Andreas Jossen to supervision and review; and Holger Hesse to funding acquisition, supervision, and review.Source: CRediT authorship contribution statement · p. 14
A. Cumulative Loss and Virtual Time Principle
The cumulative-loss method builds a dynamic loss trajectory from incremental losses under changing stress states. Virtual time relates calendar capacity loss at one point to the duration required under another point’s stress factors.
A. Cumulative Loss and Virtual Time Principle
- For calendar aging, the method divides the timeline into Δt increments, identifies incremental losses ΔL for the stress states dominating each increment, and appends their relative portions to the dynamic loss trajectory.The example describes varying SOC levels as the stress states.Source: A. Cumulative Loss and Virtual Time… · p. 14
- Virtual time t* is the duration that would be needed to reach point A’s calendar capacity loss under the stress factors present at point B.Source: A. Cumulative Loss and Virtual Time… · p. 14
B. HESS Operational Data
HESS operating conditions vary seasonally: summer brings more full charging and higher state of charge, while winter brings fewer full cycles; spring and fall show more balanced cycling.
B. HESS Operational Data
- Summer operation brings more frequent full charging and extended high-SOC periods, while winter brings fewer full cycles under lower PV generation.Spring and fall have more balanced SOC distributions, with systems regularly reaching both full and empty states.Source: B. HESS Operational Data · p. 14
- Charge and discharge rates have slightly higher means in spring and fall and are lowest in winter.Source: B. HESS Operational Data · p. 14
- Battery housing temperatures follow seasonal ambient-temperature patterns in Germany, with higher summer means linked to warm rooms and high charging power.Source: B. HESS Operational Data · p. 14
- Average hourly energy throughput is reported across the LFP-based HESS systems published by Figgener et al..Source: B. HESS Operational Data · p. 14