KIT battery aging · historical replay

Forecast capacity change three check-ups ahead

Replay 228 measured cells along their historical timelines, then compare forecasts with and without EIS for the 180 cells eligible for paired evaluation.

What is predicted? SoH loss over the next 3 check-ups Target: SoHnow − SoHthree check-ups later
Cell roster 228 180 with paired OOF forecasts; 48 outside the current modeling conditions
Replay position Check-up 10 / 28 Advance through recorded completed check-ups
Which cell should we inspect? Calculating Ranked by model disagreement for research review only
Condition-held-out development OOF

Core question: does adding EIS improve the forecast?

There is no single answer. Error deltas are D_CORE − B: negative means lower error after adding EIS; positive means higher error. These are paired development comparisons on the same snapshots, folds, and model family—not a causal test.

Paired EIS increment study · primary policy

Six paired results: EIS can help, hurt, or remain uncertain

MAE and RMSE are in SoH percentage points; R² is a supporting metric.

Why can't we simply say “EIS works” or “EIS does not work”?

Aging modes and model families point in different directions. When the 95% interval crosses zero, the current development evidence cannot establish a direction. The bootstrap resamples complete conditions; 2,000 draws do not become 2,000 new cells.

Development OOF · not final

Reveal the history one check-up at a time

“Check-up” is the observation order, not a shared number of days, equivalent full cycles, or kilometers. Each cell retains its own elapsed days. D_CORE uses EIS matched to the current check-up; this artifact does not include mileage.

228 real KIT cells · development replay

What does the fleet look like at this moment?

Paired B / D_CORE forecast available Insufficient history No third future check-up No record at this check-up Research review suggested (model disagreement)
Selected cell · Development OOF

Selected cell: from measurement to forecast

1 Read capacity history available now 2 Run the same model with B and D_CORE 3 Reveal the measured answer three check-ups later
Measured capacity now
Measured SoH nowSoH = current capacity ÷ initial capacity
B: forecast loss / future SoH
D_CORE: forecast loss / future SoH

Measured capacity history available at this moment

KIT cell history and future SoH forecasts from two feature views The solid line is measured history visible now; dashed lines connect B and D_CORE future forecast points.
MeasuredB forecastD_CORE (B + EIS) forecast

Forecast points are not measurements; future measured data to the right of the current timeline remain hidden.

Three check-ups later: compare with the measured answer

Technical details and provenance
Evidence level
KIT condition-held-out DEVELOPMENT OOF; not a protected test.
Primary target
soh_loss_pp_h3 = SoH_t - SoH_t+3
Feature views
B: capacity history + context; D_CORE: B + EIS matched to the current check-up.
Grouping
Complete experimental conditions are held out; three cells from one condition do not cross training and validation.
Model selection
selection_performed=false; the page compares B and D_CORE within the same model family.
Sensitivity
Primary and keep-last acquisition policies give the same six deltas; timestamp/retry semantics remain provisional.
Cell / fold
Source artifact

KIT historical-data replay. This page shows condition-held-out development OOF estimates—not a physical simulation, live monitor, protected external test, fault diagnosis, or battery-replacement recommendation. The difference between B and D_CORE is not a causal effect of EIS. Source: Matthias Luh and Thomas Blank (2024), KIT v2 processed result data, licensed CC BY 4.0. This interface uses transformed, derived replay data; dataset description: Scientific Data (2024).