From fleet trajectories to single-cell promises: hierarchically weighted conformal prediction for battery SOH monitoring

Xiangyu LiuarXiv 2026 · arXiv

TL;DR

Standard conformal prediction provides "fleet-level" guarantees that often mask disastrous failures for individual batteries. This paper introduces a hierarchically weighted conformal engine and a three-tier reporting protocol that ensures per-cell reliability, revealing that "battery diversity"—not just more data—is the secret to robust state-of-health (SOH) monitoring.

What Problem Does This Paper Address?

In the world of electric vehicles and grid storage, knowing a battery's State of Health (SOH) isn't just a technicality—it’s a "cash number" that drives warranty claims and replacement schedules. Engineers often use Conformal Prediction (CP) to put uncertainty intervals around SOH estimates. CP is popular because it’s "distribution-free," meaning it doesn't assume your data follows a perfect bell curve.

However, there’s a catch: CP guarantees work on average across a "calibration pool." Battery data is nested—thousands of cycles belong to one cell, and dozens of cells belong to one manufacturing batch. The authors show that a model promising 90% coverage for a fleet might actually be wrong 98.6% of the time for a specific, "worst-case" cell. If you're the operator of that specific cell, the fleet-level guarantee is useless.

Figure 1: The object of the guarantee and the three-tier reading (concept figure; numbers from panel A, SOH scale, Table 1). (a) Calibration cycles nest inside cells; cells nest inside protocol batches. The conformal guarantee attaches to a draw from the pooled cycle set; maintenance decisions attach to one cell. (b) The same experiment read at three tiers: fleet pooled coverage 0.906, minimum batch coverage 0.787, worst single cell 0.642, against nominal 0.90; 28% of cells fall below 0.85. The headline tier passes while the decision- relevant tier fails.

Methods

The authors propose a hierarchically weighted conformal engine. Instead of treating every cycle in the calibration pool as equally important, the engine re-weights them based on four factors:

  1. Cell Similarity: How much does the calibration battery look like the test battery in "feature space"?
  2. Batch/Protocol: Are they from the same manufacturing batch or charging protocol?
  3. Cell Balance Cap: Prevents "rich" cells with many cycles from dominating the statistics.
  4. Trajectory Position: Aligns cycles by their aging stage.

Crucially, they follow a "freezing protocol." All weights are calculated using information available at the time of prediction (like cycle index or initial capacity) and never "peek" at the actual error residuals. This makes the intervals auditable and legally defensible for applications like insurance or regulatory compliance.

Datasets

  • Severson 124 (Main): A public dataset of 124 Lithium Iron Phosphate (LFP) cells cycled under various fast-charging protocols. This is the primary benchmark for the "fleet vs. cell" gap.
  • Oxford (External): Eight NMC-LCO pouch cells from a different lab with different load profiles. This was used as a "stress test" to see if the model's promises hold up across different battery chemistries and environments.

Key Insights and Ideas

  • Diversity > Volume: The "repair capacity" of the model is limited by the diversity of batteries in the calibration pool, not the total number of cycles. To get better per-cell promises, you shouldn't just cycle the same batteries longer; you need to buy more different types of batteries.
  • The "Mondrian" Trap: The authors tested a simpler "Mondrian" (group-aware) approach. They found it could "cosmetically" fix batch-level statistics while leaving the worst-performing cells just as broken as before. This highlights why a three-tier (fleet, batch, cell) reporting protocol is necessary.
  • The Frozen-Signature Dead End: When the charging protocol changes mid-fleet (protocol drift), the "frozen" weights can actually harm performance. The authors argue this is a structural limitation: you can't fix a fundamental change in physics with weights that aren't allowed to look at the new errors.

Key Results

The paper's "collapse exhibit" shows just how misleading fleet averages can be. On a standard split, the fleet coverage was a healthy 0.906 (against a 0.90 target), but the worst cell was covered only 64.2% of the time. In more extreme "official" splits, the worst cell coverage plummeted to a disastrous 1.4%.

Figure 2: Per-cell coverage distribution. The "disastrous tail" shows that many cells fall far below the nominal 90% target, even when the fleet average looks good.

The proposed weighted engine successfully "repaired" much of this gap:

  • 10th-percentile cell coverage moved from 0.750 to 0.830, closing 53% of the gap to the nominal target.
  • Worst-cell coverage improved from 0.642 to 0.694.
  • Cross-chemistry transfer was a reality check: coverage collapsed to 0.286, proving that intervals calibrated on one chemistry are "meaningless" when applied to another.

Limitations and Future Work

  • Protocol Drift: The current engine cannot "repair" sudden shifts in charging protocols without a "requalification window" where new data is collected.
  • Chemistry Boundaries: The failure of cross-chemistry transfer suggests that "universal" battery SOH intervals remain out of reach for now.
  • Deferral Logic: While the engine can "defer" (refuse to predict) when evidence is short, the authors suggest that future versions should use Effective Sample Size (ESS) as a more natural trigger for these warnings.

Summary updated

Abstract

Interval predictions from conformal prediction are usually read as "90% coverage", but on nested battery data that number promises the fleet average, not the individual cell. We measure the gap on 124 public LFP cells under a preregistered three-tier reporting protocol (fleet, batch, cell): pooled coverage looks healthy (0.875-0.906 across splits) while the worst cell is covered 0.1-2% of the time, with exact binomial tests against the nominal level. We then repair part of the gap with a hierarchically weighted conformal engine whose weights consume only side information frozen at prediction time - signature features, batch labels, cycle indices - and never touch residuals. The preregistered repair criterion passes: 10th-percentile cell coverage moves 0.750 to 0.830 (53% of the gap), the worst cell 0.642 to 0.694, with zero deferrals and a flat sensitivity spectrum. What the engine does not repair is localized rather than left open: under protocol drift the frozen weights harm monotonically on the SOH scale (the absolute scale is reported in full), a structural dead end we report instead of tuning away; and repair capacity is bounded by calibration-pool battery diversity, measured on both datasets. On a cross-chemistry external set the calibrated intervals collapse (0.286), while the within-domain skeleton replicates (pooled 0.933, worst cell 0.686). The deliverable is a reporting protocol - three tiers of promise anchored to measured regimes, worst-cell statistics, deferral share - that makes per-cell interval claims auditable before deployment.
BibTeX
@misc{liu2026from,
  author = {Xiangyu Liu},
  title = {{From fleet trajectories to single-cell promises: hierarchically weighted conformal prediction for battery SOH monitoring}},
  year = {2026},
  eprint = {2610.41696},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.41696},
}
Figure