Operate · Battery energy storage safety

CellGuard

A battery management system raises the alarm when a cell crosses a fixed limit, usually 60 °C. By then an internal short has been building for half an hour. CellGuard compares every cell against its own neighbours, learns what "normal" looks like across 8 features, and spots the fault while the cell is still only a few degrees warm.

Live physics simulation · model trained on simulated normal operation
Simulated time
00:00:00
60× real time
Rack status
-
96 cells · 8 modules
ML alerts / BMS alarms
0 / 0
this session
Last lead time vs BMS
-
inject a fault to measure

Control room

Inject a fault

Applies to the selected cell (or a random one). Watch which detector fires first.

Event log

No events yet.

Rack R-01 · 8 modules × 12 cells

normal
alert

Cell M1-C1

normal

Temperature: cell vs module median (°C)

Ensemble anomaly index (1.0 = alert)

Why the model thinks so (standardised features)

Bars show how far each feature is from normal, in standard deviations. The verdict is a rule layer on top of the detector, so the alert also comes with a likely cause.

Offline benchmark

Run on the server at start-up. Each fault is injected into 3 unseen racks; "BMS" means the first cell reading above 60 °C.

Why it matters

Grid-scale storage is growing fast, and a single cell going into thermal runaway can take out a whole container. Fixed thresholds react late, and they also shut a rack down when the only problem is a faulty sensor. An earlier warning that also names the likely cause means the operator can isolate a string instead of evacuating a site.

Method

  • Peer features: each cell against its module's median, and each module against the rack.
  • A rolling online regression ΔV = a + b·I + c·t separates the OCV offset, resistance and self-discharge trend.
  • Ensemble: Isolation Forest for unusual combinations, plus a calibrated tail model for single extreme features.
  • Persistence filter (12 steps = 2 min) plus a rule layer that only escalates a diagnosis on new evidence.

Limits & next steps

  • Simulated physics only. The next step is validation on real BMS logs, e.g. open cycling datasets.
  • Peer comparison assumes most cells in a module are healthy.
  • Open question: transfer learning across chemistries (LFP vs NMC) and cooling designs.