Thermal runaway: what is predictable and what is only detectable

Thermal runaway prediction is used to describe two different things, and the difference between them is several orders of magnitude in time.

The first is the slow problem: identifying cells or modules whose condition makes them more likely to fail. That develops over weeks to months, leaves a trend, and is a prognostics problem in the ordinary sense.

The second is the fast problem: recognising that a runaway has initiated and is propagating. That develops over seconds to minutes and is a detection problem belonging to the protection system.

Vendors who move between the two in the same sentence are making a safety claim they cannot support, and the distinction is worth insisting on because the consequences of getting it wrong are not commercial.

Internal resistance growth is the most useful single indicator. It rises as a cell ages and rises faster where something is wrong, and it is estimable from ordinary operating data without a dedicated test if the duty provides enough current steps to work from.

Self-discharge behaviour separates a cell with an internal defect from one that is merely old. A cell that loses charge faster than its neighbours during a rest period is saying something specific, and rest periods occur naturally on most duties.

Thermal non-uniformity, meaning a module running consistently warmer than its neighbours under identical duty, is a strong signal precisely because it is comparative. It requires no absolute reference and no model, only enough measurement points to see a population.

Voltage relaxation after a current step, and any cell that consistently relaxes differently from the rest, is a further indicator with the same comparative property.

None of these predicts a runaway. They identify cells whose condition warrants attention, which is a different and entirely defensible claim, and it is the claim worth making.

Once a cell enters runaway the chemistry is exothermic and self-sustaining, and the sequence to venting and to propagation is fast. Nothing about the trend data from the preceding weeks changes what happens next.

This is why the framing matters. A system that identified a suspect module last month has done something useful. A system claiming it will alert before a runaway propagates is claiming a detection capability on a timescale where only physical detection operates.

The protection layers on this timescale are the ones that act without a decision: cell and module design, separation and barriers, enclosure ventilation, and automatic detection and suppression. Analytics does not sit in that path and should not be represented as though it does.

Off-gas detection generally provides the earliest indication, because venting chemistry precedes the thermal signature. The gases released as a cell begins to vent are detectable before the temperature rise has propagated to where a sensor is, which can be worth several minutes.

Temperature sensing is the conventional approach and is limited by placement. The sensor is not usually on the cell that fails, so the measurement lags the event by the time it takes heat to travel, and by then the picture has moved on.

Voltage anomaly at cell level can be immediate where cell-level monitoring exists, since a cell entering runaway behaves electrically before it behaves thermally at the sensor location.

Smoke and flame detection are late by construction. They confirm rather than warn, which is a legitimate function and not an early one.

The honest summary is that these buy notification and evacuation time reliably, and prevention much less reliably, which is exactly what a LOPA can and cannot credit them with. That is set out in applying HAZOP and LOPA to a battery energy storage system.

Mainly it separates a cell problem from a measurement problem. A module reading warm can be a module that is warm or a thermocouple that is wrong, and a set of readings can be checked for mutual consistency against what the duty and the cooling should have produced.

It also normalises for duty. A module that runs warmer because it sits where airflow is poorest is not the same as one that runs warmer at identical airflow, and separating those requires a model of the thermal path rather than a comparison of raw numbers.

And it constrains what can be claimed, which is the more valuable contribution here. A physical model can say that an observed combination of readings is not consistent with any known mechanism, which routes the finding to instrumentation rather than to an evacuation decision.

None of that changes the fast problem. It makes the slow problem more reliable, which is where the available value actually is.

The defensible claim is that continuous cell-level monitoring can identify cells and modules whose condition warrants inspection or replacement, weeks or months before they would otherwise be found, and that doing so reduces the population of cells in poor condition.

The indefensible claim is that a monitoring system prevents thermal runaway. It does not, it cannot be validated to, and asserting it invites reliance that the system will not support in the one case where reliance matters.

There is a middle claim that needs care: that reducing the number of degraded cells reduces the probability of an event. That is reasonable as a statement about population risk and it is not a statement about any individual event, and the two get conflated in marketing material regularly.

For the same reason, the fire detection and suppression system, the enclosure design and the separation arrangement remain the authoritative protection layers, and nothing modelled on top of them changes that.