The battery fire rate fell 97 percent

If you have read anything reassuring about grid battery safety in the last two years, you have read this sentence or a version of it: battery storage failures have dropped 97 percent. It appears in vendor decks, permitting hearings, insurance briefings, and the comment sections of every local fight over a proposed storage site.

It comes from a specific place, and the place is credible. EPRI, the Electric Power Research Institute, keeps the BESS Failure Incident Database, and its May 2024 white paper on that database states it plainly: "Between 2018 and 2023, the global grid-scale BESS failure rate has dropped 97%."

I have no quarrel with the number. Deployment went up by orders of magnitude, annual incident counts did not, and dividing one by the other gives you a collapsing rate. That is a genuine improvement and it reflects genuine engineering work.

What interests me is the rest of the document, because the same sixteen pages describe, with unusual candor, how little is known about the incidents in the numerator. A rate is a fraction. This one has a well-measured denominator and a numerator assembled from whatever became public.

Eighty-one incidents, twenty-six explanations

The methodology section is where the paper earns its keep. EPRI describes its own compilation as "the most comprehensive compilation of stationary BESS incidents," and immediately qualifies how it is built: "based on publicly accessible underlying data."

Then the count:

"At the time of writing, the database contained 81 incidents. Of these, 26 incidents had sufficient information to assign a root cause and to identify the element that experienced failure."

Every root-cause finding in the report, every percentage in every chart, rests on those 26. The other 55 are known to have happened and not known to have a reason. And the paper says why, in the plainest language you will find in an industry document:

"Transparency on the cause of BESS failures continues to be limited. Battery OEMs and BESS integrators are often reluctant to disclose the cause of failure, and many investigation reports are not released to the public. In several instances, legal complications prevent site owners or manufacturers from divulging information about the nature of the failure."

That is not a criticism of EPRI. It is EPRI telling you what its own instrument can and cannot see, which is exactly what a good methods section does. It is also the part that never travels with the 97 percent.

The historical split makes the shape clearer. Of the 9 incidents recorded between 2011 and 2017, none could be classified at all. From 2018 onward, 36 percent had root causes identified. The information environment improved, sharply, and 36 percent is still roughly one incident in three.

What a rate is made of

Here is the structural problem, and it is not unique to batteries. To compute failures per gigawatt-hour deployed you need two quantities that are collected in completely different ways.

The denominator, installed capacity, is close to census data. Projects get financed, interconnected, metered, and reported. Somebody is counting because somebody is paying. It is not perfect, but it is systematic.

The numerator is a media and disclosure artifact. An incident enters this database when it becomes publicly accessible: a fire department responds, a local outlet reports it, a regulator files something, a utility chooses to publish. Nothing compels a global registry of grid battery failures the way, for example, aviation compels accident reporting.

So the two halves of the fraction have different sensitivities. If reporting is more thorough in some countries than others, the rate is not comparable across them. The paper flags exactly this: the outside experts who supplied information beyond the public reports "are based in the United States, so information on incidents in other parts of the world is more limited in this report."

You can see that sensitivity in the data itself. Of the 30 incidents recorded between 2018 and 2019, 27 occurred in South Korea, during a period when Korean installations went from 30 in 2013 to 947 in 2018 under strong subsidies without matching safety regulation. A cluster that dominant is partly a real cluster of failures and partly a country that investigated and published. Both things are true at once, and the database cannot separate them.

None of this means the 97 percent is wrong. It means the honest version of the sentence is longer: the rate of publicly known failures per unit deployed fell 97 percent, in a period when deployment grew enormously and disclosure practices changed at the same time. That is still good news. It is a weaker claim than "batteries are 97 percent safer," which is how it gets repeated.

The finding underneath, which is the useful one

Now the part that changed how I think about these systems, and it comes from those 26 classified incidents rather than from the headline.

EPRI classifies each incident on two axes: the root cause (design, manufacturing, integration and assembly and construction, or operation) and the element that failed (cell or module, controls, or balance of system, where balance of system covers busbars, cabling, enclosures, power conversion, transformers, fire suppression, HVAC, and cooling).

Two results stand out.

First, the most common root cause in the analysis was not chemistry. It was Integration, Assembly & Construction. Not the cell, the act of building the plant.

Second, and more striking: "the majority (72%) of failures where the system age is known happen during construction, commissioning, or within the first two years of operation."

Seventy-two percent, in the front end of the life. The mental model most people carry, that a battery is a chemical hazard that grows more dangerous as it degrades, is close to backwards against this sample. The dangerous period, as measured, is when the thing is new and being assembled, wired, configured, and handed over.

The paper is careful about why that might partly be an artifact too, and I will be as careful. The global fleet is young. Most installed systems have not had the chance to be old, so old-system failures cannot appear in the record yet. EPRI says so directly, noting it remains to be seen whether the trend holds as today's installations age. Read the 72 percent as a fact about a young fleet, not a law of nature.

There is a matching artifact on the other axis. Manufacturing has the fewest failures attributed to it, and the paper attributes that to evidence rather than to quality: it is "most likely due to the difficulty in definitively identifying a manufacturing defect as a root cause with the loss of physical evidence after a fire or explosion." A manufacturing defect is the failure mode most likely to be incinerated by its own consequence. It is undercounted for a physical reason, not because it is rare.

That last point deserves to be sat with. This is a forensic discipline where the event routinely destroys the evidence of its own cause. Any distribution of causes assembled from post-fire investigation is biased toward the causes that survive burning.

One incident that was investigated, and what it looked like

The paper walks through the Elkhorn facility fire at Moss Landing, California, on 20 September 2022, precisely because it is one of the rare ones with a published investigation. As EPRI describes it: rainwater intrusion through the container caused electrical arcing inside the system, leading to thermal runaway in one unit. The water got in through an ingress point created when an umbrella valve was dislodged during improper installation of a vent shield. Insulation loss alarms were recorded and were not properly escalated to the operator. Two days later, smoke and fire were reported to the fire department.

EPRI classified it under two root causes at once: an integration, assembly and construction failure in the balance of system, and a design failure of the control system.

Read that chain again, because there is no chemistry in it until the very end. A part fitted wrong during installation. A path for water. Arcing. An alarm that fired correctly and went nowhere. Two days of warning that nobody acted on. The lithium is the last link, and it is the only link anyone argues about at a zoning hearing.

A footnote here that is on the nose. EPRI cites the public PG&E report on that fire, and when I followed the citation the link redirected to a general newsroom index with no report at it. The single best-documented incident in the set already has a broken trail to its primary document, two years on. That is the disclosure environment the 97 percent is measured in.

What this changes if you are actually buying or siting one

If the failures cluster in construction, commissioning, and the first two years, then the safety questions worth asking are mostly not about the cell datasheet.

• Who is doing the integration, and what is their commissioning record? This is the most common classified root cause in the set. It is also the variable a buyer has the most control over, and the one that never appears in a marketing comparison.
• What happens to an alarm at 2 a.m.? Elkhorn's insulation-loss alarms worked. The escalation path did not. A detection system whose output nobody is obligated to act on is instrumentation, not protection. I have made the same argument about alarms in buildings, and it is the identical failure.
• What is the enclosure sealing and ventilation spec, and who verifies it after installation rather than on the drawing? An umbrella valve dislodged during a vent shield install is a workmanship failure that no cell chemistry prevents.
• Does the contract oblige anyone to publish a root cause? Almost certainly not. The paper tells you why: legal complications routinely prevent disclosure. If you want the industry's data to improve, that clause is where it improves.

This is the layer I work in. I help design the AI battery-cycling systems for a veteran-owned (HUBZone) energy-storage integrator; I do not own the company and earn nothing from this link, and I flag it because integration and controls are what I actually build rather than only write about. Full policy here. The reason this paper landed for me is that it puts numbers on something integrators already suspect: the risk lives in the assembly and the alarm handling, in the same place I have argued the wear does, and not mainly in the cell.

The signal

A 97 percent decline is real and worth celebrating. It is also a ratio between a well-counted denominator and a numerator built from whatever the public was allowed to learn, in an industry whose own most authoritative compilation could explain fewer than a third of the events it had recorded.

The number tells you the trend. It does not tell you the mechanism, and the same document says out loud that nobody can, for most of these fires, because the reports are sealed, the parties are in litigation, or the evidence burned.

The useful finding was never the headline. It is that where causes could be established, the failures were concentrated in construction, commissioning, and the first two years, with integration the leading cause and one incident traced to a valve knocked loose during installation and an alarm that nobody escalated.

Grid batteries are getting safer. They are getting safer mostly at the point where somebody screws them together, which is not where the argument about them is being held.