BESS Failure Rates Fell 98% Since 2018: What EPRI's Data Means for Owners and Engineers

Lithium-ion battery energy storage failures per GWh of deployed capacity fell 98% between 2018 and 2025, according to EPRI's October 2026 analysis of its BESS Failure Incident Database. The 2025 rate stands at 0.04 large-scale failure events per GWh deployed — a level EPRI describes as comparable to the roughly 0.1% annual failure rate of oil-filled power transformers. Half of that eight-year improvement happened in a single year: failures per GWh dropped 50% from 2024 to 2025 alone.

The headline invites an easy conclusion — batteries got safe, congratulations all round — and that reading is partially right but operationally useless. The interesting question for an asset owner or a design engineer is different. Which specific design, control and chemistry changes drove the decline? Which of them are locked into a new build but absent from a plant commissioned in 2019? And what should you do differently in your simulation and financial models now that the risk distribution has shifted?

What Actually Changed — and What Didn't

EPRI attributes most of the decline to improvements between 2018 and 2020 in BESS design, codes and standards, battery management system (BMS) behaviour, and operating practices. This is a familiar pattern in industrial safety: the early failure cluster at Surprise, Arizona in April 2019 — an explosion during firefighting response that hospitalised eight first responders — was followed by a rapid, industry-wide dissemination of lessons learned. Utility Dive reported that the Arizona installation was commissioned before the initial release of NFPA 855, the standard that now governs energy storage system installation in the U.S.

What changed in practice, in the order of impact I would assign from an engineering standpoint:

• Deflagration and gas detection. Early systems had no reliable way to detect the flammable off-gas mixture that precedes thermal runaway in a sealed enclosure. Explosion mitigation was effectively absent in many pre-2019 designs. Today, flammable-gas detection inside enclosures is close to default specification.
• Separation and propagation control. Unit-to-unit spacing, fire walls, and compartmentalisation were inconsistent in early projects. NFPA 855 codified separation distances and outdoor siting, which is directly reflected in the finding that indoor facilities fail more often and sustain more damage.
• BMS state estimation and cell-level diagnostics. The move from module-level to cell-level voltage and temperature monitoring, combined with internal resistance trend tracking, converted many slow-failure modes into caught-and-corrected maintenance events rather than thermal events.
• Commissioning rigour. Factory acceptance testing (FAT) and site acceptance testing (SAT) for the DC blocks became standardised in a way they were not in 2018. For a detailed treatment of the protocol, see our deep dive on BESS factory acceptance testing and commissioning.
• Operating practice. Conservative SOC windows, charge/discharge rate limits at high ambient temperature, and scheduled thermal imaging of terminations moved from "best practice" to contractually specified performance obligations.

What demonstrably did not change is the fundamental hazard. A cell in thermal runaway releases energy at a rate determined by cell chemistry and form factor, not by industry consensus. The safety record improved because the industry got much better at preventing initiation and at limiting the consequence when initiation cannot be prevented. Those are two separate engineering problems, and conflating them is how projects get built to a failure standard that assumes the first problem never recurs.

Read the log scale carefully. Going from 2.0 failures per GWh to 0.04 is not a linear improvement of 49× — it is a change in the nature of the incident distribution. In 2018, a 20 MWh project could realistically expect to experience a large-scale event during its operational life. In 2025, a 20 MWh project occupies 0.02 GWh of the deployed base, so its expected event rate is on the order of one event per 1,250 annual GWh-equivalents — better expressed as an annual probability of roughly 0.04%, before applying EPRI's ranges for site type and vintage.

The LFP Shift: Correlation Is Not Causation

The same period saw the dominant grid-scale chemistry shift from nickel-manganese-cobalt (NMC) to lithium-iron-phosphate (LFP). Many have folded the two trends into a single causal story: LFP is intrinsically safer, therefore failures fell. EPRI explicitly declines to make that claim. The report states it did not have evidence to determine whether battery chemistry in isolation was a risk factor for system-level BESS failures, because the chemistry transition and the design/code learning curve overlap almost perfectly in time.

That is the correct epistemic position, and it matters for engineering decisions. LFP does have a materially higher thermal runaway onset temperature and a flatter, more benign heat-release profile than NMC — this is well-established cell-level physics, and it is one reason LFP dominates stationary storage. But system-level failure is not a scaled-up cell-level property. It depends on module design, propagation barriers, BMS response time, ventilation, and the quality of the installation. A badly designed LFP system can fail; a well-designed NMC system with proper propagation control and gas detection has a sound record. The distinction is worth keeping straight when you are specifying chemistry for a new project — see our comparison of LFP and NMC chemistry for stationary storage for the full trade-off including cycle life and temperature derating.

What the data does support is a compositional argument: EPRI's finding that the failure rate kept falling 50% between 2024 and 2025, long after the code-and-standard improvement curve flattened, suggests continuing gains from ongoing design refinement, BMS firmware maturity, and a deployed fleet that is increasingly new-build rather than early-vintage. That is a compound effect, not a single lever.

Indoor vs Outdoor: The Geometry of Consequence

One of the most actionable findings in the EPRI report is structural rather than electrochemical: indoor BESS facilities are overrepresented in failure incidents and sustain more damage on average than outdoor facilities. The Moss Landing incident in January 2025 is the reference case — thermal runaway in one unit spread through a 1950s-era turbine hall, ultimately destroying most of a 300 MW battery array. No injuries were reported, and subsequent soil and water monitoring found no significant threats, but the asset loss was near-total for the affected array.

The mechanism is straightforward. An outdoor container installation has an effectively infinite heat sink and an unobstructed path for smoke and pressure release. An indoor installation shares a confined volume with other units, and hot combustible gases accumulate under the roof rather than dispersing. Fire walls and compartmentalisation between units, which EPRI recommends for large indoor facilities, change the physics from "one event destroys the plant" to "one event destroys one unit."

For a design engineer, the practical implication is that building-integrated or repurposed-envelope storage carries a different risk profile from a fenced container yard, and the difference should be reflected in both the insurance case and the financial model. If your simulation or business case treats a 200 MWh indoor system and a 200 MWh outdoor system as financially identical apart from land cost, you are missing a term.

The Legacy Fleet Problem: 3% of Capacity, a Disproportionate Share of Risk

This is where the report gets uncomfortable. BESS facilities commissioned before 2021 account for only about 3% of global energy storage capacity, according to the EPRI analysis — but they fail at higher rates than installations built to current codes. A fleet in the low single digits by capacity may be carrying a much larger share of the incident count, which is exactly what a per-GWh normalised average hides.

EPRI's Lakshmi Srinivasan told Utility Dive that the knowledge gap is temporal, not technical: "There are things we know in 2026 that were not necessarily understood in 2019 or 2020, when the older systems were being put in place." The retrofit menu, in ascending order of difficulty and disruption:

1. Flammable gas detection inside battery enclosures — described as a relatively low-difficulty change, and the highest-value single addition for explosion risk.
2. Ventilation upgrades to prevent accumulation of combustible gas mixtures in confined spaces.
3. Structural pressure-relief modifications to direct pressure waves upward and away from first responders.
4. Compartmentalisation — fire walls or barriers segmenting a large indoor facility to limit propagation between units.
5. Early decommissioning, which Srinivasan notes can be the more rational choice when the retrofit cost, downtime, and warranty risk exceed the residual value of the asset.

The warranty point deserves emphasis because it is routinely underweighted in retrofit business cases. Ventilation and envelope modifications generally require changes to battery unit envelopes, which can void the manufacturer's warranty or interfere with other aspects of the system design. A retrofit that solves the explosion risk and invalidates the capacity warranty on a 2019-vintage LFP block is not obviously a net win — the retirement decision has to be modelled as a genuine option, not treated as a last resort.

What Failure Statistics Mean for Your Degradation Model

There is a modelling trap hidden in this data, and it is worth stating plainly: safety performance and capacity performance are different curves that behave differently with age.

Failure statistics improved dramatically as the fleet got newer. Capacity performance moves the other way — SOH declines monotonically, and the wear-accelerating mechanisms (SEI growth, lithium plating at high C-rate and low temperature, electrolyte oxidation at high SOC) compound over the asset life. A plant with an excellent safety record in year two can have a serious augmentation decision in year nine. Treating the EPRI safety trend as evidence that battery risk generally is declining over the asset life is a category error.

Energy Optima's battery degradation model handles the second curve explicitly: SOH and round-trip efficiency are interpolated in three dimensions — year, C-rate, and cycles per day — from real cell data spanning a 26-year projection horizon. That is the correct structure because the dominant failure mechanisms for capacity are rate- and duty-cycle-dependent, not age-only. A system cycling once per day at 0.25C ages on a different trajectory from an identical system cycling twice per day at 0.5C, and a model that treats them the same will misstate both the augmentation schedule and the terminal residual value.

The safety data feeds a different part of the model: availability and unplanned-outage assumptions. If your financial projection assumes 99.5% technical availability for a 2019-vintage asset and 99.5% for a 2026 asset, the EPRI data says that assumption is too coarse. Where a legacy system has not been through a risk assessment, a higher forced-outage rate and a non-zero probability of a total-loss event for the affected array should both appear in the sensitivity analysis. For the financial mechanics of how that flows into IRR and NPV, see our guide to IRR and NPV modelling for BESS.

Engineering Takeaways

Four things I would change in a design review or a diligence exercise after reading this report:

• Specify whether the asset is indoor or outdoor as a risk parameter. Compartmentalisation and fire walls are not optional extras on a large indoor facility — EPRI names them as the control that limits propagation. Price them in at design stage; retrofit is far more expensive.
• For pre-2021 vintage, run the assessment before the renewal decision. Gas detection is cheap and high-value. Envelope and ventilation changes are not, and may void the warranty. Model decommissioning as a real option rather than a fallback.
• Keep chemistry and safety separate in your specification. Choose LFP for cycle life, cost, and thermal margin — but do not let the chemistry choice substitute for propagation control, gas detection and BMS diagnostics. EPRI explicitly could not attribute the decline to chemistry.
• Do not let the safety trend relax your degradation assumptions. Capacity fade and thermal-event risk are different failure classes with opposite age trends. Model them in the same platform but on separate curves.

The broader picture is genuinely encouraging. A storage technology that now fails less often than a transformer fleet, with 93% of incidents producing no injuries or fatalities and no U.S. BESS-related injuries since April 2019, is no longer a nascent technology with unresolved safety characteristics. It is an industrial asset class with an industrial failure profile — which means the right response is conventional industrial asset management: baseline assessment, condition monitoring, and a retirement decision framework that is evaluated on economics rather than deferred by default.

Sources