Equipment failure rates from public databases#
For the expanded solar-specific review, including trackers, weather stations and combiner boxes, see PV device reliability: 14 reviewed sources, 371 additional evidence records, new operational methods and three gallery examples.
Estimate failure rates from observed exposure, use documented industry distributions as priors, and evaluate fault-interval models. The implementation separates input connectors → numerical methods → applied analysis. It does not equate degradation, unavailability, incident counts, and failure rates.
Included databases and coverage#
The frozen releases below were retrieved on 7 November 2026. Publication vintage and observation dates are retained separately. These are historical benchmarks, not a claim to represent today’s equipment fleet.
Source |
Stored / usable records |
Equipment and scope |
Access |
|---|---|---|---|
NRC/INL 2020 update, Table 1 |
300 / 294 |
58 published equipment groups; all table rows, including legacy estimates |
Public report; underlying IRIS records restricted |
Sandia PVROM 2018, Appendix A |
63 / 39 |
All published failure fits across 18 headings, including inverter fault groups and grid events |
Public distributions; raw work orders not included |
Carroll et al. 2016, Figure 3 |
76 / 76 |
19 offshore wind subassemblies × 4 repair categories |
Public numerical table from approximately 350 European turbines |
40 / 40 |
Every drive model in the daily CSV archive |
Public raw observations; attribution and no resale of source data |
479 stored records; 449 evaluated records. These are source/population/mode records, not 479 distinct devices. Device categories overlap across sources. Snapshot JSON files ship in the retrieval wheel; the study exports Parquet, CSV, JSON and a four-panel figure. Downloaded PDFs/archives are not committed.
NRC coverage includes actuated/manual/check/relief valves; motor-, turbine-, engine- and positive-displacement pumps; diesel/hydro/combustion generators; battery chargers, batteries, transfer switches, breakers, inverters, buses, motor-control centres, transformers, sequencers and fuses; strainers and screens; sensors, logic, relays and switches; reactor-specific equipment; dampers, air handlers, chillers and fans; compressors, dryers, accumulators, cooling-tower fans, tanks, orifices, pipes and heat exchangers. Nuclear batteries/inverters are not automatically lithium-ion BESS/PV equivalents; nuclear backup hydro generators are not a global hydro fleet.
PV coverage includes modules, combiners, inverters, fans, IGBTs, cooling, power cycling, fuse/hardware faults, ground/arc faults, AC disconnects, transformers, grid/recloser events, data acquisition, PLCs and hydraulic cylinders. Preventive maintenance remains in the audit but is excluded from failure calculations.
Wind coverage includes pitch/hydraulics, generators, gearboxes, blades, lubrication, electrical components, contactors/breakers/relays, controls, safety, sensors, pumps/motors, hubs, heating/cooling, yaw, towers/foundations, converters/power supplies, service items and other components. No underlying OEM identities are inferred from the anonymized fleet.
Run the analysis#
uv sync --locked --all-packages
uv run python -m energy_analysis.failure_reliability --output artifacts/failure-reliability
Outputs include failure_database.parquet, failure_database.json,
failure_estimates.csv, wind_budget_scenario.csv, sources.json,
summary.json and failure_reliability.png. The gallery runs offline:
from energy_data_retrieval.failure_database import load_failure_database, source_catalogue
database = load_failure_database() # excludes quarantined records
audit = load_failure_database(include_flagged=True)
pumps = database[database.device.str.contains("Pump", case=False)]
print(pumps[["device", "failure_mode", "basis", "distribution", "source_url"]])
print(source_catalogue()[["source_id", "access", "limitations"]])
Explicit live refresh, with byte limits and source SHA-256 verification:
uv sync --locked --all-packages --extra reliability
uv run python scripts/refresh_failure_database.py nrc_2020 --output artifacts/refreshed
uv run python scripts/refresh_failure_database.py pvrom_2018 --output artifacts/refreshed
uv run python scripts/refresh_failure_database.py wind_2016 --output artifacts/refreshed
uv run python scripts/refresh_failure_database.py backblaze_2013 --output artifacts/refreshed
These reproduce specific releases. An upstream document change fails the hash
check and requires review; it does not silently replace the benchmark. The optional
PDF parser uses pdfplumber; ordinary snapshot use needs no PDF dependency.
Backblaze’s 2013 archive is approximately 78 MiB compressed; expanded content is
bounded to 1 GB and read one daily CSV at a time, without filesystem extraction.
aggregate_backblaze also accepts current daily CSVs downloaded separately from
the provider. The connector does not automatically download multi-GB quarters.
Methods and units#
Input |
Estimator/model |
Output |
|---|---|---|
Failure count \(k\), summed at-risk time \(T\) |
\(\hat\lambda=k/T\); exact central Poisson count interval |
Failures per declared time unit; constant-rate MTBF \(1/\lambda\) |
Zero failures over \(T\) |
One-sided bound \(-\ln(1-c)/T\) |
Upper rate bound at confidence \(c\); zero observed failures does not prove zero risk |
Gamma industry prior \((a,b)\), new failures/exposure |
Shape/rate posterior \((a+k,b+T)\) |
Mean rate and credible interval; predictive \(1-(b/(b+t))^a\) |
Beta prior \((a,b)\), new failures \(k\) in \(n\) demands |
\((a+k,b+n-k)\) |
Per-demand probability and beta-binomial predictive probability |
Lifetimes plus survivor censoring flags |
Weibull maximum likelihood, including censored survival terms |
Shape and scale; conditional future-event probability |
Published lifetime/fault-interval fit |
\(1-S(a+h)/S(a)\) |
Probability of the next event over horizon \(h\) conditional on event-free age \(a\) |
The count estimator and zero-failure bounds follow the NIST engineering statistics handbook. Poisson intervals assume independent homogeneous fixed-exposure counts; they do not account for reporting bias, repeated-failure dependence or fleet heterogeneity. NRC’s industry-distribution percentiles are a different uncertainty quantity.
Gamma rate uncertainty uses a rate parameter with units of exposure. Gamma lifetime uses a time scale. Weibull scale has the same time unit as age. Pipe records remain per foot-hour; demand records remain per demand. No conversion to annual probability is made without an explicit exposure/demand mission.
from energy_methods.failure_rates import estimate_rate, GammaRate, BetaDemand
observed = estimate_rate(failures=2, exposure=800) # device-hours
print(observed.rate, observed.lower, observed.upper)
# Synthetic illustration: local data must be independent of the industry prior.
prior = GammaRate(shape=2, rate=100_000)
posterior = prior.update(failures=3, exposure=50_000)
print(posterior.probability_at_least_one(8760))
start_failure = BetaDemand(alpha=1, beta=999).update(failures=1, demands=500)
print(start_failure.probability_at_least_one(demands=12))
For any device absent from the supplied sources, use your own event count and
summed exposure in estimate_rate, or include all survivors in
fit_weibull_censored(times, failed). Times must start at device age zero: the
first SCADA observation is not necessarily the installation date. The fitter
does not support delayed entry, recurrent-event intervals or interval censoring.
Do not pool populations without matching duty, environment, maintenance,
failure definitions and observation windows. Do not update a prior with the same
observations used to construct it. The package deliberately has no automatic
cross-technology transfer or pooling function.
Results and interpretation#
The NRC normally-running motor-driven pump example reports a mean of 2.26 failures per million run-hours, with source 5th/95th percentiles of 0.394 and 5.38. This is a nuclear fleet benchmark for the specified failure mode; it is not a calibrated estimate for a solar-plant pump.
The wind table sums to 8.273 repair events per turbine-year (the paper’s prose rounds the fleet value to 8.3). Pitch/hydraulics has the highest subsystem event frequency, 1.076 per turbine-year. Major gearbox replacements occur at 0.154 per turbine-year in this historical sample. Under the explicitly illustrative costs below, gearboxes and generators dominate direct repair cost, despite pitch/hydraulics having more events.
Illustrative input |
Value |
|---|---|
Fleet |
100 turbines |
Minor repair |
EUR 500/event |
Major repair |
EUR 5,000/event |
Major replacement |
EUR 50,000/event |
Unknown-cost event |
EUR 1,500/event |
Expected events/year |
827.3 |
Expected direct repair cost/year |
EUR 2,280,100 |
This is a sensitivity case, not a current price quote or a complete O&M budget.
It excludes lost generation, vessel/access constraints, logistics and correlations.
Replace costs through wind_maintenance_budget(cost_per_event=...).
The complete 2013 Backblaze archive yields 724 reported failures over 5,091,501 observed drive-days, across 40 models. Calendar gaps add no exposure. There are 180 observed drive-days after a prior reported failure and zero repeated failure flags; these audit fields remain visible. Results describe recorded events and are not independent first-failure lifetime data. Tiny model cohorts and zero-event cohorts retain their uncertainty; historical model comparisons are not purchasing recommendations. Drive model strings alone are not used to infer SSD/HDD technology for arbitrary future imports.
PVROM curves describe the source’s portfolio-specific fault/failure intervals, including resets and grid-related events. They cannot establish universal warranty life or a 30-50-year replacement plan. Plotted five-year curves are extrapolations of historical fitted distributions, with no invented parameter confidence bands. Normal time models are explicitly conditioned on nonnegative time; this is an implementation adaptation, not an exact reproduction of an untruncated fit.
Data-quality exclusions and access gaps#
Six NRC rows have a distribution/unit conflict:
MDP-FTS-CCW,CTG-FTLR,CRD-FTOP,ROD-FTOP,MOD-ILS,EDC-FTR-NR. The connector preserves the published values and excludes them from forecasts pending source review.PVROM: 19 duplicate year-based fits are excluded in favour of the source’s day/hour rows; two preventive-maintenance fits are not failures; three
Lognormalrows have ambiguous mean/standard-deviation conventions. ExplicitLognormal-nuses the underlying normal parameters as documented.The PVROM 2020 Table 3-1 summary is catalogued but not numerically ingested: reported means and scale conventions need clarification, and the statistic concerns first occurrence at system level rather than identifiable devices.
OREDA@Cloud requires subscription/member access. No public numeric endpoint or licensed export was available for this contribution.
EPRI’s BESS incident database is an incident source. Without a matched at-risk fleet denominator and reporting completeness, its counts do not determine a component failure rate. Battery degradation/cycling datasets likewise do not establish fleet failure incidence.
Sandia CREW publishes benchmarks but its underlying operating data are restricted.
Sources retain their own terms. The repository’s software license does not relicense third-party publications or datasets. Backblaze attribution and its restriction on selling the source data itself apply to its extract.