Solar PV device reliability: evidence, estimation and operational use#
Research updated 7 November 2026. This extension reviews 14 sources and adds 371 evidence records, with two reproducible PDF-table connectors. It complements the existing Sandia PVROM connector rather than duplicating its 63 stored / 39 usable fault-interval fits. These are published observations and benchmarks, not access to proprietary maintenance databases.
Evidence inventory#
Source |
New records |
What the numbers mean |
|---|---|---|
Baschel et al., 2018, Table 1 |
52 |
Alternative component rates; mixed historical field and literature inputs. 47 pass the internal-consistency screen. |
Lillo-Bravo et al., 2018, Tables 2–3 |
288 |
176 event-count and 112 normalized-rate records, including overlapping totals; 15 fixed-tilt plants, 15 months. |
2 |
Published first-event survival points from 2,121 inverters; not reconstructed individual histories. |
|
1 |
Mean tracker uptime from power-correlation classification, 83 inverter groups / 34 sites. |
|
8 |
Mean/median tracker availability under two tracking windows and two missing-data assumptions. |
|
14 |
Seven individually described weather-station ages and seven sensor drift observations. |
|
6 |
Stuck-row prevalence at two sites on specific satellite observation dates. |
The other seven sources provide taxonomy, access constraints, interpretation or existing data: IEA PVPS tracker report (2024), Lindig et al. (2023), Gallardo-Saavedra et al. (2019), Jordan et al. (2020), Stanchev and Hinov (2026), PVROM 2018 and PVROM 2020.
pv_reliability_sources() records the reviewed scope and access limits of each.
The 371 rows are not 371 distinct devices, and cannot be averaged together.
Publication year is not manufacturing vintage. No current OEM ranking is inferred.
Coverage by device#
Device or subsystem |
Evidence available |
Appropriate use / remaining gap |
|---|---|---|
PV modules |
Field/literature rates, module-event counts, PVROM fits |
Separate discrete failures from gradual power loss and warranty thresholds. |
String / central inverters |
Component rates, fault intervals, first-event survival, categorized events |
Match topology, size, environment and event definition; resets are not replacements. |
Optimizers / MLPE |
BFH survival study and contextual results |
Individual histories are unavailable; no numeric curve digitization or invented microinverter rate. |
Inverter IGBTs, capacitors, cooling fans, control boards, breakers |
Baschel rates and PVROM fault fits |
Use a component model or a complete inverter model; adding both double-counts. |
Combiner / junction boxes (CBX) |
Existing PVROM fits; published box events; disjoint part-rate scenario |
Lillo reports five box events but no box inventory, so no per-box rate is inferred from that count. |
CBX fuses, terminals, monitoring units, DC switches |
Separate component-rate alternatives |
A single branch failure need not disconnect the entire box. |
DC/AC cables, connectors, disconnectors |
Published rate alternatives and wiring events |
Match connector/cable-unit definitions; no unreported cable-length conversion. |
Transformers, switchgear, grid connection |
Rates and station/grid events |
Weather/grid interruptions are distinct from internal transformer hardware failure. |
Trackers: complete rows |
2018/2026 uptime studies, satellite prevalence |
Estimate operational availability; missing angles and model error remain visible. |
Slew drives, motors, gearboxes, bearings, encoders |
IEA failure-mode evidence |
No exposure-based component rate found in the reviewed public material. Log confirmed faults and installed row/component time. |
Tracker controllers, communications and batteries |
IEA mechanisms; tracker operational evidence |
Separate control, communication, power-supply and mechanical causes. No transfer of nuclear battery rates. |
Weather stations (WST) |
Long-term compact-station cases |
Small, unmaintained, non-PV cohort; not a utility met-station fleet prior. |
Temperature / RH sensors |
Published individual drift observations |
Compare with a reference and retain calibration/replacement histories; drift is not failure incidence. |
Anemometers / rain gauges |
Documented wear, corrosion and measurement failure |
Sensor validity and stow availability matter; no universal failure rate derived. |
Pyranometers, reference cells, module-temperature probes |
PV monitoring failure mechanisms / maintenance context |
No defensible exposure-based rate found here. Collect redundant-reference QC, cleaning and calibration records. |
DAQ, PLC and monitoring networks |
Existing PVROM fits and monitoring event counts |
Keep communication availability separate from device hardware reliability. |
Findings that change the analysis#
Inverters: the BFH cohort reports 65.7% first-event survival at year 15 and 59.1% at year 18, with a reported interval of 52.1–67.0% at year 18. Only 15 units remain at risk at that age. This is neither a constant annual failure rate nor a prediction that every inverter must be replaced at year 15. The source’s indoor/outdoor direction conflicts between abstract and body; this extension does not encode a directional multiplier.
Trackers: Martins starts with 64 sites (about 2.1 GWp), rejecting 15 for data quality. The reported median changes from 66% to 87% for all tracking, and 83% to 89% for core tracking, when missing angles are counted as available instead of unavailable. These are scenario results, not confidence limits. The abstract’s 88% upper value conflicts with Section 3.2; the extract uses the section-specific figures. The older 99% benchmark uses different populations and a clear-sky power-based method, so the difference is not a measured deterioration over time.
WST: a station can transmit while individual channels are wrong. The compact station study distinguishes functioning, ceased-transmission and drifting sensors. Its text says six units but names seven, and contains inconsistent VP2/TX labels. The seven named ages remain flagged; the Vaisala observation is greater than 13 years, not an exact failure time. No manufacturer ranking or pooled lifetime fit is made from this heterogeneous sample.
CBX and monitoring: whole-plant ticket frequency is a poor substitute for equipment exposure. Monitoring events may auto-reset, while a small number of power-path failures can affect more generation. Plan work using the affected capacity, expected duration and verified cause as well as event frequency.
Latest SCADA/ML research: health indices, classification scores and anomaly alerts are useful monitoring signals. They need independently verified failures, time-at-risk and prospective calibration before becoming failure probabilities. The 2026 inverter case study is catalogued on this basis, without treating its health/risk indices as physical survival probabilities.
Runnable methods and cases#
The reliability algorithms guide adds a full method selection table, equations and a runnable example for censored Weibull fitting, paired-bootstrap uncertainty, competing causes, redundancy and repair-aware availability/energy loss, alongside the existing Poisson and Bayesian tools.
uv sync --locked --all-packages
uv run python -m energy_analysis.pv_device_reliability --output artifacts/pv-device-reliability
Outputs: typed Parquet/JSON evidence, source registry, quality audit, component-rate scenarios, combiner parts, a labelled synthetic repair/energy sensitivity CSV, summary and a four-panel figure. Three offline gallery examples cover connectors, methods, and analysis.
from energy_data_retrieval.pv_reliability_database import load_pv_reliability_evidence
from energy_analysis.pv_device_reliability import combiner_box_case
rates = load_pv_reliability_evidence(metric="failure_rate")
trackers = load_pv_reliability_evidence(source_id="martins_2026")
parts, case = combiner_box_case(boxes=500)
print(case)
The illustrative combiner case has 24 string fuses, 48 terminal lugs, one string-monitoring unit and one DC switch per box. Using their source rates and 4,015 operating hours/year gives 0.013691 component events/box-year, or 6.85 events/year for 500 boxes. This assumes independent, constant-rate, disjoint components; it is not a whole-box outage, fire-risk or replacement forecast. No outage cost or lost energy is invented. Configuration and source vintage must be reviewed before use for an actual site.
Method |
Input |
Output and boundary |
|---|---|---|
|
Aligned actual/reference angles, POA, interval durations, verified expected-stow flags, optional core-tracking mask |
Duration-weighted availability bounds, observed availability, coverage, and unknown-eligibility hours. Default threshold 5°, configurable. |
|
Channel readings, scheduled interval durations and externally assessed QC mask |
Data completeness versus valid-data fraction; neither becomes a hardware failure count. |
|
First-event/censor ages measured from commissioning, failure flags |
Risk sets, survival, Greenwood log-log pointwise intervals. All survivors included; no unsupported tail extrapolation. |
|
Non-overlapping component rates, integer multiplicities, explicit operating hours |
Expected component events and model probability of at least one event. |
Existing |
Confirmed events/exposure, or failed operations/demands |
Rate estimates and updates; stow-on-demand reliability needs demand counts, not calendar hours. |
For the tracker KPI, POA must exceed zero. Confirmed expected stows count as available following Martins. Equality at the angle threshold is available. Unknown irradiance is reported outside the evaluated denominator. Missing angles widen the bounds; they do not become confirmed mechanical faults. Durations are supplied explicitly: reindex to the complete schedule and mark missing intervals, rather than carrying the last observation across a gap. The optional core mask must reflect the actual geometry/control strategy.
The KM fitter assumes independent right censoring and commissioning-age origin. It does not handle delayed entry, interval-censored failures, repeated events, competing causes or confounding adjustment. Boundary intervals at survival zero or one remain undefined; they are not zero-width guarantees. The method example uses a clearly labelled synthetic cohort, not reconstructed BFH data.
For stow risk, the wind sensor, power supply, network, controller and drive form a dependent chain. Shared failures and storms break an independence assumption. Record failed stow demands separately and use verified outage causes; the CBX independence scenario is not a safety assessment of that chain.
Reproducibility and source-quality audit#
from energy_data_retrieval.pv_reliability_database import refresh_pv_reliability_source
# Install energy-data-retrieval[reliability] first; explicit network calls.
rows = refresh_pv_reliability_source("baschel_2018")
rows = refresh_pv_reliability_source("lillo_2018")
Both connectors limit downloads to 20 MB, verify a registered SHA-256 hash, validate table layout, and return records without overwriting snapshots. Other extracts are reviewed transcriptions, not advertised as live APIs. Original PDFs and restricted records are not redistributed in the repository. Third-party publications retain their own licenses and attribution requirements.
Five Baschel rows are excluded from default scenarios because source MTBF or survival values conflict with the stated rate, or the original reference is missing. Published values remain available. The Lillo extract retains per-park total inconsistencies; the published 10.4 monitoring events/plant-year also differs from 193 / (15 × 1.25) = 10.2933. No silent correction replaces the source. Aggregate rows overlap individual parks and must never be summed with them. A source’s quoted annualized rate cannot supply an unreported confidence interval or hardware replacement count.
To improve these priors with an operating fleet, collect serial/model, component hierarchy, commissioning/removal times, true failure and restoration timestamps, verified cause, event versus reset versus maintenance, observable operating exposure, censoring, environment and firmware. Retain never-failed devices. This provides the denominator that public tracker, WST and CBX event reports often lack, and supports site-specific estimates using the existing rate tools.
For a complete database-to-forecast workflow, see the 14-step failure-prediction walkthrough, including your own CSV inputs, temporal validation and uncertainty.