Reliability methods and algorithms#

Use the PV evidence database for attributed historical benchmarks and your asset register/CMMS for device exposure and verified outcomes. Choose an estimator by the question and available data. Do not combine all published rates into one fleet average: event definitions, clocks and populations differ. The implementation uses NumPy/SciPy/Pandas with no additional dependency.

Implemented algorithm catalogue#

Algorithm / Python API

Required data

Result / PV application

Poisson MLE + exact Garwood interval: estimate_rate

Confirmed events and summed at-risk device exposure

Annual or hourly event rate and uncertainty for inverters, CBX or sensors with a valid denominator.

Zero-event bound: zero_failure_upper_bound

Exposure with no observed failures

One-sided upper rate bound; zero events do not establish zero risk.

Gamma–Poisson update: GammaRate

Comparable prior shape/rate and new event/exposure counts

Posterior rate, credible interval and predictive event probability. Prior strength must be justified separately.

Beta–binomial update: BetaDemand

Failed and successful demands

Failed-stow or failed-switching probability per demand, not per calendar hour.

Kaplan–Meier + Greenwood log-log interval: kaplan_meier_survival

First-failure or censor age for every device

Empirical first-event survival without imposing a lifetime distribution.

Right-censored Weibull MLE: fit_weibull_censored

Positive first-event/censor ages and failure flags

Shape and scale; decreasing, constant or increasing model hazard.

Conditional lifetime risk: LifetimeModel.conditional_failure

A lifetime model, current age and forecast horizon

Probability of first failure in the horizon given survival to the current age.

Paired bootstrap: bootstrap_weibull_survival

Individual first-event/censor histories

Pointwise percentile uncertainty in the fitted Weibull survival curve.

Aalen–Johansen: aalen_johansen

First-event ages with mutually exclusive cause codes

Cause-specific cumulative incidence, accounting for competing first failures.

Independent k-out-of-n dynamic program: k_out_of_n_reliability

Component survival probabilities at the same mission age

Series, active-parallel and threshold redundancy scenarios.

Two-state Markov: repairable_availability

Calendar-up-time failure rate, mean restoration time, time intervals, counterfactual power

Transient/interval/steady availability, repeated-event expectations and expected lost energy.

Tracker and sensor observability: tracker_availability, sensor_serviceability

Complete interval schedule, angles/POA or readings/QC

Time-weighted availability bounds, coverage and valid-data fraction.

Component assembly: assembly_event_scenario

Disjoint part rates, multiplicities and operating exposure

Expected component events in CBX or other assemblies.

APIs are in energy_methods.failure_rates, energy_methods.pv_device_reliability and energy_methods.reliability_algorithms. The algorithm gallery demonstrates the whole workflow with explicitly synthetic observations.

Rates and Bayesian updating#

For independent, homogeneous Poisson events with fixed exposure E, the maximum-likelihood rate is k/E. Include never-failed devices and exclude periods when a failed component is no longer at risk. With zero events, the one-sided 95% upper rate is -log(0.05)/E.

With a Gamma(shape=a, rate=b) prior, new data give Gamma(a+k, b+E). Expected future events in exposure H are H(a+k)/(b+E), while the probability of at least one is 1 - ((b+E)/(b+E+H))**(a+k). The two quantities differ. A published point estimate alone does not identify prior strength. Never count the same underlying field data once as prior evidence and again as new data.

Lifetime fitting, conditional risk and uncertainty#

The Weibull survival model is S(t)=exp(-(t/scale)**shape). The censored log-likelihood sums event log densities and survivor log survival. The fitter profiles out scale and optimizes log shape; it rejects unidentified fits. Conditional first-failure risk is 1-S(age+horizon)/S(age). Shape below/at/above one corresponds to decreasing/constant/increasing hazard; it does not by itself identify a physical failure mechanism.

The bootstrap resamples complete (observed_age, failed) pairs, refits, evaluates the curve and takes percentile bounds. Default 500 resamples and a fixed seed make examples reproducible; increase resamples and check interval stability for an actual decision. Percentile bounds are approximate and can have poor coverage in small samples. They are pointwise, not simultaneous confidence bands, and do not cover model misspecification. Any unidentified resample stops the calculation instead of silently selecting only successful fits. Shared-site/weather cohorts need cluster resampling, which is outside this implementation.

Extrapolating beyond maximum observed follow-up requires explicit opt-in. A plotted 30–50-year Weibull tail remains a model assumption until supported by long-term field data; the two published BFH summary survival points cannot replace individual lifetime observations for fitting.

Competing first-failure causes#

Use 0 for right censoring and positive integer codes for mutually exclusive first-event causes. An unresolved cause must retain a positive “unknown cause” code. At each observed event age, the algorithm adds S(previous)*d_cause/n_risk to that cause’s cumulative incidence, then updates all-cause survival. Survival plus all cause incidences sums to one. Events precede censoring at ties. This implements the first-event special case of Aalen–Johansen, not a general multi-state or recurrent-event model. It requires independent censoring but not independent latent causes. No cause-specific confidence interval is claimed.

For an inverter, cooling, power-stage and control-board first failures may be competing causes. Treating the other causes as simple censoring and reporting 1-KM would answer a different question from observed cause incidence.

Redundancy and repair-aware energy loss#

The k-out-of-n algorithm convolves Bernoulli working-state probabilities and sums the probability of at least k working units. k=n is series; k=1 is active parallel. Inputs must describe the same mission and independent components. One working inverter does not imply full plant output; threshold reliability does not calculate proportional generation. Shared wind sensors, supplies, networks, storm damage, load sharing and standby switching need explicit models.

The two-state Markov model uses up-to-down rate lambda and down-to-up rate mu=1/mean restoration hours. It solves dA/dt=mu*(1-A)-lambda*A and integrates availability exactly over each supplied interval. For initial availability A0:

A(t)=mu/(lambda+mu) + (A0-mu/(lambda+mu))*exp(-(lambda+mu)*t).

Expected lost energy is the sum of counterfactual interval power times expected down-hours. Counterfactual kW represents only affected capacity and is treated as constant within each interval. Expected repeat events are lambda times expected up-hours. Both transition rates use calendar hours in this API. An operating-hour source rate must not be inserted unchanged. The constant-rate, exponential-restoration, independent-power assumptions exclude ageing, weather-correlated failures, constrained shared repair crews and partial derating. Do not add parent and child outage energy for the same interruption.

Run the examples#

uv sync --locked --all-packages
uv run python examples/methods/plot_reliability_algorithms.py
uv run python -m energy_analysis.pv_device_reliability --output artifacts/pv-device-reliability

The study exports synthetic_repair_sensitivity.csv: an assumed 10 MW fleet, 0.04 events per unit calendar-up-year, a repeated artificial daylight power profile, and 8/24/72/168-hour mean restoration scenarios. This demonstrates the energy calculation and is labelled synthetic in the CSV and summary. No repair time, financial price or production profile is asserted to come from a paper.

Algorithm sources and verification#

Tests compare bootstrap bounds with SciPy, redundancy with exhaustive state enumeration, and repairable availability with independent numerical ODE integration. Hand-calculated cause risk sets verify ties and probability mass; the single-cause result reduces to Kaplan–Meier. Earlier methods retain their Poisson/Bayesian/censoring tests. No fitted predictor of field failures is claimed without a held-out fleet cohort and calibration assessment.