POET Methodology

POET converts observed reactor operating data into empirical inputs for Probabilistic Risk Assessment (PRA): initiating event frequencies, duration-based exposure, plant operating-state fractions, planned-outage statistics, peer benchmarking, and comparison to NRC industry-average baselines. The objective is to replace generic assumptions with traceable, reproducible, plant-specific values.

What POET does not do: it does not infer thermal plant conditions (hot shutdown, cold shutdown, mid-loop, and so on). Those cannot be determined from daily percent-power data. Instead, shutdown phases are defined strictly by time elapsed since outage start. State labels (EARLY, MID, EXTENDED) are time-based approximations, not thermal states.

NRC source files are retained for provenance and downloadable: ParameterEstimates2020.xlsx, initiating-events-summary-2023.xlsx, LOOP-events-summary-2023.xlsx, CCFParameterEstimates2020Rev1.xlsx. Original data: nrcoe.inl.gov.

The INPO Sustained Fuel Reliability values shown in POET are an example dataset for illustration. INPO performance-indicator data belongs to its members, so the figures are presented as a sample and the raw source file is not downloadable. For current values specific to your own fleet, import your INPO PIC export of the Sustained Fuel Reliability dataset, grouped by owner.

Each outage is an unbroken run of zero-percent-power days. Outages carry a type derived from the NukeWorker outage pipeline:

• Scheduled (planned refueling and major projects) is treated as REFUELING.
• Unscheduled and Scram are treated as unplanned (forced) outages and form the basis of initiating event frequencies.

Unplanned outages are bucketed by duration into SHORT, MEDIUM, and LONG. These three classes drive the initiating event and exposure outputs.

SHORT / MEDIUM / LONG thresholds are derived statistically from the empirical tertiles of the fleet-wide unplanned-outage duration distribution (Unscheduled + Scram, full history), then fixed as constants for reproducibility.

Fixed thresholds in use:

The upper break (5 days) coincides with the existing forced-short versus forced-extended operational boundary in the outage database; the lower break (2 days) coincides with the EARLY_SHUTDOWN state-phase boundary. The live re-derivation above currently reproduces the fixed constants.

Every observed day with non-null power is assigned exactly one state:

IF day is within a Scheduled outage:
    STATE = REFUELING
ELSE IF day is within an Unscheduled or Scram outage:
    elapsed = days since outage start
    IF elapsed <= 2:  STATE = EARLY_SHUTDOWN
    ELSE IF elapsed <= 7:  STATE = MID_SHUTDOWN
    ELSE:             STATE = EXTENDED_SHUTDOWN
ELSE (not in an outage):
    IF power > 90%:   STATE = FULL_POWER
    ELSE:             STATE = TRANSITION

EARLY approximates high-activity early-outage conditions, MID intermediate conditions, and EXTENDED long-duration stable shutdown. These are explicitly not hot or cold shutdown labels.

Two exposure bases are computed from the daily power record over the analysis window:

• Reactor-calendar-year (rcaly) = (count of observed non-null power days) / 365.25. This is the primary normalization base for frequencies, exposure, and refueling.
• Reactor-critical-year (rcry) = (count of days with power greater than 0) / 365.25. Used only for the NRC comparison, because NRC initiating event frequencies are reported per reactor-critical-year.

State fractions use observed days directly: each state's day count divided by total observed days, so the six fractions sum to 1.

For each unplanned class, frequency is the event count divided by the calendar exposure:

IE-SHORT-TRANSIENT  = (count of SHORT unplanned outages)  / rcaly
IE-MEDIUM-TRANSIENT = (count of MEDIUM unplanned outages) / rcaly
IE-LONG-OUTAGE      = (count of LONG unplanned outages)   / rcaly

Events are attributed to the window by outage start date. IE-ALL-UNPLANNED is the total forced-outage frequency.

Exposure is total class outage-days per reactor-year, which equals frequency times average duration:

EXPOSURE-X = (sum of durations of class-X outages) / rcaly
           = IE-X frequency  x  average class-X duration   (days per reactor-year)

Refueling statistics: REFUEL-FREQ = (count of Scheduled outages) / rcaly, and REFUEL-AVG-DAYS = mean Scheduled outage duration.

Initiating-event occurrences are modeled as a Poisson process with rate λ. The conjugate prior is a Gamma distribution, so the posterior is also Gamma and closed-form. For x events observed over exposure T (reactor-years):

Jeffreys (plant data only):   posterior = Gamma(x + 0.5, T)
NRC industry prior:           posterior = Gamma(alpha_NRC + x, beta_NRC + T)

POET reports the posterior mean and the 5th / 95th percentiles (a 90% credible interval) for every frequency. This is the gamma-Poisson conjugate model of NUREG/CR-6823 ("Handbook of Parameter Estimation for PRA"): for the Poisson rate, NUREG/CR-6823 gives the Jeffreys noninformative prior as gamma(½, 0), so the posterior is exactly gamma(x + ½, T) — the formula POET uses — and the conjugate update of an informative gamma prior is gamma(α + x, β + T). It is the same model used by the NRC/INL RADS calculator and the NUREG/CR-6928 industry-average estimates.

Median, error factor, and a frequentist cross-check

Alongside the posterior mean and 5th/95th, POET reports the posterior median and an error factor, EF = sqrt(95th / 5th). Many PRA models carry a frequency as a lognormal distribution defined by its median and EF, so a POET frequency can be dropped straight into a basic event as median applied as ×/÷ EF (this is exactly what the "PRA parameters" export provides, next to the Gamma α/β for tools that take a gamma directly). For comparison, POET also reports the exact (Garwood) frequentist Poisson 90% confidence interval, computed from the chi-square / gamma relationship; it is shown in the report and the exports next to the Bayesian credible interval so the analyst can see both.

Bayesian update against the NRC prior

The NRC SPAR dataset publishes the actual prior parameters (α, β) for each initiating event. POET uses them to update the General Transient industry prior with the plant's own operating experience, yielding a plant-specific posterior. This pulls a sparse-data plant toward the industry distribution and tightens the interval as plant exposure grows, exactly as RADS does. The update inherits the General Transient proxy caveat (see section 12).

Validation

The gamma machinery is validated by reproducing each NRC baseline's own published 5th and 95th percentiles from its (α, β). Across all 49 NRC initiating events, the worst relative error is 4.82% (most are well under 1%; small residuals reflect rounding in the published tables and a few rows that use external priors such as NUREG-1829 for LOCAs).

Empirical-Bayes fleet model

Beyond the NRC prior, POET fits its own population prior from the fleet. A Gamma prior for the all-unplanned rate is estimated by exposure-weighted method of moments with a Poisson sampling-variance correction: the pooled mean is the prior mean, and the between-plant variance is the observed weighted variance of per-unit rates minus the average within-unit Poisson sampling variance. The fitted prior is then conjugate-updated with a single plant's data, giving a fleet-informed posterior that shrinks sparse-data plants toward the fleet and tightens as exposure grows. If the fleet shows no over-dispersion beyond Poisson sampling, a weakly-informative prior is used and the result is flagged. This is the parametric empirical-Bayes / moment-matching approach of NUREG/CR-6823, Chapter 8; POET uses a simple method-of-moments fit (the rigorous Kass-Steffey and full hierarchical-Bayes variants in NUREG/CR-6823 are a possible future refinement).

Initiating-event cause attribution

To narrow the general-transient proxy, POET classifies the actual cause of each event from its NRC event description using a transparent, priority-ordered keyword lexicon (specific causes such as loss of offsite power, loss of feedwater, loss of condenser heat sink, turbine trip, and loss of reactor coolant flow are matched before generic reactor-trip phrases). Every classification is traceable to a matched phrase. Causes that map to an NRC category can then be compared directly to the NRC baselines instead of only through the duration proxy. Coverage is the subset of events that carry NRC event text (scram notices, 2011-present), so cause counts are a floor, not the total. This is consistent with INL's AI/ML operating-experience work (NUREG/CR-7294; INL/CON-23-72154), which classifies event text into initiating-event categories; a trained-classifier upgrade (e.g. TF-IDF with a supervised model) is a possible future enhancement to the current explainable keyword rules.

Reference data library

POET also packages the NRC/INL industry-average reference parameters used as PRA starting points: component reliability (failure probabilities per demand and rates per hour, with distributions), common-cause failure alpha-factors, and the empirical LOOP offsite-power non-recovery curve (probability that offsite power is not yet restored by a given time). All are reproduced verbatim from the public SPAR datasets.

Fuel reliability

POET shows a sample of the INPO Sustained Fuel Reliability index per unit: a graded performance index on a 0 to 10 scale (10 is best) that reflects fuel-cladding integrity sustained over the last two operating cycles. Because it is a graded health index and not a Poisson count, POET deliberately does not give it the gamma-Bayesian rate treatment used for initiating events. Instead it is presented as a benchmarked metric: the unit's value and grade, its monthly trend over the sample window, and its standing against the NukeWorker-tracked U.S. fleet (median, share of units at the clean 10.0 ceiling, and the unit's percentile when below the ceiling). The INPO color and peer quartile are shown as reported; the U.S.-fleet percentile and range are computed by POET over the mapped operating units. For PRA this informs fuel-barrier and source-term judgments and enterprise-risk screening rather than an initiating-event frequency. New units carry no score until about two fuel cycles of operation accrue, and the index is reported, not modeled, so no curve fitting is applied to it. The values shown are an example dataset; INPO members can import their own PIC export (grouped by owner) to use current, fleet-specific figures.

Peer groups can be the entire U.S. fleet, the same reactor type, the same NSSS vendor, the same owner (utility), or a custom filter (reactor class, vendor, region, age range). Peer values are pooled aggregates (total events divided by total exposure across the group), which is the defensible estimator, rather than a mean of per-plant ratios. The peer group includes the selected plant.

The transient multiplier expresses the plant's combined short-plus-medium frequency relative to the peer-group pooled average:

TRANSIENT-MULTIPLIER = (plant IE-SHORT + plant IE-MEDIUM)
                       / (peer  IE-SHORT + peer  IE-MEDIUM)

A value of 1.00 means the plant is at the peer average; greater than 1 means more frequent short and medium transients than peers.

NRC rows other than General Transient (Loss of Offsite Power, Loss of Condenser Heat Sink, Loss of Main Feedwater) are shown as reference baselines with no plant value, because POET cannot attribute outages to those specific causes from daily power data. NRC frequencies are reproduced verbatim from the SPAR 2020 Parameter Estimates and are reactor-class specific (BWR or PWR) where the NRC distinguishes them.

Every output is reproducible from the cited data window, classification logic, and fixed thresholds. Each result carries a traceability block (time range, observed data range, plants included, exposure basis, thresholds, data source, generation timestamp). Exported files embed the same metadata.

POET is designed to support risk-informed work consistent with NRC Regulatory Guide 1.174 (use of PRA in risk-informed decisions) and RG 1.200, Revision 3 (December 2020), which endorses, with exceptions and clarifications, the ASME/ANS RA-Sa-2009 Level 1/LERF PRA Standard (Appendix A). POET supports that standard's Initiating Event (IE) and Data Analysis (DA) elements by supplying empirical, plant-specific initiating event frequencies and operating-state exposure; the parameter-estimation methods follow NUREG/CR-6823, "Handbook of Parameter Estimation for Probabilistic Risk Assessment," and uncertainty is treated consistent with NUREG-1855. POET provides inputs and traceability, not a PRA: it is the analyst's responsibility to confirm applicability to the plant's design and model, perform prior-data consistency checks where NRC or fleet distributions are used as priors, and document all assumptions in the PRA of record.

Back to the tool

A typical workflow for justifying a plant-specific initiating-event input:

1. Select the plant and a defensible window. Use a window long enough for a stable estimate (POET defaults to the most recent 10 complete calendar years). Decide whether to include decommissioned units in peer groups (default: operating only).
2. Read the point estimates (IE-SHORT/MEDIUM/LONG-TRANSIENT, EXPOSURE-*, plant-state fractions, REFUEL-FREQ) and their Bayesian credible intervals. Use the posterior mean and 90% interval, not the bare point value, for a small number of events.
3. Choose the prior. For a plant-specific estimate you can present either the Jeffreys (plant-only) posterior or a fleet-informed posterior (NRC industry prior, or the NukeWorker empirical-Bayes fleet prior). Document which prior you used and why.
4. Benchmark. Report the transient multiplier and fleet percentile to show where the plant sits relative to peers; use the threshold-sensitivity table to demonstrate robustness to the duration cutoffs.
5. Cross-check against NRC. Compare to the NRC/INL industry-average baselines, and use the cause-attribution breakdown for category-specific comparisons (treating the text-coverage counts as a floor).
6. Export with traceability. Download the full report (PDF/Word) or the data files; each embeds the window, fleet definition, exposure basis, thresholds, data source, and timestamp, plus a ready-to-paste citation.

Example citation: "Initiating event frequencies were derived from operating experience using the NukeWorker PRA Operating Experience Tool (POET), [plant], analysis window [range], generated [date]; NRC industry-average values per the SPAR 2020 Parameter Estimation Update (NUREG/CR-6928 series), nrcoe.inl.gov."

POET supplies empirical inputs and traceability; it does not replace analyst judgment. Confirm applicability to the plant's design and PRA model, perform prior-data consistency checks (ASME/ANS RA-Sa-2009 DA-D4c) where NRC distributions are used as priors, and document all assumptions in the PRA of record.