POET Method Basis & V&V

This page documents POET as a newly developed method (NDM) for estimating initiating-event frequencies and data-analysis parameters from operating experience, in the form a PRA peer-review team expects (NEI 17-07, Rev. 2, Sec. 2.3; PWROG-19027-NP, Rev. 2). It is intended to accompany a focused-scope peer review of the method as incorporated into a plant PRA.

In scope: empirical IE frequencies (short/medium/long unplanned transients), duration-based exposure, time-based plant operating-state fractions, refueling frequency/duration, Bayesian credible intervals (plant-only and prior-updated), fleet and NRC industry-average benchmarking, an empirical-Bayes fleet prior, threshold sensitivity, and keyword-based cause attribution. A reference library of NRC/INL component-reliability, common-cause-failure, and LOOP non-recovery data is provided.

Explicitly out of scope: POET is an input aid, not a PRA. It does not build accident sequences, quantify CDF/LERF, infer thermal plant states, perform a complete IE identification and grouping analysis, or replace the analyst's judgment, integration, or documentation.

Parameter estimation (NUREG/CR-6823)

Initiating events are modeled as a Poisson process; the conjugate prior is a Gamma distribution. POET uses the formulas of NUREG/CR-6823, "Handbook of Parameter Estimation for PRA": the Jeffreys noninformative posterior is gamma(x + ½, T); an informative prior updates to gamma(α + x, β + T). The same model underlies the NRC/INL RADS calculator and the NUREG/CR-6928 industry-average estimates.

Empirical-Bayes fleet prior

A fleet population prior is fit by exposure-weighted method of moments with a Poisson sampling-variance correction (parametric empirical Bayes; NUREG/CR-6823, Ch. 8), then conjugate-updated with the plant's data.

Prior/data consistency (SR DA-D4c)

Before any prior is used, POET computes the prior-predictive (gamma-Poisson / negative-binomial) tail probability of the observed count and reports whether the plant data are statistically consistent with the prior (NUREG/CR-6823 Sec. 6.2.3.5).

Duration thresholds (data-derived)

SHORT/MEDIUM/LONG cutoffs are the empirical tertiles of the fleet-wide unplanned-outage duration distribution: P33 ≈ 2 d and P67 ≈ 5 d (n = 2,763), fixed as constants (SHORT ≤ 2 d, LONG > 5 d) for reproducibility. A sensitivity view shows the effect of alternative cutoffs.

Operating-state classification (time-based)

States are assigned from daily percent power and outage records by time elapsed since outage start, not by inferred thermal condition.

• IE proxy. POET's "unplanned outage" frequency is a daily-power-derived proxy for the reactor-trip / general-transient IE rate. Bias is bidirectional: it may overcount (forced maintenance outages that are not trips) and undercount (trips that recover sub-daily). The NRC ratio is a benchmarking indicator, not a taxonomic equivalence.
• Cause attribution is keyword-based and covers only events carrying NRC event text (scram notices, 2011-present); counts are a floor, not a complete IE categorization.
• Empirical-Bayes uses a method-of-moments fit (Kass-Steffey / full hierarchical Bayes are a possible refinement).
• No thermal-state inference and no complete IE identification/grouping (the completeness analysis remains the analyst's responsibility).
• Normalization. Frequencies are per reactor-calendar-year (primary) and per reactor-critical-year (for NRC comparison); both bases are reported.

• Statistical engine validated against NRC published values. POET's gamma quantiles reproduce each NRC baseline's published 5th/95th percentiles from its (α, β): worst relative error 4.82% across 49 initiating events.
• Cross-checks against source data. All computed counts and exposures reconcile to direct database queries (independent harness: 47/47 assertions passed across PWR, BWR, decommissioned, and new-unit cases).
• Invariants enforced. Operating-state fractions sum to 1; exposure = frequency × average duration; empirical-Bayes posterior interval ≤ plant-only interval (shrinkage); percentiles ordered and bounded.
• Edge cases guarded. One-day windows, pre-data windows, zero-exposure and brand-new units, invalid dates, and unknown units are handled without error.
• On-demand self-test. An admin self-test page re-runs these checks live after each data import.
• Reproducibility. Fixed thresholds and constants; NRC baselines regenerated from retained source files by a documented script; every output carries a traceability block.

Before crediting POET outputs in a PRA, the analyst is responsible for:

1. confirming applicability of the data window and peer group to the plant's design and model;
2. reviewing the prior/data consistency result (SR DA-D4c) and selecting an appropriate prior (plant-only Jeffreys, NRC industry, or fleet empirical-Bayes) with justification;
3. confirming IE completeness independently (POET does not perform full IE identification);
4. documenting the inputs, assumptions, and any deviations in the PRA of record; and
5. including the method in the scope of the PRA peer review.

POET supports the Initiating Event (IE) and Data Analysis (DA) elements of the ASME/ANS RA-Sa-2009 PRA Standard, as endorsed (with exceptions and clarifications) by NRC Regulatory Guide 1.200, Revision 3 (December 2020), Appendix A. It is documented here as a newly developed method so that, when incorporated into a plant PRA, it can be reviewed under NEI 17-07, Rev. 2 (Sec. 2.3) by reviewers independent of the method's developer; PWROG-19027-NP, Rev. 2, governs whether its adoption is classified as PRA maintenance or a PRA upgrade. Uncertainty treatment is consistent with NUREG-1855.

POET does not itself "conform" to the Standard — conformance is a Capability-Category determination made by peer review of a plant's PRA. POET provides defensible inputs and the documentation needed to support that determination. A crosswalk of POET outputs to specific IE and DA supporting requirements (ASME/ANS RA-Sa-2009) is provided.

• Thresholds are fixed constants (re-derivable; the live derivation is shown on the Methodology page).
• NRC baselines are regenerated from retained source spreadsheets by a versioned script; the source files are downloadable from the Methodology page.
• Data vintage is stamped on every result and report.
• Outputs are reproducible from the bookmarkable analysis URL (plant, window, peer group).

• ASME/ANS RA-Sa-2009, "Standard for Level 1/Large Early Release Frequency PRA for Nuclear Power Plant Applications."
• NRC Regulatory Guide 1.200, Revision 3 (December 2020), "Acceptability of PRA Results for Risk-Informed Activities."
• NEI 17-07, Revision 2, "Performance of PRA Peer Reviews Using the ASME/ANS PRA Standard."
• PWROG-19027-NP, Revision 2, "Newly Developed Method Requirements and Peer Review."
• NUREG/CR-6823, "Handbook of Parameter Estimation for Probabilistic Risk Assessment."
• NUREG/CR-6928 series / SPAR 2020 Parameter Estimation Update; INL/NRCOE, nrcoe.inl.gov.
• NUREG-1855, "Guidance on the Treatment of Uncertainties Associated with PRAs in Risk-Informed Decisionmaking."
• INL/MIS-22-70252, "U.S. Nuclear Operating Experience Program for PRA Parameter Estimations" (the program that produces the SPAR parameter estimates POET draws on).
• NUREG/CR-7294 (INL/EXT-21-61117) and INL/CON-23-72154, INL work applying AI/ML to U.S. nuclear operating experience, including text classification of events into initiating-event categories — the direction underpinning POET's cause attribution.

Full methodology →   ← Back to the tool