ML112770176

South Texas Project, Units 1 and 2 - Licensee Handouts from 10/3/11 Meeting STP LOCA Frequency (TAC Nos. ME5358 and ME5359)

ML112770176
Person / Time
Site:	South Texas
Issue date:	10/03/2011
From:	Mosleh A South Texas
To:	Singal B K Plant Licensing Branch IV
Singal, B K, NRR/DORL, 301-415-301
Shared Package
ML112770162	List: ML112770170 ML112770176 ML112770203 ML112770230 ML112770237 ML112770246 ML112770251
References
TAC ME5358, TAC ME5359
Download: ML112770176 (24)

Text

Technical Review of STP LOCA Frequency EtitiMthdlEstimation MethodologyAli MoslehWork Performed for South Texas Project Electric Generation StationOct 3, 20111 Scope of ReviewLOCA Frequency Estimation Model Interpretation and use of NUREG-1829 information Other key assumptions and computational steps used ypppto generate the numerical resultsTwo rounds of review were done, one on an earlier draft (Early August, 2011), and one on the final report (dated Sept 2011) 2 LOCA Frequency Estimation Model Overview Current report has focused on estimation of the frequency of LOCAs initiated at or near the location of pipe and nozzle weldsFtkillildLOCAdtifiltFuture work will include LOCAs due to pipe failures at other locations and non-pipe related failures in the RCS pressure boundary3 Basic Estimation ModelOverview The model for estimating the frequency of a LOCA of a given size is given by the following equations: P(RF)IF(LOCAx)=miixix refers to the various break size ranges such as those used in NUREG-1829 to describe the 6 LOCA categories. jx=ix=ikkP(RxFik)Iikik=nikik=nikfikNiTi4 Definitions5 Definitions6 Basic Estimation ModelOverview Different locations are characterized and assigned rupture frequencies based on 45 combinations of system type, weld type, degradation mechanisms, and pipe size. The estimation process is a "bottom-up" approach building the LOCA estimates by combining various contributing components, failure mechanisms, and conditional probabilities. 7 Basic Estimation ModelComments and ObservationsThe general decomposition formula follows known principles of calculus of probabilityThe formula represents the commonly made assumption of constant failure rate model for the selected period in plant lifelifeThe formulation enables the consideration of known susceptibilities to the damage mechanisms in quantification of the failure rates and probabilitiesIt allows tailoring the probabilities for component-specific vulnerabilities to degradation mechanisms8 Basic Estimation ModelComments and ObservationsParameter Iik, determined through a previously developed and peer-reviewed Markov model, provides the ability to account for the effects of integrity management programs on reducing likelihood of catastrophic failuresAs a physical model the decomposition formula views the pipe rupture(LOCA)processasatwostageprocessof(1)stochasticrupture (LOCA) process as a two stage process of (1) stochastic occurrence of degraded functional conditions, and (2) stochastic rupture events given a degraded conditionThis is a reasonable first order approximation of a physical process model. Similar formulations are seen elsewhere in PRA such as Precursor Event methodology and some of the commonly used methods for common cause failure analysis analysis9 Basic Estimation ModelComments and ObservationsA key advantage of the approach is the ability to make use of existing operational dataAssumptions in implementation Various degradations (crack, small leak, leak etc) are lumped together as one homogenous class irrespective of their implied severity This means that the same conditional rupture probability applies to all degraded states. The effects is overestimation for some (e.g., crack) and underestimation for other degraded conditions (e.g.,

leak) 10 Basic Estimation ModelComments and ObservationsThe contributions of various DMs at a specific piping location are combined linearly to determine the total LOCA frequency distribution (for each weld /location) This is a first order approximation to a significantly more complexcasewheresynergeticeffectsofmultipleDMsarecomplex case where synergetic effects of multiple DMs are considered explicitly. However, the state of the art in terms of experimental results and theoretical models for the synergistic effects is quite limitedTo some extent the the net effect of synergistic phenomena are accounted for via the use of actual data in the failure frequency estimation estimation stage11 Failure Rate Development Steps 1.1Determination of component and weld types1.2Perform data query for failure counts1.3Estimate component exposure1.4*Develop component failure rate prior distributions for each damage mechanism (DM) 1.5*Perform Bayes' update for each exposure case (combination of weld count case and DM susceptibility case)1.6*Develop mixture distribution to combine results for different exposure hypotheses to yield conditional failure rate distributions given STP specific applicable DMs1.7Calculate total failure rate over all applicable damage mechanisms

• Subject of specific comments on next viewgraph 12 Observations on Failure Rate Development Steps STEP 1.4 Develop component failure rate prior distributions for each damage mechanism (DM) From earlier work in EPRI RI-ISI, peer reviewedVery wide lognormal distributions, based on existing pipe failure rates, engineering judgment, and use of dataSTEP15PfB'dtfhSTEP 1.5 Perform Bayes' update for each exposure case (combination of weld count case and DM susceptibility case)Standard use of Bayesian updating (mostly with 0 failures), well established and accepted in PRAsComputations were done with aid of a validated and widely used computer code13 Observations on Failure Rate Development Steps STEP 1.6 Develop mixture distribution to combine results for different exposure hypotheses to yield conditional failure rate distributions given STP specific applicable DMsUse of (equally weighted) mixture of distributions to generate a composite distribution for different data assumptions is an accepted (advanced) methodology within the Bayesian framework. The report has applied this methodology in a very systematic and rigorous wayThe applied Monte Carlo sampling technique and numerical procedures were not subjects of this reviewWhile the process and assumptions are clearly stated (see for example table 3-1) the scope of this review did not include an evaluation of the technical basis for assumptions made in the process of estimating population size for each class (weld, pipe segment, ect,) and in assessing the level of susceptibility of each class to various failure mechanisms14 Conditional Rupture Probability Estimation MethodologyThe goal is to develop a set of conditional rupture probabilities (CRPs) vs. break size for each component category CRPs are estimated by anchoring the the generic total frequency of each of the six LOCA classes on two distinct sources of NUREG-1829 dataBase case analyses of specific PWR components (hot leg, surge line, HPI line for a specific 3-loop PWR) developed by B. Lydellusing methodology similar to that used by the current study Estimates of 9 experts of LOCA frequencies vs. break size for many PWR components for the entire fleet of U.S. PWRs15 Conditional Rupture Mode Probability Model Steps 2.1 Select components to define conditional rupture probability (CRP) model categories*2.2 Obtain expert reference LOCA distributions from NUREG-1829*2.3 Obtain expert multiplier distributions for 40yr LOCA frequencies from NUREG-1829*2.4 Determine 40yr LOCA distributions (product of steps 2.2 and 2.3) for each expert, fit to lognormal*2.5 Determine geometric mean of expert distributions from Step 2.4 (lognormal)* Subject of specific comments on next viewgraph 16 Observation on CRP Estimation StepsSTEPS 2.2 and 2.3 Some NUREG-1829 experts had provided asymmetric inputs for lower, middle, and upper values, inconsistent with lognormal distribution characteristics NUREG-1829 used "split lognormal distributions" to treat this tThSTPhfitthtittddasymmetry. The STP approach fits the asymmetric cases to standard lognormal distribution in part to simplify the procedure Of the three possible alternatives, this reviewer recommended use of Upper and Mid values to determine the corresponding distributionpreserves the central tendency of the distribution andkeeps the expert's upper bound estimate which carries more risk significance 17 Observation on CRP Estimation StepsSTEP 2.4 Determine 40yr LOCA distributions (product of steps 2.2 and 2.3) for each expert, fit to lognormalThis is done for each expert's estimate prior to aggregation of theresultof9experts(correctapproachamongalternatives)the result of 9 experts (correct approach among alternatives) The procedure follows basis probability rules for developing distribution of product of lognormallydistributed quantities18 Observation on CRP Estimation StepsSTEP 2.5 Determine geometric mean of expert distributions from Step 2.4The report investigated two approaches for forming composite distributions, Mixture Distribution Method (with equal weight for all experts)Geometric Mean Method (by taking the geometric means of two parameters of the experts lognormal distributions, medians and range factors)The Geometric Mean method, which is the approach taken by NUREG-1829 was also adopted by the reportIn this reviewer's opinion, in general the mixture distribution approach to expert opinion aggregation has a stronger technical basis. However, an equally important factor is the engineering assessment by NUREG-1829 and the STP report of the overall suitability of the results produce by each approach. 19 Conditional Rupture Mode Probability Model Steps2.6a Benchmark LydellBase Case Analysis for selected components2.6 b Determine failure rate distribution for LydellBase Case analysis in NUREG-1829; fit to lognormal 2.6c Apply LydellCRP model from Base Case Analysis26dDtiLOCAfditibtifLdllBCAli2.6d Determine LOCA frequency distribution from LydellBase Case Analysis*2.7 Determine mixture distribution of NUREG-1829 GM (from Step 2.5) and LydellLOCA frequency (from Step 2.8) to obtain Target LOCA frequency Distribution for each CRP category component*2.8 Apply formulas to calculate CRP distributions to be used as prior distributions for each component assigned to each CRP category*2.9 For each component in CRP category, perform Bayes update with evidence of failure and rupture counts from service data.* Subject of specific comments on next viewgraph 20 Observation on CRP Estimation StepsSTEP 2.7 Determine mixture distribution of NUREG-1829 GM (from Step 2.5) and LydellLOCA frequency (from Step 2.8) to obtain Target LOCA Frequencydistribution for each CRP category componentThe report has viewed NUREG-1829 expert opinions and Lydell'sapproach as two largely separate sources of yppgypinformation on LOCA frequencies. This reviewer agrees that the overlap of the two sources are less pronounced than the differences in corresponding modeling and quantification approachesOf possible options for use of these two sources of information, the report has chosen a mixture distribution of GM and Lydellresults, following a recommendation form this reviewer. 21 Observation on CRP Estimation StepsSTEP 2.8 Apply formulas to calculate CRP distributions to be used as prior distributions for each component assigned to each CRP categoryThis step is straightforward conceptually, but requires careful numerical (MC) procedures. This review however did not look into the details of computational implementationSTEP 2.9 For each component in CRP category, perform Bayes update with evidence of failure and rupture counts from service data.This step follows standard approach to Bayesian updating (beta prior and binomial likelihood) Updating is done with marginal distributions (beta) of each of the 6 CRP categories given a failure. The joint distribution of CRPs is a multinomial distribution but separate updating with marginal (beta) distribution for each CRP is a reasonable simplification with no visible impact on results give the low values of CRPs22 Summary and ConclusionsComments and suggestion on earlier version have been addressed in the final report. Examples include Addition of a more detailed summary of the methodology and charts explaining the steps BetterexplanationofassumptionsBetter explanation of assumptionsMore consistent application of probabilistic methods in some areas23 Summary and ConclusionsThe modeling and parameter estimation approaches are found to be sound, acknowledging a number of common approximations and certain simplifying assumptions, some imposed by the limitations in the state of the artThe report's careful and comprehensive treatment of known sources of uncertainty as well as relatively broad rages assigned to distributions are expected to cover much of the impact of the assumptions and limitations in the state of the art24