Techniques for uncertainty and sensitivity analyses
In radioactive waste studies, particular difficulties arise in Probabilistic Safety Assessment (PSA), since the distribution function of parameters may vary over several orders of magnitude, means are therefore unstable, and models are strongly non-linear. Uncertainty Analysis based on a Monte Carlo approach is well suited to handle input data uncertainties. Faster convergence of the mean may be achieved by importance sampling or by a Laplace approach. Where the distribution function is unknown, or non-normal, a non-parametric test based on Tchebycheff's Theorem can establish confidence bounds on the mean. Alternatively these may be calculated on the inverse cumulative distribution using the Kolmogorov Statistic. A convergence analysis performed on the results of the Level O PSAC intercomparison of PSA codes found no evidence of systematic errors in any of the results. Four techniques for Sensitivity Analysis (SA) are discussed: Spearman's coefficient, partial correlation and standardised regression coefficents and Smirnov's tests, complemented by input/output scatterplots based on rank. SA for the level E PSAC intercomparison showed that although non-parametric tests are an effective tool, they can lead to erroneous results, due to e.g. non-monotonicity of the model. An expert's judgement is needed to verify the results.
Bibliographic Reference: Paper presented: Risk Analysis in Nuclear Waste Management, Ispra, (IT), May 30-June 3,1988
Availability: Available from (1) as Paper EN 34339 ORA
Record Number: 198910141 / Last updated on: 1994-12-01
Original language: en
Available languages: en