NUMERICAL ILLUSTRATION: EFFECTS OF UNCERTAINTY AND BIAS ON RISK PREDICTIONS

摘要

Risk predictions determined from performance assessment analyses are used to evaluate whether a site meets regulatory requirements. For example, risk predictions based on performance metrics for a high-level radioactive waste repository at Yucca Mountain, Nevada, were proposed by the U.S. Environmental Protection Agency in 40 CFR Part 197 and the implementation of that proposed rule presented by the U.S. Nuclear Regulatory Commission in 10 CFR Part 63. Performance metrics include peak-of-the-mean, mean-of-the-peaks, quantiles, and cumulative release, and may be affected differently by uncertainties and model biases. In analytical chemistry, spiked samples are used to study the accuracy and precision of chemical analyses. We apply a similar approach to Monte-Carlo based probabilistic risk assessment in order to ascertain how well the most commonly used probabilistic technique can reproduce the risk from a nominal scenario. The stability and accuracy of these performance metrics in relation to the nominal risk at a range of uncertainties and model biases were investigated in previous analyses. The previous analyses concluded the least stable and least accurate risk predictions were from the peak-of-the-mean metric, whereas the most stable and the most accurate risk predictions were from the cumulative release metric. Furthermore, risk dilution (i.e., a decrease in the predicted risk with increased uncertainty) was exhibited in the peak-of-the-mean metric. A criticism of our earlier work is that the observations may have been valid only for the model repository considered for the nominal case. In this work we show the effects of bias for single parameters (i.e., release rate and groundwater travel time) on the predicted risk. The results demonstrate that, within the bounds of what we have tested to date, the conclusions summarized above and in previous work, are more general and not a product of the model system chosen.