MODEL VALIDATION
The retrospective assessment of doses and risks necessarily involves the interplay of observations or measurements at specific times or locations and a construct of the events that presumably underlie the observations. This construct, or model, is used to predict what happened at times or locations for which direct observations are not available or are inadequate to provide reasonable assurance that observed values are appropriate. Various potentially applicable models can be envisaged a priori in any given situation, but the choice of the appropriate one centers on accuracy and precision in predicting measured values and hence, presumably, values that are necessary but not available. Testing of such reliability is known as model validation, and it is central to all reconstructions. In the end, model validation requires a comparison of predicted values with independent measured values, within the limits of uncertainty for both. Well-established measurements are often referred to as “benchmarks.” In the present case, few relevant measurements were available for either the airborne or the liquid pathways, and the extent of model validation can be considered only qualitative.
Uranium, thorium, radium, plutonium, neptunium, cesium, ruthenium, technetium, and strontium radionuclides were all supposedly considered in the mathematical model describing transport of these radionuclides into groundwater, air, and soil (for example, see table 11 of the RAC report for groundwater). However, none of the fission products were actually followed in any detail after a screening assessment showed them to be of minor consequence. Multiple site-specific receptor locations were examined, and the examination considered the multiple release points within the FMPC production area, the physical and chemical characteristics of the release sources, the rate of diffusion as a function of the distance between the site-specific receptor locations and the multiple release points, the size distribution of the wet and dry particles, the resuspension into air of material previously deposited on the ground, the runoff and leaching of material deposited on soil, the rate of radioactive decay, and the biologic pathways through exposed persons. The atmospheric-dispersion models used to estimate ground-level air concentration at specific locations
incorporated local meteorologic data available after August 1986, when a meteorologic tower began full operation. To account for the meteorologic conditions during the years before 1987, 4 types of approximations were used to provide information about wind conditions during the period from the beginning of operation of the plant until 1987. It was concluded that the observed data from the meteorologic tower after August 1986 were the most representative of all those before 1987. The mean wind speed at the Cincinnati airport was not considered reliable for use in the model because direct comparison indicated mean wind speeds frequently twice as high as the measurements made at FMPC.
The short period for which actual meteorologic records were made is a severe limitation. Nevertheless, the comparisons of other sites and the use of simulated environmental conditions afford a degree of confidence that there are no large errors in the meteorologic data or in the atmospheric-dispersion models used.
To compare model-predicted and observed concentrations of uranium, the differences between the 2 values at each of multiple sites were expressed as a predicted:observed (P/O) ratio, where
geometric bias (all sites) = exp [∑ ln (Pi/Oi)/n],
in which the summation is across all times and sites and Pi is the predicted concentration at the location at time i, Oi is the observed concentration, and n is the number of locations or times.
Whenever parameter values are estimates rather than measurements, the use of mathematical models leads to larger uncertainty in the calculated results. Such uncertainty also results from errors in observed measurements and in the extrapolation from measurements to predictions of missing measurements. Uncertainty is expressed by calculating the range of releases, doses, or risks of health effects. The results have been expressed as the 5th to 95th percentiles; that is, 90% of RAC's predicted values fall within these limits. The 5th and 95th percentiles describing the release of uranium span a range of a factor of 5 in the estimates in the case of soil deposition of uranium. For example, on page C-8 of volume II, the releases are 720,000 kg (95%) and 130,000 kg (5%), a factor of 5-6. That was also the case with
respect to uncertainty in the estimation of airborne releases of uranium. For another example, the validation results for uranium were as follows (table 17):
Time |
Geometric bias |
Uncertainty in bias |
|
Air (perimeter) |
1958-1971 |
1.0 |
0.6-1.8 |
Air (boundary) |
1972-1988 |
1.0 |
0.6-1.6 |
Gummed film |
1954-1964 |
0.4 |
0.3-0.7 |
Soil |
1959, 1971-1988 |
1.1 |
0.7-1.7 |
Modeling the deposition of uranium on the ground was more difficult than modeling air concentrations. Observations based on gummed-film monitoring were compared with predicted values. Deposition predicted for 1954-1956 was in poor agreement with observations, but deposition predicted for 1957-1964 was in better agreement with observations made at that time.
It was possible to compare measurements of uranium in the soil and water with those predicted by the mathematical models, which of necessity included best estimates and approximations of parameters in the model. The release estimates were based, for example, on production information. An example of validation is the comparison of the uranium air-monitoring results from 1986-1988 with the concentrations predicted by the model. Most of the environmental monitoring from which data are available for the early days of the operation of the plant involved only uranium.
Differences between the recent calculations and those reported in 1986 by Stevenson and Hardy were adequately addressed and can be accounted for by the differences in the conditions examined, for example, the depth of soil that was considered. Stevenson and Hardy (1986) considered the top 5 cm of soil, whereas RAC considered the top 10 cm.
Measurements of radon at boundary monitoring stations before and after the silo domes were sealed were proportional to predicted decreases based on the model (see page 69 of the RAC report). Unfortunately, few observation data were available for validation of the release of radon from the silos during the periods of highest releases. From 1980 until the present, the observed radon concentrations in air at multiple boundary air-monitoring stations
were similar to those calculated with the radon-dispersion model. For 1986-1991, hourly measurements of radon in air agreed well with predicted large fluctuations related to time of day.
A recapitulation of uncertainties is given in volume II figure M-9, covering both uranium and radon concentrations in air. Well-known statistical methods, some of recent origin, were used in the assessments. For example, Gaussian plume-dispersion models, representation of uncertainties with probability distributions, and Monte Carlo techniques were used.
In general, for both uranium and radon in soil, air, and water, the agreement between predicted and observed concentrations was acceptable. RAC's overall uncertainty in the validation of the releases was a factor of about 5.