No abstract

No abstract

Atmospheric transport models and observations from monitoring networks are commonly used aids for forecasting spatial distribution of contamination in case of a radiological incident. In this study, we assessed the particle filter data-assimilation technique as a tool for ensemble forecasting the spread of radioactivity. We used measurements from the ETEX-1 tracer experiment and model results from the NPK-Puff atmospheric dispersion model. We showed that assimilation of observations improves the ensemble forecast compared to runs without data assimilation. The improvement is most prominent for nowcasting: the mean squared error was reduced by a factor of 7. For forecasting, the improvement of the mean squared error resulting from assimilation of observations was found to dissipate within a few hours. We ranked absolute model values and observations and calculated the mean squared error of the ranked values. This measure of the correctness of the pattern of high and low values showed an improvement for forecasting up to 48 h. We conclude that the particle filter is an effective tool in better modeling the spread of radioactivity following a release.

No abstract

In the case of a nuclear accident, forecasting the spread of contamination is important for determining whether contamination thresholds are exceeded. In this study we explore ensemble modeling for forecasting threshold exceedance. This involves defining probability density functions for the most import model drivers and parameters and creating an ensemble of models by drawing from them. We test two ensemble modeling techniques, a simple Monte Carlo simulation (MC) and the particle filter. The particle filter extends on MC by assimilating observations into the model as they become available in real-time. In this paper we show that using a deterministic model run provides a false sense of accuracy. Using ensemble modeling we can visualize the uncertainty in threshold exceedance by classifying the 95% prediction interval at each grid cell relative to the threshold into either higher, lower or not distinguishable. In addition, we classify the grid cells relative to 4 multiples of the threshold (0.5, 1, 2 and 4), showing the sensitivity of the classification. Large changes between multiples indicate a small prediction interval. By comparing MC to the particle filter we observe a reduction by a factor of up to 10.6 in uncertainty in the PDF of the spread of contamination. We also aggregate the results to the level of a municipality, which might prove more informative to decision makers. Finally, we demonstrate that errors in the PDFs of the most important model settings can degrade the performance of the particle filter.