Bayesian updating of atmospheric dispersion models for use after an accidental release of radioactivity

TLDR: In this paper, the authors describe a stochastic Bayes linear model for the forecasting of the spread of the contamination in the event of an accidental release of radioactivity, which is used to guide both short-term counter-measures, such as sheltering and evacuation, and also longerterm countermeasures such as food bans and relocation.

Abstract: We report on the development of the Bayesian forecasting and uncertainty handling components of a decision support for emergency management in the event of an accidental release of radioactivity. In particular, we focus on the forecasting of the spread of the contamination. We describe a simple but novel form of stochastic Bayes linear model. This closely mirrors well-developed (Gaussian) puff atmospheric dispersion models, but admits effective Bayesian learning procedures on the values of uncertain variables. Several features of the model are highlighted and its workings illustrated on (partially simulated) test data. In the event of an accidental release of radioactivity, the dispersal of the contaminated material will be of prime concern in determining appropriate countermeasures. A European consortium, funded by the Commission of European Communities (CEC), is currently developing a decision support system (DSS) to aid emergency management in such events. It is envisaged that this DSS will be used to guide both short-term counter- measures, such as sheltering and evacuation, and also longer-term countermeasures, such as food bans and relocation. At its heart needs to be an atmospheric dispersion model to provide a forecast of the spread of the plume. Currently, such models, despite their complexity, are essentially deterministic. They neither allow a simple means of assimilat- ing monitoring data nor provide realistic measures of the uncertainty in their forecasts. Our role in the consortium is to investigate the feasibility of using the Bayesian metho- dology to provide data assimilation and uncertainty management in the DSS. We have begun by focusing on atmospheric dispersion models. Subsequently, our efforts will be directed to models used to forecast the efficacy of countermeasures and the evaluation of different possible strategies. There are essentially four sources of uncertainty associated with forecasting atmospheric dispersion. * Data concerning actual emissions or the source term: If the release occurs in a 'controlled' way through the stack at a power plant, knowledge of the mass and spectrum of the released contaminants will be quite precise. On other hand, in the event of a Chernobyl-like accident, direct information on the emissions will be very sparse. * Errors inherent in the air and ground contamination readings: Monitoring observations taken to locate the spread of the plume are subject to variability both from observational error (counting statistics and human error) and from the natural heterogeneity of the contamination caused by local geographic features, wind behaviour and precipitation. * Uncertainty associated with the meteorological forecasts: Weather forecasting is an uncertain science. Also, dispersal other than over a flat terrain in stable weather conditions depends on local wind patterns, which in turn depend in complex ways on the local geography. * Modelling error inherent in the use of a particular atmospheric dispersion model: Such models, although extremely complex, will typically fail to incorporate enough contingencies to enable them to work well in all circumstances.