Uncertainty Estimation of an Air Pollution Model

This paper addresses the problem of uncertainty analysis of the predictions of an atmospheric dispersion model. The proposed method uses Bayesian conditioning of the predictions on the available observations and enables consideration of the uncertainties which influence model predictions such as uncertainty of observations and limitations of the model. As an illustration, the proposed methodology is applied to a simple, Gaussian short range atmospheric model (Clarke, 1979). This is commonly referred to the R91 methodology, after the name of the report on which it is based, and shall be referred to as such throughout this paper. The choice of the model was dictated by its availability, simplicity and widespread use. However, the proposed approach has much wider application and can be used for different air dispersion models such as UKADMS (Carruthers et al., 1992) or any other model with comparatively short times of computations and for which observations are available. The simplifications used in dispersion modelling are very substantial and comprehensive data are very limited. This means that it is important to make best possible use of the information which is available -in the form of distributions, value ranges, specialist opinions and common sense. Also many uses in nuclear safety rely on the concept of risk and it is important that uncertainty modelling is applied at all.