Abstract

The vertical pollutant dispersion is quite sensitive to the eddy diffusivity, KV. Therefore, good estimations of KV are essential for improving the predictive performance of Eulerian dispersion models; especially in urban areas where literature based KV correlations are not always accurate. Here, we present a methodology to obtain a more accurate, but site-specific, KV correlation. It is based on using artificial neural networks (ANN) to find the best KV function for a particular urban area by minimizing, in a least-squares sense, the difference between ambient measurements of carbon monoxide and dispersion simulations of this tracer species. The resulting ANN-KV correlation is a function of three parameters namely, the stability parameter (z/L), the height within the mixing layer (z/h), and the scaled height (zfC/u*)-hence the Monin-Obukhov (L), mixing (h) and Ekman (u*/fC) lengths are used to predict KV across the atmospheric boundary layer. We then assess how such an ANN-KV model improves the capability of a dispersion model (CAMx) to predict peak concentrations of ambient carbon monoxide in a large city. The evaluation has been performed with a set of eight air-quality meteorological stations evenly spread across the city of Santiago, Chile, during springtime. Results show that with the ANN-KV model, CAMx achieved better predictions of peak CO concentration levels than has been hitherto possible. Typically root-mean-square errors are reduced to half their original values. The resulting ANN-KV model - without any additional training - was then used to predict CO ambient concentrations at another period (summertime) and also to predict ambient concentrations of total carbon (PM 2.5) at both periods. A much-improved agreement was observed. Furthermore, the ANN formulation allowed for the quality of the urban emission inventory to be critically assessed indicating that the weekend emissions in Santiago are most likely underestimated. © 2005 Elsevier Ltd. All rights reserved.