The real-time prediction of pollution episodes in urban areas close to industrial plants is of great importance for regulatory control, so that emissions can be reduced when high pollution levels are observed. For this purpose, an 'on-line' practical model must be used with a limited number of input meteorological data. However, two questions arise. What are the ability and the accuracy of the available models? How should the models be used for this regulatory purpose? For the latter question, there is the problem of the knowledge and integration into the procedure of large-scale meteorological forecasts; this means that one must know in advance the meteorological conditions over the urban zone of interest. This paper provides a brief account of a comparison between predictions made with a Gaussian model and concentration measurements obtained by the air pollution network AIR NORMAND. Particular attention has been paid to the determination of the dispersion parameters. A comparison is made between the use of Pasquil-Gifford stability classes and the use of wind fluctuation measurements as input data.