Air pollution dispersion modelling of tall stacks that emit infrequently during periods of a year poses various problems for dispersion modellers. One of the most significant issues is that the predictions of peak concentrations become increasingly Susceptible to the prevailing meteorology as the amount of time a source emits during a year decreases. A probabilistic approach has been adopted by randomly sampling the predicted concentrations at different receptor locations over 5 different years. This approach is used to explore how source-operating time affects high-percentile concentration predictions of SO2 from a single stack and a network of four stacks. For locations that are rarely downwind of sources and for low operational times, the inter-annual variability of predicted concentrations is shown to be high. Furthermore, the range of possible concentrations for a particular year is wide and suggests that model results under such conditions could easily be atypical, even when several years of meteorological data are used. The modelling also highlights how the probability distributions are affected by plant operating patterns for two cases: first, where sources are assumed to emit simultaneously and second, where they emit independent of one another. By considering ensembles in this manner, it is possible to derive median predicted concentrations and information concerning the probability of exceeding a certain concentration, thus providing decision makers with a richer source of information. (C) 2002 Elsevier Science Ltd. All rights reserved.