Modelling the Concentration Distribution of Non-Buoyant Aerosols Released from Transient Point Sources into the Atmosphere

Thumbnail Image
Cao, Xiaoying
puff dispersion modelling, Gaussian dispersion
Neural network models were developed to model the short-term concentration distribution of aerosols released from point sources. Those models were based on data from a wide range field experiments (November 2002, March, May and August 2003). The study focused on relative dispersion from the puff centroid. The influence of puff/cloud meandering and large-scale gusts were not considered, the modelling was limited to studying the dispersion caused by small-scale turbulence. The data collected were based on short range/time dispersion, usually shorter than 150 s. The ANN (Artificial Neural Network) models considered explicitly a number of meteorological and turbulence parameters, as opposed to the Gaussian models that used a single fitting parameter, the dispersion coefficient. The developed ANN models were compared with predictions generated from COMBIC (Combined Obscuration Model for Battlefield Induced Contaminants), a sophisticated model based on Gaussian distributions, and a traditional Gaussian puff model using Slade’s dispersion coefficients. Neural network predictions have been found to have better agreement with concentration measurements than either of the other two Gaussian puff models. All models underestimate the maximum concentration, but ANN predictions are much closer to observations. Simulations of concentration distributions under different stability conditions were also checked using the developed ANN model, and it showed that, for a short time, Gaussian distributions are a good fit for puff dispersion in the downwind, crosswind and vertical directions. For Gaussian puff models, the key issue is to determine appropriate dispersion coefficients (standard deviations). ANN models for puff dispersion coefficients were trained and their average predictions were compared with the results of measurements. Very good agreement was observed, with a high correlation coefficient (>0.99). The ANN models for dispersion coefficients were used to analyze which input variables were more significant for puff expansions. Dispersion time, particle position relative to the centroid, turbulent kinetic energy and insolation showed the most significant influence on puff dispersion. The Gaussian puff model with dispersion coefficients from the ANN models was compared with COMBIC and a Gaussian puff model using Slade’s dispersion coefficients. Generally speaking, predictions generated by the Gaussian puff model with dispersion coefficients generated by ANN models showed better agreement with concentration measurements than the other two Gaussian puff models, by giving a much higher fraction within a factor of two, and lower normalised mean square errors.
External DOI