Modelling the Concentration Distributions of Aerosol Puffs Using Artificial Neural Networks

A neural network model was developed to predict the short-term (<150 s) concentration distributions of aerosols released from point sources over very short time periods (approximately 2 s). The model was based on data from field experiments covering a wide range of meteorological conditions. The study focused on relative dispersion about the puff centroid, with puff/cloud meander and large-scale gusts not being considered. The artificial neural network (ANN) model included explicitly a number of meteorological and turbulence parameters, and was compared with predictions from two Gaussian-based puff models to the measurements of four independent trials representing different stability conditions. The performance of the neural network model was comparable (in stable conditions) or better (in unstable and neutral conditions) than these two models when high concentration predictions were considered. Simulations of concentration distributions under different stability conditions were also generated using the developed neural network model, with the result that Gaussian distributions provided good descriptors for puff dispersion in the downwind and crosswind directions, and for particles close to the centroid in the vertical when dealing with short dispersion times.