Application of artificial neural networks to modeling the transport and dispersion of tracers in complex terrain

Simulation of the transport and dispersion of chemical tracer in complex terrain has been performed using artificial neural networks (ANN). The ANN method has been applied to relatively high temporal resolution data (hourly averages—long time series), and lower-resolution data (daily averages—short time series). The meteorological input consisting of surface and upper-air data was selected in such a way that it optimally represents the spatial inhomogeneity of the flow field, atmospheric stability, and synoptic conditions. In both cases, the inclusion of previous tracer concentrations as input has significantly improved the ANN performance. For the daily average case, several isolated single-point sharp peaks that were recorded in the series of daily concentrations were not resolved by the ANN. An improved correlation with measurements (from 0.946 to 0.997) was obtained after simple smoothing of the tracer concentrations. Because the number of data samples was small, a “leave-one-out” method was used. The hourly averages provided more cases and thus more significant input for ANN training; however, it brought more uncertainty into the selection of appropriate inputs because of the transport time due to the separation between the source and receptor. Here, training was performed using the first 85% of cases; the rest was used for testing. The ANN-simulated hourly concentrations agreed well with the measured concentrations and yielded correlation coefficients for the training and testing sets of 0.844 and 0.896, respectively. The sensitivity analysis revealed that previous concentration data contributed to resolving peaks in simulated concentrations while meteorological data provided more information on the temporal characteristics of the simulated tracer concentrations. A rudimentary comparison with traditional statistical methods revealed that the ANN performed better and showed fewer limitations as a tool for tracer modeling, especially for long-term prediction.