Pollution Transport Simulation and Machine-Learning Aided Source Detection in Metropolitan Areas

Abstract

Air pollution is one of the world's largest environmental health threats. This study aims to use simple remote signals to locate the source of a pollution release, which will significantly enhance our readiness to counter its threat. In urban areas, the flow structures advecting the pollution are extremely complex: boundary layer separation generates vortical structures that increase the spread of pollutants and break the plume into smaller patches by dispersion effect. Furthermore, the blocking and mixing in urban areas make it more obscured to locate the pollution sources. Flow structures were obtained by solving the two-dimensional NavierStokes equations using Computational Fluid Dynamics in a simplified scenario with imaginary urban architectures. We applied the canonical neural network to relate characteristics in the remote pollutant detector signals to the actual location of the pollutant release. The proposed algorithm identifies the source location and its uncertainty through a Monte Carlo analysis. When the number of training samples is small, as limited by the number of trial-releases we can perform in reality, data augmentation is done by introducing noisy measurement as new training samples. While the source localization is reasonable in the cross-flow direction, it is much harder to locate the streamwise location of the source due to signal similarity. The data augmentation technique we applied reduced the uncertainty of the source location by introducing under-fitting phenomena into the model. Furthermore, sensors away from the center line of the flow outperforms the ones near the center line, especially for detecting off-center sources. This indicates a pronounced effect of blocking and mixing right behind the building on the center line, which blurs sensor measurements from different source locations, and thus hinders the ability to trace back to the source.