Evaluation of fast atmospheric dispersion models in a regular street network

The need to balance computational speed and simulation accuracy is a key challenge in designing atmospheric dispersion models that can be used in scenarios where near real-time hazard predictions are needed. This challenge is aggravated in cities, where models need to have some degree of building-awareness, alongside the ability to capture effects of dominant urban flow processes. We use comprehensive large-eddy simulation (LES) and wind-tunnel data of flow and dispersion in an idealised urban canopy to highlight important dispersion processes and evaluate how these are reproduced by representatives of the most prevalent modelling approaches: (1) a Gaussian plume model, (2) a Lagrangian stochastic model and (3) street-network dispersion models. The poorest model performances were associated with mean over-predictions of concentrations of approximately a factor of 2 and with a relative scatter larger than a factor of 4 of the mean, corresponding to cases where the mean plume centreline deviated significantly from the LES and wind-tunnel observations. This was linked to a low accuracy of the underlying flow models/parameters that resulted in a misrepresentation of pollutant channelling along streets and of the uneven plume branching observed in intersections. Model performances greatly improved by increasing the accuracy of the driving flow field. When provided with a limited set of representative velocity parameters, the comparatively simple street-network models performed equally well or better compared to the Lagrangian model run on full 3D wind fields. The study showed that network models capture the dominant building-induced dispersion processes in the canopy layer through parametrisations of horizontal advection and vertical exchange processes at scales of practical interest. At the same time, computational costs and computing times associated with the network approach are ideally suited for emergency-response applications