Jonathan Seaquist

Wind is a complex phenomenon and a critical fac-tor in assessing climatic conditions and pedestrian comfortwithin cities. To obtain spatial information on near-groundwind speed, 3D computational fluid dynamics (CFD) model-ling is often used. This is a computationally intensive methodwhich requires extensive computer resources and is time con-suming. By using a simpler 2D method, larger areas can beprocessed and less time is required. This study attempts tomodel the relationship between near-ground wind speed andurban geometry using 2.5D raster data and variable selectionmethods. Such models can be implemented in a geographicinformation system (GIS) to assess the spatial distribution ofwind speed at street level in complex urban environments atscales from neighbourhood to city. Wind speed data, 2 mabove ground, is obtained from simulations by CFD model-ling and used as a response variable. A number of derivativescalculated from high-resolution digital surface models (DSM)are used as potential predictors. A sequential variable selec-tion algorithm followed by all-possible subset regression wasused to select candidate models for further evaluation. Theresults show that the selected models explain general spatialwind speed pattern characteristics but the prediction errors arelarge, especially so in areas with high wind speeds. However,all selected models did explain 90 % of the wind speed vari-ability (R2≈0.90). Predictors adding information on width andheight ratio and alignment of street canyons with respect towind direction are suggested for improving model perfor-mance. To assess the applicability of any derived model, theresults of the CFD model should be thoroughly evaluatedagainst field measurements