Combining CityGML files and data-driven models for microclimate simulations in a tropical city

Abstract

This research first developed a solution to convert LOD 2 CityGML files to LOD 1 files. Then, by using LOD 1 files as geometric inputs, this study developed a python-based microclimate simulation program based on a customized machine learning algorithm to better understand the relationship between urban morphology and outdoor temperature. The integration with the 3D geometry makes this prediction model not only useful to the scientific research, but also as an urban planning tool for environmentally sustainable design. For model development, meteorological data and urban morphology features were used to predict the daily maximum temperature, daily minimum temperature, daily daytime average temperature and daily night-time average temperature. Voting regression (VOT) based on ordinary least square and random forests was used as the regression model. The average CV-R2 and the CV-RMSE predicted by VOT for the outdoor air temperature were 0.84 and 0.52°C. Compared with ordinary least squares, VOT improved R2 by an average of 18% and RMSE by an average of 24%. Sensitivity analysis indicated the daytime temperature was inversely proportional to the aspect ratio of urban canyon streets. In tropical climates, greenery can be used as a cooling measure. A negative logarithmic relationship was found between the daily maximum temperature, daily minimum temperature and green plot ratio. This shows that as the green coverage increases, the cooling effect of greenery increases, and it is more effective to adopt greenery early in areas with less vegetation.