Urban spatial microclimate prediction with morphological map-based ensemble deep learning method

Weather data play a crucial role in shaping building energy consumption in cities. Conventional building simulation software relies on a homogeneous Typical Meteorological Year (TMY) weather file to simplify the weather conditions within the whole city without considering the inter-city variation of microclimate.

This study intends to develop a spatial prediction of urban microclimate conditions based on the close relationships of urban morphology and microclimate. A novel urban morphology description method, multi-layer urban morphological mapping, is developed to generate graphic data for roads, vegetation, and buildings. A multi-modal deep ensemble learning module is proposed to establish spatial prediction with inputs of morphological maps and microclimate data from a limited number of weather stations.

The proposed model provides higher prediction accuracy than the conventional model with single-site microclimate data inputs. A validation experiment was conducted in a campus environment. The proposed map-based ensemble DL model reduces air temperature prediction errors by 78.1%, 24.5%, 19.4%, 13.2%, and 8.9% compared to TMY data, the Kriging model, the DL model, the factor-based DL model, and the map-based DL model, respectively.