Assessing Urban Micro-climate Conditions by Building Morphological Machine Learning-based Analysis Methods

Abstract

Cities occupy only 0.63% of the Earth’s surface, yet they account for roughly 76% of global energy consumption and greenhouse gas (GHG) emissions. Driven by rapid urbanization over recent decades, urban buildings have become the largest contributors to this growth, responsible for nearly 40% of total energy use and about 33% of GHG emissions. Consequently, the building sector holds the greatest potential for energy savings and achieving carbon neutrality. To enhance building energy efficiency and support city management, accurate and realistic estimation of building energy dynamics is essential. However, urban areas are highly complex, shaped by both natural and built environments. This complexity creates heterogeneous microclimates with significant impacts on building energy demand, most notably through urban heat island (UHI) effects. Despite this, most building energy assessment and simulation tools simplify reality by relying on city-scale Typical Meteorological Year (TMY) weather files, which assume uniform climatic conditions across an entire city. In practice, this overlooks crucial microclimate variations. Therefore, developing precise and reliable microclimate data is critical for producing realistic and accurate building energy consumption estimates.

In urban environments, heterogeneous microclimate conditions arise from complex interactions among man-made structures, the natural environment, and human activities. As the dominant component of cities, buildings exert a strong influence on microclimate, most notably through the urban heat island (UHI) effect. Research has consistently shown that morphological attributes are the primary drivers of a building’s impact on the local environment, affecting shading, ventilation, and reflection. To examine the influence of building morphology on microclimate, various quantitative methods have been developed to describe building arrangements and geometries. Traditional approaches typically rely on average morphological attributes within a building patch, such as mean height, footprint area, or volume. However, such aggregated measures overlook the spatial distribution of buildings and often show weak correlations with urban microclimate conditions. This highlights the need for more detailed and effective methods of characterizing building morphology for microclimate assessment. Recent advances in machine learning (ML) and deep learning (DL), including artificial neural networks (ANN) and long short-term memory networks (LSTM), have demonstrated strong performance in handling nonlinear relationships and complex input features. Incorporating detailed building morphological features with ML or DL approaches therefore holds significant potential for improving urban microclimate assessment.

To address this research gap, this thesis presents four main contributions. First, an urban vertical wind speed estimation method was developed that incorporates building morphologies. A pie-shaped segmentation approach was introduced to decompose buildings within a microclimate zone according to the approaching wind direction, enabling the calculation of direction-dependent, factor-based morphological indicators. Machine learning techniques were then applied to estimate the key parameters of a modified power law wind function, which facilitates the conversion of Typical Meteorological Year wind speed into urban vertical wind speeds. Wind tunnel data collected in Hong Kong were used to validate the proposed method. The results show that incorporating morphological inputs significantly improves the estimation of vertical wind speeds across different wind directions and heights. Furthermore, deep neural networks and support vector regression demonstrated superior performance.

Secondly, this thesis proposed a deep transfer learning neural network that integrates morphological features with the depth distribution of frontal areas for time-series microclimate prediction. To support this, stripe-enabled building morphology calculation methods were developed to decompose the frontal area of building groups using horizontal, vertical, and two-dimensional grid cutting. Each cutting method was applied in four directions (north, northeast, east, and southeast). A validation experiment was conducted on a university campus using a wireless sensing system. The results demonstrate that stripe-enabled building morphologies significantly enhance the accuracy of local microclimate prediction, with the combined vertical and horizontal cutting approach achieving the best performance.

Third, a sun-path-dependent deep transfer learning neural network approach was developed to assess built microclimates using frontal projection maps of building groups. Two projection map generation methods, parallel projection gridding and surrounding projection gridding, were introduced. By combining seasonal sun-path trajectories with frontal projection maps, the proposed network better captures the shading effects of buildings on local microclimates. A case study conducted on a university campus with wireless environmental sensing systems validated this approach. The findings show that incorporating sun-path data with frontal projection maps effectively improves the accuracy of microclimate prediction.

Finally, this thesis introduced multi modal deep ensemble learning neural networks for spatial prediction of microclimate conditions, enabling the regression of urban morphological impacts over large areas with limited data collection points. Unlike the previous studies, this approach employed multi-layer urban morphological maps to represent buildings, roads, and vegetation in graphic formats. The ensemble mechanism was designed to reduce prediction errors associated with single point microclimate measurements. A campus-based case study using local weather stations validated the method, and the results indicate that the proposed model outperforms conventional interpolation techniques as well as deep learning models relying solely on factor based morphological inputs.

In summary, this thesis systematically investigated the influence of building morphology on local microclimates and advanced urban built-environment analysis through urban model dimensionality reduction. A series of novel building object decomposition methods were developed to generate more effective and relevant morphological factors for data-driven microclimate prediction. These contributions enable more accurate microclimate estimation, which is essential for improving building energy performance evaluation.

Date of Award	4 Sept 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Xiaowei LUO (Supervisor) & Jiayu CHEN (Supervisor)