基于深度学习的城市建成环境热舒适快速预测方法

Rapid Prediction Method for Thermal Comfort in Urban Built Environments Based on Deep Learning

关键词:  风景园林  城市建成环境  生成式对抗网络  pix2pix算法  热舒适

Key words:  landscape architecture  built environments  GAN  pix2pix algorithm  thermal comfort

摘要:

传统的城市热舒适性评估方法因缺乏多层次综合预测能力，难以满足城市建成环境形态景观优化中大量方案快速迭代的实时反馈需求。提出基于生成对抗网络(GAN)的双层次城市热舒适快速预测模型，整合高分辨率城市形态数据与深度学习技术，分为“建筑布局”宏观模型和“绿化布局”微观模型。通过pix2pix架构，分别实现建筑整体布局下的热舒适性能预测与建筑和绿化系统下的细化预测。结果表明，宏观建筑布局模型在测试集上的结构相似性指数(SSIM)平均值为89.3％，而微观绿化布局模型为94.8％；拟合优度R2分别为0.958和0.980。说明模型在不同尺度下均具有较高的预测精度，支持了城市环境中景观与形态的协同优化。

Abstract:

The Urban Heat Island (UHI) effect has become one of the most pressing challenges for improving urban environmental quality. On the one hand, UHI exacerbates extreme summer heat events, significantly reducing outdoor thermal comfort and posing risks to public health; on the other hand, it increases building cooling loads and energy consumption, thereby constraining sustainable urban development. Existing studies indicate that urban morphology, particularly building configuration and greenery, plays a decisive role in regulating microclimate conditions. Building height, density, and material properties affect local thermal environments through shading, airflow, and radiative processes, while urban greenery mitigates heat via shading, evapotranspiration, and enhanced surface reflectivity. These insights provide a theoretical foundation for mitigating UHI through design optimization. However, traditional observation methods such as meteorological stations and remote sensing, as well as numerical simulation tools including ENVI-met, PHOENICS, and Ladybug Tools, remain constrained by computational cost, spatial resolution, or processing speed. They are often unable to deliver the real-time, fine-grained predictive feedback required for rapid iteration of urban design schemes. To overcome these limitations, this study develops a rapid prediction method for outdoor thermal comfort based on deep learning. We propose a dual-level Generative Adversarial Network (GAN) framework that integrates high-resolution morphological data with environmental simulation. The framework consists of a macro-scale building layout model and a microscale greenery layout model. In the first stage, the building layout model, trained using the pix2pix architecture, predicts baseline Universal Thermal Climate Index (UTCI) maps directly from urban morphology inputs. In the second stage, the greenery model refines these predictions by incorporating vegetation distribution as an additional multi-source input, thereby capturing the combined effects of building configuration and landscape design. This dual-level structure enables the macro-scale model to guide and constrain micro-scale predictions, significantly improving both accuracy and efficiency. The case study area is Tianjin's Nankai District, characterized by compact building clusters and limited open space. Using 300 m × 300 m sampling windows, we collected detailed three-dimensional building data, street networks, meteorological records, and vegetation layouts from open platforms such as OpenStreetMap and domestic mapping services. Ladybug Tools was employed to simulate UTCI values under different climate scenarios, producing 2,000 paired datasets of morphology/ greenery inputs and UTCI outputs for training and validation. The GAN models were trained with 70% of the data used for training, 20% for validation, and 10% for testing. The results demonstrate excellent predictive performance. The building layout model achieved R2 = 0.958 and an average Structural Similarity Index (SSIM) of 0.893 on the test set, while the greenery layout model reached R2 = 0.980 with an SSIM of 0.948. Compared to conventional simulation, which required approximately 4.5 minutes per case, the trained GAN framework reduced prediction time to 0.03 seconds per case, enabling near real-time feedback. Cross-city validation using building layouts from Beijing, Shanghai, Qingdao, Zhengzhou, and Shenzhen further confirmed the generalization capability, with SSIM values consistently above 0.84. These findings verify that the dual-level model not only captures the physical mechanisms of thermal comfort formation but also achieves strong transferability across different urban contexts. To facilitate practical application, the proposed framework was embedded into a Rhino & Grasshopper plug-in. This tool allows urban designers to interactively modify building height, density, and vegetation distribution, and immediately visualize their impact on UTCI distribution. In conclusion, this research contributes a novel deep learning approach to the field of urban climate modeling by integrating macro- and micro-scale morphological factors within a unified predictive system. The dual-level GAN framework achieves both high accuracy and efficiency, offering a scalable solution for rapid scenario testing in urban design. This method not only expands the applicability of deep learning in environmental performance prediction but also provides a powerful decisionsupport tool for optimizing thermal comfort, enhancing greenery quality, and promoting sustainable urban development