LES-Based Optimization of Urban Thermal Comfort Using Machine Learning

Thermal exposure risks in high-density urban environments directly impact pedestrian health. In the precincts surrounding Taipei 101, super-tall buildings induce complex street-level vortices that significantly attenuate near-ground wind speeds. This study investigates the coupling effects between anthropogenic heat emissions from air conditioning systems (AC) and urban street flow fields. AC exhaust heat is frequently trapped at the pedestrian level (1.5 m) by micro-scale vortices, and improper configurations lead to a pronounced Thermal Retention Effect.

To establish a data-driven evaluation framework and translate simulation outputs into actionable building heat-discharge strategies, this research employs Large Eddy Simulation (LES) to conduct transient thermo-fluid flow-field modeling, validated through on-site measurements. For the 1.5 m near-ground micro-environment, Physiological Equivalent Temperature (PET) is adopted as the thermal stress index, integrating air temperature (Ta), wind speed (V), and mean radiant temperature (Tmrt).

Recognizing that conventional CFD approaches struggle to provide quantitative weighting of heat-accumulation factors and to identify dominant thermal mechanisms, this study introduces a Machine Learning (ML) model to quantify the relative contributions of different AC installation positions to localized heat accumulation, thereby establishing an efficient and predictive thermal-load assessment framework.

Results indicate that leeward vortices in dense residential districts produce a heat-retention rate of 60%, driving pedestrian-level PET into the Extreme Heat Stress category. ML feature-importance analysis reveals that the interaction between discharge positioning and vortex circulation is the dominant thermal driver. Optimization of AC configurations reduces the spatial extent of high heat-accumulation zones from 60% to 25%, significantly mitigating thermal vulnerability and localized heat hazards caused by Thermal Trapping.

This study confirms that the coupling between street vortices and anthropogenic heat emissions is the primary physical driver of near-ground microclimate deterioration. By optimizing AC placement (windward, leeward, crosswind, or rooftop) via the ML model, localized heat loads within the urban canyon can be effectively removed through ventilation-enhanced heat dissipation, enabling a functional decoupling between building heat discharge and ambient flow fields.

Keywords: Human Biometeorology; Urban Canyon; LES; Machine Learning; Physiological Equivalent Temperature (PET); Anthropogenic Heat; AC Configuration Optimization