Rapid Prediction of Regional Thermal Environment Parameters: A Method Based on Numerical Solution and Machine Learning Extension

Outdoor microclimate is critical in assessing thermal comfort and building energy efficiency. However, traditional numerical simulations are limited by low computational efficiency due to their large model scales and complex heat-mass exchange processes. Artificial neural networks (ANNs), leveraging rapidity and accuracy in high-dimensional nonlinear mapping, offer a promising solution. This study analyzes a neighborhood in Guangzhou by establishing a three-dimensional (3D) microclimate model and applying a deep neural network (DNN) to predict temperature and humidity ratio distributions. We compared the performance of a single-output model (DNN-S) against a dual-output model (DNN-D) under both two-dimensional (2D) and 3D data conditions. The results indicate that the 3D DNN-S achieved superior accuracy, effectively replicating the effects of airflow inlets and surface influences. In regions with lower-level complex airflow, 90.7% of the studied area exhibited temperature prediction errors below 0.5°C, while over 83.2% of regions at all heights maintained humidity ratio errors under 0.3 g/kg. The DNN reduced prediction time to just 2 s, significantly enhancing computational efficiency. This approach provides an effective tool for urban planning, climate assessment, and coupled simulations, demonstrating broad application potential.