The distribution of outdoor microclimate airflow is critical for analyzing outdoor thermal comfort and building energy consumption. However, traditional models are often complex, and the intricate processes of heat and mass exchange typically result in lengthy computation times. To address this challenge, this study proposes a deep neural network (DNN) to rapidly predict the three-dimensional temperature field of urban microclimates using 36 high-fidelity microclimate simulation datasets. The DNN model has demonstrated remarkable computational efficiency, with CPU prediction times of approximately 2 s, significantly reducing the acquisition time compared to traditional computational fluid dynamics (CFD) simulations, which typically require around 30 min. The results indicate that the DNN could achieve highly accurate predictions, particularly in critical areas within 5 m above the ground. Specifically, only 12% of the well-trained DNN's area predictions exhibited a root mean square error (RMSE) exceeding 0.5°C at a height of 1 m, while the majority of other sections show temperature prediction deviations generally below 0.5°C. The mean error across 10 repeated temperature predictions was less than 0.5°C, with a mean absolute percentage error (MAPE) of less than 1.2%, underscoring the reliability of the DNN in predicting microclimate temperatures.
Li et al. (Fri,) studied this question.