Evaluating a deep learning-based surrogate model for predicting wind distribution in urban microclimate design

Abstract

Wind environment assessment in urban microclimate design is crucial for enhancing pedestrian wind comfort and managing pollutant dispersion. Deep learning-based models have shown the potential to replace traditional computationally intensive numerical simulations for accelerated assessment. However, the effectiveness and characteristics of the models in predicting urban wind environments for different building configurations remain unclear. Moreover, there is a lack of knowledge about the influence of the training dataset composition on the model performance. This study aimed to comprehensively evaluate the performance and characteristics of a deep learning-based surrogate model for fast prediction of wind distribution for urban microclimates. Based on a dataset of 4,000 simulations, an end-to-end model was first trained, and the model's performance and domain transferability were then evaluated. The trained model achieved average mean absolute percentage errors (MAPE) ranging from 1.74 % to 12.49 % for unseen configurations containing 1 to 4 buildings, offering a speed-up of 3–4 orders of magnitude over traditional CFD methods. The model exhibits limited domain transferability, as it can learn transferable wind flow patterns. As a result, the model can accurately predict wind flow patterns around isolated buildings, while it struggles to capture the complex wind flow caused by a dense arrangement of multi-building cases. To improve the robustness and applicability of the model, integrating building configurations with different numbers of buildings into the training stage could be an effective strategy.