Accelerating urban street canyon wind flow predictions with deep learning method

Predicting wind flow statistics in urban areas is important for various environmental and engineering applications. Currently, building-resolved computational fluid dynamics (CFD) simulations are the most commonly used and reliable methods to simulate urban wind flows but they are time-consuming which limits their use in real applications. Therefore, our objective is to develop a surrogate model based on deep learning (DL), which can be used as a faster alternative to CFD methods for urban flows. The proposed model hypothesis is that the spatial distributions of the time-averaged flow quantities within urban canopies are highly correlated to the local urban geometries. To test this hypothesis, we developed a model to predict the flow in uniform urban street canyons by constructing a geometry reading filter to convert local urban geometry information around the targeted locations into a numerical array as DL model inputs. A standard feedforward DL model is then trained using large-eddy simulation (LES) results to predict the mean wind and turbulence within uniform street canyons. Our results show that the model can give fast and accurate predictions compared to LES results. The prediction errors are found to range from 5.8% to 36%, and the normalized mean bias magnitudes range from 6.6×10⁻³ to 1.6×10⁻¹ for the different flow quantities. The DL model is also found to predict the flow patterns reasonably well, consistent with experimental data similar to the results of coarse-resolution LESs. This model has the potential to be further developed into a robust and practical tool for fast urban flow predictions.