A rapid indoor 3D wind field prediction model based on conditional generative adversarial networks

The prediction of building performance during the early design phase is essential for architects and engineers. Given the complex nature of parameter inputs and the need for time efficiency, surrogate models have become a preferred method for predicting building performance. However, most surrogate models for indoor airflow could not predict the wind flow field information in three-dimensional (3D) space (named 3D wind field). The few advanced 3D data prediction models are often computationally expensive. This paper proposes a surrogate model based on Conditional Generative Adversarial Networks for the prediction of indoor 3D wind fields under natural ventilation. The core innovation lies in compressing indoor 3D wind field information into 2D planes via image encoding and subsequently obtaining wind field maps of arbitrary planes through data post-processing. By taking prefabricated houses as the case study, a database is constructed and the model is trained to predict the wind field at any cross-section within the space. The resulting surrogate model can generate predictions within a 3–5 s timeframe. To evaluate the accuracy of the model prediction, 21 testing planes were selected. Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) were used to assess the numerical accuracy, and Structure Similarity Index Measure (SSIM) was used to comprehensively evaluate the visualization results of the wind field images. The results indicate that the model exhibits outstanding prediction performance for the planes, with an MAE of 0.1233, a MAPE of 12.20 %, and an SSIM of 0.9492 on the test set. Compared to simulation methods, this approach can improve prediction speed by 350 times-450 times, significantly enhancing the efficiency of obtaining 3D wind fields during the early design stages. © 2025 Elsevier Ltd