A METHOD FOR REAL-TIME PREDICTION OF INDOOR NATURAL VENTILATION IN RESIDENTIAL BUILDINGS

Against the backdrop of energy crises and climate change, performance-oriented architectural design is increasingly gaining attention. Early-stage assessment of natural ventilation performance is crucial for optimizing designs to enhance indoor environmental comfort and reduce building energy consumption. However, traditional numerical simulations are time-consuming, and existing data-driven surrogate models primarily focus on predicting partial indicators in indoor airflow or single-space airflow. Predicting the spatial distribution of airflow is more advantageous for addressing global issues in building layout design. This paper introduces a surrogate model based on Generative Adversarial Networks. We constructed a dataset of floor plans, with 80% of the data generated using parameterized methods and 20% sourced from real-world examples. We developed a 3D encoding method for the floor plans to facilitate machine understanding of spatial depth and structure. Finally, we conducted airflow simulations on the dataset, with the simulated results used to train the Pix2pix model. The results demonstrate that the Pix2pix model can predict indoor airflow distribution with high accuracy, requiring only 0.8 seconds. In the test set, the average values for MAPE, SSIM, and R-2 are 2.6113%, 0.9798, and 0.9114, respectively. Our research can improve architectural design, enhance energy efficiency, and enhance residents' comfort, thereby contributing to the creation of healthier indoor environments.