Fire scenario simulation method for residential buildings based on generative adversarial network

Abstract

Fire scenario simulation in residential buildings is crucial for fire safety design, risk assessment, and emergency management. Traditional CFD-based methods face challenges, including long computation times and reliance on expertise, limiting their use for real-time prediction and rapid design optimization. This study introduces a novel simulation method using Generative Adversarial Networks (GANs). A database of 50 residential layouts encompassing a wide variety of apartment configurations is constructed, with high-resolution spatiotemporal data on temperature and soot visibility generated via CFD. The GAN-based model uses layouts, ignition locations, and fire development times as inputs to predict temperature and soot fields. Experimental results show the model achieves an average Structural Similarity Index (SSIM) of 95.7 % compared to CFD and reduces prediction time to 2.56 s—an efficiency improvement of 80,000 times. This method provides an efficient tool for fire risk assessment, evacuation planning, and intelligent fire protection system design in residential buildings.