Compartment fire predictions using transpose convolutional neural networks

Abstract

This paper presents a data-driven approach to predict spatially resolved temperatures and velocities within a compartment based on zero-dimensional zone fire modeling using a transpose convolutional neural network (TCNN). A total of 1333 Fire Dynamics Simulator (FDS) simulations of simple two-compartment configurations with different fire locations, fire sizes, ventilation configurations, and compartment geometries were used in training and testing the model. In the two-compartment test cases 95% of TCNN predicted temperatures and velocities were within ±17.2% and ±0.30 m/s of FDS predictions. Although the model was trained and tested using a simple two-compartment configuration, the TCNN approach was validated with two more complex multi-compartment FDS simulations by processing each compartment individually. Overall, the flow fields in the multi-compartment validation tests agreed well with FDS predictions with 95% of TCNN predicted temperatures and velocities within ±11% and ±0.25 m/s of FDS predictions. Coupling a zone fire model with the TCNN approach presented in this work can provide spatially resolved temperature and velocity predictions without significantly increasing the computational requirements. Since the approach is based on a zone fire model, the TCNN approach presented in this work is limited to simplified geometries which can be sufficiently modeled using a zone fire model.