A novel artificial neural network fire model for prediction of thermal interface location in single compartment fire

Thermal interface is the boundary between the hot and cold gases layers in a compartment fire. The height of the interface depends predominantly on the mass of air entrained into the fire plume. However, the analytical determination of the air mass flow rate is complicated since it is highly nonlinear in nature. Currently, computer models including zone models and field models can be applied to predict fire phenomena effectively. In the zone model computation, the compartment on fire is commonly divided into two layers to which conservation equations are applied to evaluate the fire behaviour. However, the locations of the fire bed and the openings are ignored in the computation. Computational fluid dynamics techniques may be employed, but a major shortcoming is the requirement for extensive computational resources and lengthy computational time. A unique, new and novel artificial neural network (ANN) model, denoted as GRNNFA, is developed for predicting parameters in compartment fires and is an extremely fast alternative approach. The GRNNFA model is capable of capturing the nonlinear system behaviour by training the network using relevant historical data. Since noise is usually embedded in most of the collected fire data, traditional ANN models (e.g. feed-forward multi-layer-perceptron, general regression neural network, radial basis function, etc.) are unable to separate the embedded noise from the genuine characteristics of the system during the course of network training. The GRNNFA has been developed particularly for processing noisy fire data. The model was applied to predict the location of the thermal interface in a single compartment fire and compared with the experiments conducted by Steckler et al. (Flow induced by fire in a compartment, NBSIR 82-2520, National Bureau of Standards, Washington, DC, 1982). The results show that the GRNNFA fire model can predict the location of the thermal interface with up to 94.5% accuracy and minimum computational times and resources. The trained GRNNFA model was also applied to rapidly determine the height of the thermal interface with different locations of fire on the compartment floor and different widths of the opening against field model predictions. Among the five test cases, four of them were predicted well within the minimum error range of the experiment results. It also demonstrated that the prediction accuracy is related to the amount of knowledge provided for network training.