Real-time flashover prediction model for multi-compartment building structures using attention based recurrent neural networks

This paper presents the development of an attention based bi-directional gated recurrent unit model, P-Flashv2, for the prediction of potential occurrence of flashover in a traditional 111 m2 single story ranch-style family home. Synthetic temperature data for more than 110 000 fire cases with a wide range of fire and vent opening conditions are collected. Temperature limit to heat detectors is applied to mimic the loss of temperature data in real fire scenarios. P-Flashv2 is shown to be able to make predictions with a maximum lead time of 60 s and its performance is benchmarked against eight different model architectures. Results show that P-Flashv2 has an overall accuracy of ∼ 87.7 % and ∼ 89.5% for flashover predictions with a lead time setting of 30 s and 60 s, respectively. Additional model testing is conducted to assess P-Flashv2 prediction capability in real fire scenarios. Evaluating the model again with full-scale experimental data, P-Flashv2 has an overall prediction accuracy of ∼ 82.7 % and ∼ 85.6 % for cases with the lead time of setting 30 s and 60 s, respectively. Results from this study show that the proposed machine learning based model, P-Flashv2, can be used to facilitate data-driven fire fighting and reduce fire fighter deaths and injuries.