Real-Time Prediction of Smoke Spread Affected by Multiple Factors in Subway Tunnel Using CAERES-DNN Model

Abstract

High-accuracy and real-time smoke spread models are vitally important for firefighting in subway tunnels. This paper proposes a data-driven model, named CAERES-DNN, for predicting smoke spread in subway tunnel fire. CAERES-DNN constitutes of CAERES (convolutional autoencoder with residual blocks) architecture and DNN (deep neural network) architecture. CAERES has 26 convolutional layers, 4 full link layers and 2 reshape layers in total. Specially, 6 residual blocks introduced in the CAERES to avoid gradient vanishing. DNN, a six-full-link-layer neural network, is built to regress data. Both in CAERES and DNN, binary cross entropy loss with L2-norm and stochastic gradient descent optimizer with momentum are used to update and calculate the model parameters. Therefore, CAERES-DNN performs better than conventional convolutional neural network based on the dataset (500, 100 and 48 in training, validation and testing set) from 54 FDS simulations in the fired subway tunnel. The 95.3% soot visibility values predicted by CAERES-DNN is within ± 10% of FDS prediction. In the 200 m tunnel, the prediction error of spread distance is always within ± 5 m. Meanwhile the prediction time is in seconds by the pre-trained model, which is 105 times faster than FDS. Due to its accuracy and speed, CAERES-DNN can help in firefighting during a confined and long subway tunnel in real-time.