Abstract
Tunnel fires pose a serious threat to life and property. The prediction of tunnel fires could reduce the risk and loss from thermal disasters. Computational fluid dynamics (CFD) provides a strong tool for quantitatively analyzing tunnel fires. However, CFD calculations are time-consuming, and reverse prediction from physical fields to key control parameters using the governing equation is impossible. To improve the prediction efficiency of tunnel fire information and solve the reverse prediction problem of key control parameters in tunnel fires, this paper proposes a deep learning model for fast bidirectional prediction between the entire physical fields and key control parameters of tunnel fires.
In this study, a deep learning model based on an encoder and a decoder is constructed, in which the encoder is used to construct the mapping from the physical fields to the key control parameters, and the decoder is used to construct the mapping from the key control parameters to the physical fields. In the model training process, the input of the encoder and the output of the decoder are required to be as close as possible, and the output of the encoder and the input of the decoder are also required to be as close as possible. The mathematical differences between them are therefore defined as the loss function. In this way, the encoder and the decoder form a cyclic structure. Data processing approaches are proposed so that all physical fields have a unified format and all key control parameters have the same distribution.
The proposed model is trained using a large high-resolution numerical database with different cases under various key control parameters. The data learning ability and prediction capacity of the deep learning model are evaluated. With the increase of the training epoch, the calculated temperature field and key control parameters increasingly agree with the true temperature field and key control parameters. After 100 training epochs, the loss function almost converges, and the proposed bidirectional prediction model with the constructed dataset achieves good training convergence. In addition, the physical fields and key control parameters can be reproduced on the training set. After the completion of model training, the prediction performance of the deep learning model is tested. The average temperature field of tunnel fires and the six key parameters of tunnel fires are accurately predicted, and the predictions encompass the geometric and physical information of the tunnel.
Overall, this article proposes a deep learning network model based on the characteristics of tunnel fires for predicting various physical fields and key control parameters of tunnel fires. This study can be applied for the rapid acquisition of the full physical fields of tunnel fires, which helps design ventilation systems in tunnels and risk evaluation. In addition, another application is to retrieve the key control parameters of tunnel fires, which helps to quickly obtain the key control parameters according to the recorded infrared temperature field in the postinvestigation of tunnel fires. The above application scenarios can provide theoretical bases and new ideas for the prevention and control of tunnel fires.
