Predicting fire dynamics in ventilated tunnels with spatio-temporal buffer network

Abstract

Tunnel fires pose significant threats to public safety and urban infrastructure. Prompt responses and the implementation of effective firefighting measures are crucial for reducing losses from tunnel fires. However, current approaches lack high-precision and long-term prediction methods for various physical fields in tunnel fires. To address this issue, this paper proposes a deep learning model capable of predicting the development of multiple physical parameters in tunnel fires, such as temperature and visibility: Multi-Granularity Spatiotemporal Buffer Network (MUST-BN). MUST-BN simplifies the integration of multiple physical fields by using multi-scale wavelet convolution modules and global Fourier transform modules. It effectively captures the inherent spatiotemporal relationships in fire dynamics. Additionally, the model employs a spatiotemporal attention mechanism to seamlessly merge information from the two modules. It also introduces a buffer mechanism that connects the neural network with data streams to improve long-term prediction accuracy. This paper constructs a FIT dataset by simulating various tunnel fire scenarios, providing learning data for the model. Deep learning numerical results show that MUST-BN exhibits small prediction errors in forecasting tunnel fire-related parameters, with prediction errors for temperature and visibility below 5%. Furthermore, comparative analyses with other state-of-the-art models reveal that MUST-BN significantly outperforms them in temperature prediction, achieving a mean squared error of 0.0342 and a structural similarity index measure of 0.9117. These results indicate that MUST-BN effectively models the interactions of multiple physical fields and provides reliable long-term predictions, thereby offering valuable support for fire safety planning and emergency response in tunnel environments. We provide our complete code and dataset at the following URL: https://github.com/Fire-pre/MUST-BN/tree/main.