Study on a real-time prediction framework for the heat release rate and ceiling heat flux profile in a tunnel fire with lateral smoke extraction using deep learning

Tunnel fires pose significant energy safety risks due to the narrow and restricted spaces and complex ventilation conditions. A typical scenario is a train roof fire triggered by a pantograph malfunction, which not only releases substantial thermal energy but also results in the wastage of electrical and fuel energy. This study focuses on the impact of the coupled lateral smoke exhaust system, which combines longitudinal ventilation and lateral exhaust, on fire heat release rate and flow field evolution during train fires, while also exploring an AI-based approach for predicting train fire dynamics. Compared with traditional ground fires in tunnels, the burning rate of train roof fires in interval tunnels shows non-monotonic changes, and the evolution characteristics of the flow field also differ significantly. To reveal the prediction process of the full-period changes in tunnel fires over time and overcome the temporal limitations of traditional fire physical models, this paper proposes a full-period, real-time prediction framework and the novel algorithm for heat release rate (HRR) and ceiling heat flux profile of tunnel fires using deep learning. The framework integrates ResNet-18 and ViT-Small for flame image feature extraction, with physical prior information incorporated as auxiliary input features to enhance the model's predictive performance and physical interpretability. This enables rapid and accurate full-period prediction of the HRR and the heat flux profile beneath the tunnel ceiling under the coupled conditions of longitudinal ventilation and lateral extraction in a tunnel.