运动列车火灾条件下隧道内温度分布快速预测方法

[Objective] In case of catching fire, a train moving in a tunnel section might lose power and come to an emergency halt while approaching the next Station for emergency rescue. Predicting the distribution of smoke and temperature in tunnels under such fire scenarios is difficult because of the influence of moving trains. This difficulty in prediction might seriously threaten the safety of passengers and impede metro evacuation management. [Methods] This study considers influencing factors, including train speed, fire source location, and fire heat release rate, to solve this prob lern, and 75 different fire scenarios are designed and simulated. Train movement in the simulated scenarios is realized using the equivalent piston wind method. The Simulation results of the smoke and temperature distributions collected using sensors near the tunnel ceiling are used to construct a dataset for deep learning. Accordingly, a deep learning model comprising long short-term memory networks, a convolution (Conv) module, and a deconvolution (DeConv) module is then proposed for rapid prediction of temperature distribution in tunnels under moving train fire conditions. The train speed, train braking time, and temperature time-series Information from the sensors together are fed as inputs to the model. [Results] The results indicated that: (1) Under various train movement states, the model was able to predict the temperature distribution of the lateral evacuation platform in the tunnel 30 s in advance using the current sensor data, with a mean absolute error (MAE) of only 2. 2 °C and a mean absolute percentage error (MAPE) of 4%, indicating high accuracy. (2) In a stark contrast with the week-long time taken to obtain temperature distribution in a fire dynamics Simulator (FDS), this deep learning model could make prediction within only 0.08 s, hence representing a computational efficiency improvement of four Orders of magnitude versus the computational fluid dynamics method. (3) Validation with fire scenarios in none of the training, Validation, and test datasets resulted in model MAE and MAPE values of 3. 1 °C and 5%, respectively, indicating a strong generalization ability. (4) Considering the possibility of sensor failure within tunnels, this study investigated the influence of simulated sensor failures on the model's prediction accuracy by varying sensor spacing. The model continued to exhibit an effective predictive ability even at a sensor spacing of 4. 00 m. At 8. 00 m sensor spacing, the model's errors were larger albeit at very few time frames (with a maximum MAE lower than 10. 0 °C and a maximum MAPE lower than 15%). However, for other sensor spacing cases, the model's MAE and MAPE were less than 5. 0 °Cand 10%, respectively. Hence, it could be concluded that the model has strong robustness. [Conclusions] This study constructs a comprehensive dataset for a tunnel with moving train fire conditions using an FDS and leverages advanced deep neural networks to completely exploit the extensive information within the dataset, ultimately resulting in a high-precision, robust model for rapid prediction of temperature distribution in tunnels under moving train fire conditions. These advancements are highly important for effective emergency management and response planning in tunnels under these challenging conditions.