隧道火灾预测的传统方法是使用传感器，但存在设备老化、误报率高等缺陷，因此需要开发更加高效的火灾预测手段。该文提出一种利用在隧道外安全区域可观测的外部烟气图像和深度学习算法，对隧道内的火源功率和火源位置进行同步预测的方案，首先通过FDS软件构建了100 m长隧道的外部烟气图像数据库，再利用VGG16神经网络框架建立了烟气图像与火源参数间的联系。结果表明，该文所提方案可对隧道火灾进行有效预测；基于隧道双侧正向视角的烟气图像训练所得模型的预测精度最高，对火源功率的预测误差小于25%，对火源位置的预测误差小于10 m；此外，当火源以0~2 m/s速度移动时，该文所提方案依旧可进行有效预测。该文成果可为隧道火灾的智能预测技术提供参考。

Objective: Tunnel fires pose remarkable challenges for evacuation and fire rescue operations due to inadequate ventilation and associated hazards, such as smoke accumulation, elevated temperatures, rapid heat release rates (HRRs), and severely reduced visibility. While various monitoring techniques, such as thermocouples, fibers, and CCTV cameras, have been proposed to monitor fire development trends and assist in firefighting and evacuation efforts, obtaining critical tunnel fire information, specifically real-time fire HRR and fire source locations, remains challenging. These difficulties arise mainly because conventional detection methods are often disrupted by high temperatures or obstructed by dense smoke, hindering effective information transmission. Hence, an improved method to predict tunnel fires is urgently needed. Methods: In this study, external smoke images, i.e., the smoke structure observed from outside the tunnel gate, and CNN-based deep-learning algorithms are used to predict real-time fire HRR and location within the tunnel. A 100-m full-scale tunnel is selected as the target, and its behavior is simulated using the Fire Dynamics Simulator to form an image database. During simulation, different fire parameters, such as maximum HRR, soot yield rate, and location, are varied based on typical vehicle types found in real tunnels, resulting in approximately 900 different tunnel models that generate diverse external smoke morphologies. The simulated smoke images are captured at 1 s intervals from four observation angles: front and side views from the left and right tunnel gates. As a result, approximately 388, 800 smoke images are collected in the database. For the deep-learning algorithm, the VGG16 model, proposed by the Oxford CNN team, is employed as the target AI model for tunnel prediction. During model training, the VGG16 model continuously refines its internal parameters to minimize the error between AI predictions and the FDS simulation. Results: Results show that the proposed method can effectively predict real-time variations in fire HRR variation and location. The model trained using front-view images from both tunnel gates achieved the highest prediction accuracy, with an HRR error of less than 25% and a location error of less than 10 m. Additional tunnel simulations were conducted to further validate the robustness of the proposed method. In these simulations, the fire source is not stable but continuously moving within the tunnel at velocities ranging from 0 to 2 m/s, simulating a scenario where a vehicle catches fire but does not stop immediately. The results show that, although trained on stable fire cases, the AI model still maintains high accuracy in predicting the moving fire source, with small HRR and location errors, thus confirming the effectiveness of the smoke image-based detection method. Conclusions: Notably, further efforts are still necessary for the application of this method in real tunnels because the current work does not consider the complex background interference in actual smoke images, nor does it consider the impacts of environmental factors such as wind, sprinklers, and exhaust systems on the external smoke structure. However, this study represents an important first step toward predicting tunnel fires based on external smoke, which could play a valuable role in future smart fire prediction and firefighting applications.