An Improved Updatable Backpropagation Neural Network for Temperature Prognosis in Tunnel Fires

Abstract

Because it is impossible to predict the temperature in advance, specific fire scenes (fire type, fire location, tunnel geometry, etc.) are unknown in traditional tunnel fire research. To address this difficulty, this work developed a novel algorithm to achieve temperature prognosis in tunnel fires that includes an updatable backpropagation (BP) neural network and a smoothing procedure. The data-driven algorithm is not limited to a specific fire scene, which makes it easy to fit real complex tunnel fire disasters. In addition, a full-scale fire test was conducted and utilized to verify the algorithm. Two innovations, including the updatable BP neural network and the smoothing procedure, made the predicted results match well with the experimental results. We can achieve a real-time precise temperature prediction 20 s in advance at a high accuracy of about 85.6%. If there is no sudden external factor intervention, the accuracy is about 99.4%. The algorithm provides an effective numerical tool for early fire warning and firefighting decision making that can address the temperature prognosis of tunnel fires.