Critical Wind Speed Prediction in Tunnel Fire by Attention-Based Neural Network

Tunnel fires are one of the serious threats to public safety and infrastructure. When a tunnel fire occurs, the emission of smoke is crucial for controlling the loss of the tunnel fire. In order to better control the tunnel fire and reduce casualties and financial losses, it is necessary to accurately predict the critical wind speed. The neural network prediction model can play a good predictive role. Although long short-term memory (LSTM) can achieve good prediction results in dealing with sequential data, as the length of the input sequence increases, the model tends to forget the previous information seriously. This article combines the attention mechanism with the LSTM network to improve information transmission and solve long-sequence modeling problems. In this study, a dual attention mechanism-based LSTM neural network is proposed, with attention weights introduced to the encoder and the decoder. The LSTM and the CNN+LSTM models are also compared to our proposed one. It is shown that the loss of our proposed model is the smallest, indicating the best prediction performance among the three models. This proves that adding attention to the LSTM encoding-decoding architecture improves the prediction accuracy of tunnel fire critical wind speed.