Scholarlink

<p>The heat release rate is a crucial indicator of fire behavior in the confined environments of tunnels, making it crucial to quickly assess and predict fire conditions. This study investigates the prediction of heat release rates of tunnel fire using flame images through deep learning techniques. Initially, a series of tunnel fire experiments were conducted under various ignition points with forced ventilation to capture flame images across different scenarios. Next, according to the temporal changes in the flame images, the Deep Residual Network (ResNet18) was used to extract flame features. Subsequently, the Long Short-term Memory (LSTM) model was employed to track the temporal evolution of these flame features and relate them to the heat release rates at corresponding time intervals. The results indicate that the ResNet18-LSTM model effectively captures the dynamics during the growth and decay phases of the fire, although its predictive capability is limited around the peak heat release rate. Additionally, comparisons with three other models (CNN-GRU, CNN-LSTM, ResNet18-GRU) demonstrate that the ResNet18-LSTM model maintains a level of predictive accuracy. Specifically, it achieves an average R<sup>2</sup> increase of 15.49%, 28.12%, and 26.15% over CNN-GRU, CNN-LSTM, and ResNet18-GRU, respectively. Moreover, the average MAE is reduced by 14.88%, 27.56%, and 29.13%, while the average RMSE decreases by 19.17%, 30.21%, and 28.93%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantification of Heat Release Rate in Tunnel Fires: An Intelligent Real-Time Prediction Model Based on Flame Images

  • 期刊论文

  • 发表日期：

    2025年7月21日

查看全文
  • Gongyousheng Cui,
  • Yuchun Zhang,
  • Xinyi Liu,
  • Haowen Tao,
  • Keran Li

摘要

The heat release rate is a crucial indicator of fire behavior in the confined environments of tunnels, making it crucial to quickly assess and predict fire conditions. This study investigates the prediction of heat release rates of tunnel fire using flame images through deep learning techniques. Initially, a series of tunnel fire experiments were conducted under various ignition points with forced ventilation to capture flame images across different scenarios. Next, according to the temporal changes in the flame images, the Deep Residual Network (ResNet18) was used to extract flame features. Subsequently, the Long Short-term Memory (LSTM) model was employed to track the temporal evolution of these flame features and relate them to the heat release rates at corresponding time intervals. The results indicate that the ResNet18-LSTM model effectively captures the dynamics during the growth and decay phases of the fire, although its predictive capability is limited around the peak heat release rate. Additionally, comparisons with three other models (CNN-GRU, CNN-LSTM, ResNet18-GRU) demonstrate that the ResNet18-LSTM model maintains a level of predictive accuracy. Specifically, it achieves an average R2 increase of 15.49%, 28.12%, and 26.15% over CNN-GRU, CNN-LSTM, and ResNet18-GRU, respectively. Moreover, the average MAE is reduced by 14.88%, 27.56%, and 29.13%, while the average RMSE decreases by 19.17%, 30.21%, and 28.93%.