Deep learning-based real-time prediction of tunnel cable fire physical fields: Combustion characteristics and temperature profile

The escalating electricity demand, driven by AI and new energy industries, intensifies the fire risk in critical urban infrastructure like cable tunnels, where conventional detection methods often fail. Existing research on fire prediction largely relies on empirical models or machine learning trained on simulated data, limiting its real-world applicability. This study aims to develop and validate robust deep learning models for real-time fire stage detection and the prediction of key physical parameters using a novel high-calibre dataset from a full-scale cable tunnel experimental platform. A dual-module flame stage detection framework, integrating an improved YOLOv8s model with chromaticity recognition, was proposed. Furthermore, four mainstream machine learning algorithms (LSTM, XGBoost, BPNN, and HOPR) were comparatively evaluated for predicting flame burning duration, height, and centreline temperature. The results demonstrate that the dual-module detection framework achieves a detection accuracy exceeding 90% across all fire stages, significantly outperforming the standalone YOLOv8s model. For parameter prediction, the Long Short-Term Memory (LSTM) model consistently showed the highest accuracy, achieving a coefficient of determination (R 2 ) of 0.9986 and an average error as low as 2.67°C for centreline temperature prediction. While LSTM offers maximum precision, XGBoost and polynomial regression models provide a superior balance of accuracy and computational efficiency, making them highly suitable for practical engineering deployment. This work provides a validated, data-driven framework for intelligent fire safety management, offering a critical technological solution to enhance the operational reliability of power systems and mitigate catastrophic fire risks in underground cable tunnels.