To reduce the loss of tunnel fire accidents, a rapid detection and 3D temperature field prediction model was constructed for tunnel vehicle fires based on deep learning to realize advanced tunnel fire prediction. An improved YOLOv8 network model combined with the OpenCV module was employed to achieve vehicle information detection and position prediction. A novel triangle rotation convolutional neural network (TRCNN) model and Triangle framework were proposed, which were integrated with the high-order polynomial regression (HOPR) model and Informer-long short-term memory (Informer-LSTM) network to form the HOPR-Informer-LSTM hybrid model for fire smoke recognition and precise 3D fire temperature prediction. The research results demonstrate that the vehicle classification and license plate recognition achieve mean average precision at 50% intersection over union (mAP@50) of 98% and 99% respectively, vehicle speed prediction reaches a coefficient of determination (R²) of 94%, and fire smoke recognition accuracy reaches 82%. Validation of the temperature prediction model using small-scale test bench measurements shows that the 3D temperature field prediction error remains within ±5 ℃. The research results provide an advanced tunnel fire prediction method, offering technical support for tunnel fire prevention and control.