Improved tunnel fire prediction method based on PatchTST

Abstract

Accurate forecasting of multi-channel parameters—temperature, carbon monoxide concentration, and wind speed—is essential for emergency ventilation and evacuation during tunnel fires. Fire temporal curves, however, couple slow convective evolution with abrupt, localized fluctuations driven by transient physical disturbances, and existing deep learning models struggle to separate these multi-scale dependencies, causing errors to accumulate over long horizons. We propose PatchTST-multilayer perceptron (MLP), a hybrid model for multi-channel tunnel fire forecasting. The architecture uses patch-based embedding to capture local temporal context at low computational cost. A context-aware decomposition mechanism then splits the embedded sequence into a long-term trend and a transient-disturbance component; the Transformer encoder models global long-range dependencies while the MLP encoder captures high-frequency fluctuations. We constructed a full-scale tunnel fire dataset of 111 cross-sectional channels under varied heat release rates and ventilation conditions using fire dynamics simulator, with a strict sequential partitioning strategy to prevent data leakage. PatchTST-MLP achieves competitive or superior performance against state-of-the-art time-series baselines across short- and long-term horizons. Ablation studies show that 2 Transformer blocks and 2 MLP layers give the best accuracy–complexity trade-off making it practical for real-time tunnel fire safety management.