Physics-informed machine learning-driven active utilization of ventilation and heat exchange in large underground tunnels

Underground ventilation tunnels have significant potential for shallow geothermal utilization. Traditional methods struggle to provide accurate real-time conditions for calculating tunnel ventilation heat exchange. This study presents a physics-informed machine learning (PIML) framework to predict tunnel air parameters, using an LSTM-based Seq2Seq model with an attention mechanism for real-time multi-step predictions of temperature and humidity at the tunnel outlet. Based on year-round simulation data from 16 intake tunnels and field data from the target tunnel, the framework achieves accurate predictions of air temperature and humidity. Results show that the average real-time prediction errors for temperature and relative humidity are below 0.4 °C and 2 %, demonstrating the method's reliability. Attention heatmaps and feature ablation tests reveal the model's interpretability: it adapts to focus on key time steps and daily cycles, with physical factors like time cycles, enthalpy, and wind speed crucial for accuracy. During the measurement period, the 1950m-long tunnel's average heat exchange rate was 199 kW, indicating its substantial energy regulation potential. This research provides a feasible approach for real-time heat exchange prediction in large tunnels and lays the foundation for efficient ventilation heat use in underground power stations, energy storage, mines, and subways.