Abstract: Aiming at the spontaneous combustion risk caused by accumulated heat concentration in the gangue pile of subsidence areas in Shendong Shangwan Coal Mine, focusing on optimizing the accumulated heat extraction efficiency of superconducting gravity heat pipes, an RF-Attention-LSTM hybrid machine learning model is proposed. A dataset is constructed based on on-site measured data, features are selected and parameters are initially optimized via RF, and the Attention mechanism is introduced to enhance LSTM’s learning of time-series features, realizing precise optimization of heat pipe parameters. Results show that the model’s coefficient of determination R² reaches 0.97, 18.6% and 23.3% higher than standalone RF and LSTM models respectively; the optimized parameters are 8.5 m in length, 5.8 m in burial depth and 3.2 m in spacing. Applied to areas ≥1 300 m², the surface temperature is stably ≤35 ℃, and the accumulated heat extraction rate is over 35% higher than traditional parameters, providing an efficient technical solution for spontaneous combustion prevention of gangue piles in subsidence areas.