Abstract: Under the condition of shallow fully mechanized mining, the toxic and harmful gases generated by the spontaneous combustion of gangue piles above the subsidence area are likely to pour into the underground through mining-induced fractures, which seriously threatens mine safety. However, the existing monitoring systems have problems such as single-point threshold dependence, high false alarm rate, and lack of spatiotemporal correlation mining. To address this issue, it proposes an online monitoring and fire prevention method integrating a Multiscale Spatiotemporal Graph Convolutional Network （MSTGCN） and superconducting gravity heat pipe accumulated temperature export technology. This method captures data anomalies through multiscale feature extraction, uses MSTGCN to mine spatial dependencies and risk propagation paths, incorporates physical prior constraints to construct the MSTGCN-PI hybrid model to enhance prediction stability during data-sparse periods, and finally couples heat pipe technology to carry out accumulated temperature export fire prevention and extinguishing tests. Taking the gangue pile in the overlying subsidence area of the 22106 working face in Shangwan Coal Mine as the test site, arrange 107 monitoring points to collect data for 12 months, and set up 110 heat pipes in the 1 300 m² test area to verify the effect. The results show that the MSTGCN-PI model advances the early warning time by 6.8 hours, which is 6.3 hours longer than that of the traditional threshold method, reduces the false alarm rate to 2.6%, and achieves F1-score and AUC values of 0.93 and 0.97, respectively. After heat pipe treatment, the deep temperature （at 5 m and 10 m） and CO concentration of the gangue pile decrease significantly, and spontaneous combustion is effectively suppressed. This technology constructs a full-process closed loop of “monitoring-early warning-location-governance” and provides an efficient engineering scheme for similar prevention and control tasks.