Improved MF-LSTM based multi-support prediction method for roof pressure in coal mines

Coal mine roof pressure prediction plays a crucial role in ensuring the safety of coal mine production and preventing roof accidents. In recent years, coal mine roof accidents in China account for 23%-30% of the total number of coal mine accidents. How to realize the precursor identification of top plate disaster through accurate prediction has become a core issue in the industry. Currently, the research in this field faces a double challenge: on the one hand, the existing prediction models mostly use the pressure data of a single support as the training source, abandoning the coupling relationship between adjacent supports; on the other hand, the traditional time series models such as Long Short-Term Memory Networks (LSTM) are susceptible to gradient attenuation when predicting long-term pressure fluctuations, resulting in the lack of ability to capture the cyclic pressure characteristics of the roof slab. For this reason, this paper innovatively proposes a Multi-Frame Fusion LSTM model, which realizes the data fusion of different scaffolds by introducing an attention mechanism. Experiments on the self-constructed mine pressure dataset show that the MF-LSTM model reduces the MAE and RMSE by 11.43% and 25.03%, respectively, compared with the baseline model, which verifies the effectiveness of the multidimensional feature fusion mechanism in time series prediction. This study not only provides a new methodological framework for mine pressure prediction, but its multi-source heterogeneous data fusion mechanism also opens up a new technical path for underground engineering safety monitoring.