Multi-step prediction of roof pressure based on multi-scale contextual fusion network

Roof accidents are very harmful to the normal production of mines and the safety of underground workers, and the realization of accurate prediction of roof pressure is of great significance for the safe and intelligent mining of coal seams. To solve this problem, a methodology based on multi-scale contextual fusion network is proposed to predict the roof pressure. Firstly, to effectively filter out the noise in the mine pressure data and better maintain the characteristic trend of the original data, Savitzky-Golay filtering is used to preprocess the raw pressure data. Then, in order to adapt to the speed and magnitude of changes in the mine pressure data over different time periods, we spatially group the data and perform feature extraction at different scales in a multiscale long and short-term memory network. Compared with previous methods of obtaining features at a single scale, this design has extremely high advantages and performance. Meanwhile, considering the influence of different contextual information on the current moment, an efficient multiscale residual network is designed to extract input features comprehensively, which makes up for the disadvantage of considering only the information above. Finally, the pressure data from a working face of FuCun coal mine in ZaoZhuang City, Shandong Province is chosen to be fed into the model and the experiments are conducted. The results show that at 12 steps after prediction, the multistep prediction model of roof pressure based on multi-scale fusion network reduces the RMSE by about 72% and 43%, and the MAE by about 71% and 37%, respectively, compared with LSTM and GRU networks. The proposed model has a leading position as well as core competence in multi-step prediction of roof pressure.