Data-driven prediction framework of surrounding rock pressure in a fully mechanized coal face with temporal-spatial correlation

Surrounding rock pressure prediction for fully mechanized coal face (FMCF) roof management and control is of great significance. The key challenge is to effectively combine various factors to evolve the surrounding rock pressure trend with high temporal-spatial correlation and heterogeneity. The hydraulic support load is an important indicator that reflects the change in the surrounding rock pressure. Inspired by the research on 'support-surrounding rock' and the massive data collected by sensors, this paper proposes a data-driven prediction framework for surrounding rock pressure in FMCFs. The framework includes the multidimensional working condition matrix (MWCM) building module, the hydraulic support group temporal-spatial feature fusion (HTSFF) network, and the adaptive deployment strategy (ADS). The MWCM generating methods of hydraulic support in combination with the mining process and the data quality and quantity requirements of the conventional supervised learning method are weakened by the treatment of missing data adapting to the working conditions. The HTSFF network focuses on capturing the spatial correlation of the surrounding rock pressure along the mining direction and the temporal periodicity along the advancing direction of the FMCF and fully considers the dependencies between them. ADS aims to solve the problem of missing data in model deployment and reduce the interference of abnormal data in the model prediction process by adaptively generating MWCM. The rationality and practicability of various designs of the proposed framework were verified on our dataset. Quantitative experiments show that the average error of the surrounding rock pressure prediction is 1.2406 MPa, and the accuracy, precision, and recall rate of periodic pressure prediction are all above 95%. This method can effectively assist FMCF roof management and control. Although the method is effective, it may require further fine-tuning of the model to adapt to different geological conditions and varying sensor quality, ensuring broader applicability.