ConvLSTM for Predicting Short-Term Spatiotemporal Distribution of Seismic Risk Induced by Large-Scale Coal Mining

Coal-burst is a typical dynamic disaster that raises mining costs, diminishes mine productivity, and threatens workforce safety. To improve the accuracy of coal-burst risk prediction, deep learning is being applied as an emerging statistical method. Current research has focused mainly on the prediction of the intensity of risks, ignoring their evolution in a spatiotemporal distribution. A spatial distribution model based on the seismic energy attenuation law was established to describe quantitatively the relative spatial evolution of seismic risk at the work face. Spatiotemporal sequence samples were constructed using seismic events that occurred during the extraction of LW250105 in Huating Coal Mine. A deep learning model based on a convolutional long short-term memory network (ConvLSTM) was constructed to predict the short-term spatiotemporal distribution of seismic risks. A new loss function and metric were used in the deep learning model to improve its performance. The results showed that (1) the optimal performance of the ConvLSTM model improved on the test set by 14.7%, compared with the baseline; (2) the prediction results of the ConvLSTM model were a correction of the current RSR distribution; and (3) the multistep prediction results outperformed the baseline, with improvement rates ranging from 20 to 27%. The effectiveness of the ConvLSTM model in predicting the spatiotemporal evolution of seismic risk was demonstrated by comparing the predicted distribution in several cases. The proposed method can explicate the evolution and distributions of the key characteristics associated with seismic risk and provide effective guidance for the early warning of coal-bursts.