Aiming at the prominent rockburst risk in deep mining of Kongzhuang coal mine with high in-situ stress and thick hard roof, this study takes microseismic monitoring time-series data as the research object, and constructs a set of rockburst prediction and early warning method based on LSTM multivariate deep learning. According to the evolutionary characteristics of cumulative energy and cumulative apparent volume, the rockburst development process is divided into three stages, among which the abnormal acceleration stage is taken as the key early warning interval. On this basis, univariate and multivariate parallel prediction models are established by using improved LSTM networks including S-LSTM, Bi-LSTM and CNN-LSTM. The results show that S-LSTM has the highest accuracy in univariate prediction, and CNN-LSTM performs best in multivariate parallel prediction. Furthermore, an early warning model of MCNN-LSTMs is constructed for rockburst stage identification, and its overall accuracy is both higher than those of MLP, FCN, GRU, TCN, and Transformer models. Combined with field engineering application, the rockburst early warning criterion and grading warning mechanism are formulated. The field verification in Kongzhuang coal mine proves that the proposed method can accurately predict the evolution trend of microseismic parameters and effectively identify precursor anomalies before rockburst, which significantly reduces false alarms and missing alarms. This study provides a reliable technical approach for time-series prediction and dynamic early warning of rockburst in deep coal mines.
Guan et al. (Thu,) studied this question.