实现煤矿冲击地压智能预警对于保障矿井安全作业具有重要意义。以新疆某矿急倾斜巨厚煤层的掘进巷道冲击地压发生时序智能分级预测作为背景，分析了急倾斜巨厚煤层巷道掘进期间各微震信息指标的时空演化规律，利用遗传算法(Genetic Algorithm，GA)优化的随机森林(Random Forest，RF)对预测冲击发展趋势性能较高的多项指标进行了优选，基于相空间重构技术(Phase Space Reconstruction，PSR)将数据映射至高维空间进行重构，结合长短期记忆神经网络(Long Short-Term Memory，LSTM)训练学习高维度数据特征，构建了基于深度学习与多元混沌时序的急倾斜巨厚煤层冲击地压预测模型(PSR-LSTM)，依据现场实际对模型的预测性能进行了评价。结果表明：急倾斜巨厚煤层巷道掘进下各微震信息指标对冲击预警的敏感性较强，彼此之间具有显著的相关性；优选出了预测冲击发展趋势性能较高的6项微震信息指标；多项指标的时间序列具有混沌特性，经过相空间重构后再进行LSTM学习训练，可有效增强模型的数据利用率与预测精度，所构建的PSR-LSTM模型在指定预测时长为1 d的情况下，预测准确率可达0.913 5、F₁值可达0.911 6，均优于未经重构的LSTM模型。模型较好地预测了急倾斜巨厚煤层掘进巷道发生冲击危险的时序趋势及危险等级，研究方法可为急倾斜巨厚煤层掘进巷道冲击地压发生的智能预测预警提供借鉴与参考。

Realizing the intelligent warning of rock burst in coal mine is of great significance to ensure the safety of mine operation. Based on the intelligent classification prediction of rock burst occurrence time series in roadway of steeply inclined ultra-thick coal seam in a mine in Xinjiang, the spatio-temporal evolution of each microseismic information index during roadway excavation was analyzed. The Random Forest optimized by Genetic Algorithm (GA) was used. RF selects a number of indicators with high performance in predicting the development trend of impact. Based on the Phase Space Reconstruction (PSR) technology, the data is mapped to the high-dimensional space for reconstruction. LSTM is trained to learn the characteristics of high dimensional data, and a prediction model of steeply inclined ultra-thick coal seam rock burst (PSR-LSTM) based on deep learning and multiple chaotic time series is constructed. The results show that each microseismic information index is sensitive to shock warning and has significant correlation with each other. Six microseismic information indexes with high performance in predicting the development trend of shock are selected. The time series of multiple indicators has chaotic characteristics. After phase space reconstruction, LSTM learning and training can effectively enhance the data utilization rate and prediction accuracy of the model. When the prediction time of the constructed PSR-LSTM model is specified as 1 day, the prediction accuracy can reach 0.9135, and the F₁ value can reach 0.9116. All of them are better than the unreconstructed LSTM model. The model can well predict the time series trend and danger level of the rock burst in the excavation roadway of steeply inclined ultra-thick coal seam. The research method can provide reference for the intelligent prediction and early warning of rock burst in the excavation roadway of steeply inclined ultra-thick coal seam.