现有的矿山巷道液压支架载荷预测方法通常假设静态时空开采布置，忽略了远场围岩的动载荷，影响了冲击地压预测的准确性。为了确保安全开采，必须对潜在的冲击地压进行实时和精确的预测评估。提出一种基于数字孪生和机器学习的麻雀搜索算法–随机森林(SSA–RF)预测方法。通过分析支护系统与围岩的相互作用，建立了两柱式支架的数字孪生体模型，基于数据驱动实现物理实体与其数字孪生体之间的交互映射与同步反馈。对比分析两柱式支架升柱过程中姿态变量的计算值与真实值发现，与支架物理实体相比，数字孪生体模型的角度平均误差为0.14°，长度平均误差为6.15 mm，符合精度要求。此外，麻雀搜索算法用于优化随机森林中决策树和节点特征的数量。与使用单一预测模型相比，SSA–RF预测模型的建立提高了收敛速度和优化能力。试验结果表明，与长短期记忆(LSTM)、随机森林(RF)和支持向量机(SVM)等预测算法相比，笔者提出的SSA–RF方法表现最优，其在中部支架和端头支架数据集上的预测准确率分别达到85.89%和91.09%。此外，研究发现，工作面中部支架所在区域顶板易出现断裂失稳现象，会破坏工作面中部区域垂直方向的应力支撑条件，从而导致中部支架载荷变化范围更大预测精度略低于端头支架。上述研究结果将对进一步研究煤矿冲击地压发生机理和准确预测潜在冲击地压提供一定理论借鉴。

Existing methods for predicting hydraulic support loads in mine roadways usually assume a static spatio-temporal mining arrangement, which ignores the dynamic loads of the far-field surrounding rocks and affects the accuracy of rockburst prediction. In order to ensure safe mining, real-time and accurate predictive assessment of potential rockburst is necessary. In this paper, a Sparrow Search Algorithm-Random Forest (SSA–RF) prediction method based on digital twin and machine learning is proposed. By analyzing the interaction between the support system and the surrounding rock, a digital twin model of the two-column support is established, and the interaction mapping and synchronous feedback between the physical entity and its digital twin are realized based on data driving. By comparing and analyzing the calculated and real values of the attitude variables during the column lifting process of the two-column support, it is found that compared with the physical entity of the support, the digital twin model has an average error of 0.14° in angle and 6.15 mm in length, which is in line with the accuracy requirements. In addition, the Sparrow Search Algorithm was used to optimize the number of decision trees and node features in the Random Forest. Compared with using a single prediction model, the SSA–RF prediction modeling improves the convergence speed and optimization ability. The experimental results show that the SSA–RF method proposed in this paper performs optimally compared with prediction algorithms such as Long Short-Term Memory (LSTM), Random Forest (RF) and Support Vector Machine (SVM), and its prediction accuracy reaches 85.89% and 91.09% on the central support and end support data sets, respectively. In addition, it is found that the roof in the area of the central support is prone to fracture instability, which will destroy the vertical stress support conditions in the central area of the working face, thus leading to a larger range of load variations in the central support with a slightly lower prediction accuracy than that of the end support. The above results provide some theoretical reference for further research on the occurrence mechanism of rockburst in coal mine and accurate prediction of potential rockburst.