Abstract: With the accelerated transformation of China’s coal mining industry toward deep, high-intensity, and intelligent development, the traditional “general prevention and control” model for coal mine dynamic disasters faces bottlenecks such as imprecise decision-making, delayed response, and unclear target identification. To address key challenges—including the difficulty of dynamic characterization of geological anomalies, limited integration between data and physical mechanisms, and delays in the optimization and adjustment of control parameters—this study proposes a full-time and spatial early warning and disaster control optimization method based on digital twins. By constructing a technical system of “disaster geologic body dynamic analysis — physical-virtual joint deduction of disaster evolution — disaster control decision-making optimization,” this research innovatively promotes a shift from experience-driven to “data–model” dual-driven disaster prevention and control. It enhances both the visualization of disaster evolution and the traceability of control processes. The main innovations of this study include: ① establishing a large language model-guided geological knowledge extraction method, which overcomes technical challenges in converting unstructured geological data into intelligent 3D geological models. ② proposing a variational geometry-constrained dynamic meshing algorithm and a three-stage optimization architecture of boundary conditions—material parameters—constitutive models, which breaks through theoretical bottlenecks in real-time model updating during mining operations while incorporating field monitoring data. ③ constructing a CNN–LSTM–Attention hybrid neural network architecture to reveal the correlation mechanism between the spatiotemporal evolution of the mining stress field and microseismic event clusters. ④ developing a dynamic optimization method for disaster control parameters based on deep learning surrogate models, enabling the precise regulation of measures such as pressure relief through borehole drilling. This study highlights the coupling among field data flow, numerical model flow, and disaster control flow, enabling forward-looking disaster prevention through continuous model calibration and predictive analysis. The “Coal Mine Power Disaster Digital Twin Platform” developed herein has been deployed and validated in typical disaster-prone mines, forming a closed-loop control framework of “geological anomaly identification → disaster process reconstruction → prevention and control scheme iteration.” This system effectively addresses issues of delayed feedback and fragmented information in conventional disaster prevention decision-making. It significantly improves the visualization of disaster evolution and the timeliness of control decisions, enhances the efficiency of mine safety operations, and provides new technological approaches and theoretical support for the intelligent prevention and control of coal mine dynamic disasters. Furthermore, it offers a digital and intelligent reference path for managing major hazards in the broader context of deep energy development.