Abstract: Accurate decision-making on the residual support force of hydraulic supports during pressurized moving under fragmented roof conditions is crucial for improving intelligent mining efficiency in ultra-thin coal seams and ensuring operational safety. To address this challenge, this study proposed a novel decision-making method based on a Deep Hybrid Kernel Extreme Learning Machine (DHKELM) optimized by an Improved Dung Beetle Optimization (IDBO) algorithm. The DHKELM model was constructed by incorporating an Extreme Learning Machine Autoencoder (ELM-AE) into the Hybrid Kernel Extreme Learning Machine (HKELM) framework, enhancing its feature extraction capability and nonlinear mapping efficiency for complex inputs. Furthermore, the Dung Beetle Optimization (DBO) algorithm was enhanced with ICMIC chaotic mapping, Lévy flight, and a greedy strategy, yielding the IDBO algorithm with improved optimization accuracy and faster convergence. The IDBO algorithm was further employed to optimize the hyperparameters of the DHKELM model, forming the IDBO-DHKELM model. Using field-measured data from hydraulic supports during pressurized moving in a fully mechanized ultra-thin coal seam mining face, key influencing factors of residual support force—including support number, support force before pressurized moving, pushing cylinder inlet pressure, and pushing cylinder stroke variation speed—were identified through visualization and correlation analysis. A residual support force decision-making dataset was subsequently constructed, and the IDBO-DHKELM model was trained and evaluated. Experimental results demonstrate that the proposed IDBO-DHKELM model achieves high decision-making accuracy, with a root mean square error (RMSE) of 0.143, a mean absolute error (MAE) of 0.119, and a coefficient of determination (R²) of 0.971.