To address the demand for base-station traffic forecasting in ultra-dense 5G/6G deployments, this study proposes an Enhanced Randomly Mixed Kernel K-Nearest Neighbors algorithm (ER-MKKNN). By fusing a radial basis function kernel with a white-noise kernel into a hybrid kernel function, it overcomes the trade-off bottleneck between nonlinear relationship modeling and noise suppression inherent in single-kernel approaches. Innovatively introducing dual random subsampling on both samples and features, together with a randomized hyperparameter‐interval strategy, markedly improves generalization stability in high-dimensional, sparse scenarios. A dynamic weight‐allocation mechanism based on out-of-bag (OOB) error inversion enhances the algorithm’s robustness to abrupt traffic fluctuations. The accompanying multi-level parallel architecture offers a scalable prediction solution for ultra-dense network topologies. Experimental results show that ER-MKKNN outperforms the best deep-learning models by 4.6%, 63.5%, and 8.6% on RMSE, MAPE, and MAE, respectively, charting a new technical pathway for intelligent network operations and maintenance.