Located in the transition zone from the Qinghai-Tibet Plateau to the Sichuan Basin, the Minjiang River Basin is characterized by strong spatiotemporal variability in runoff and is heavily regulated by a dense cluster of reservoirs. To address the limitations of existing data-driven models that frequently overlook spatial hydrological connectivity and the impacts of artificial regulation, this study proposes a Random Forest-based mid- to long-term runoff forecasting model that integrates cascade relationships and reservoir regulation features, focusing on key hydrological control stations within the region. The model employs an out-of-fold cross-validation strategy to prevent information leakage, thereby facilitating the cascading transmission of upstream simulated streamflow, associated uncertainties, and reservoir operation information to downstream nodes. The results demonstrate that: (1) the synergistic coupling of the Asian Polar Vortex, the East Asian Trough, and the Tibetan Plateau indices dominates the interannual variability of runoff fluctuations in the region. (2) The cascade mechanism effectively enhances both physical consistency and forecasting accuracy, with the basin-wide average Nash-Sutcliffe Efficiency (NSE) increasing from 0.616 to 0.638 during the validation period. (3) Error propagation exhibits a spatially divergent pattern, where errors accumulate longitudinally along single river reaches, whereas forecasting robustness is enhanced at multi-tributary confluence nodes due to the mutual cancellation of biases from different branches. (4) The proposed model yields the most significant improvements in low-flow simulations, reducing the mean absolute deviation from 15.08% to 8.79% and accurately capturing artificial flow compensation signals during the dry season. This study not only effectively improves runoff forecasting accuracy in complex basins but also provides a transferable technical framework for regions heavily impacted by human activities.