Predicting flood discharge plays a pivotal role in water resource management, particularly within highly complex hydrological basins. Hydrological models frequently encounter significant challenges in replicatingthe non-linear interactions between atmospheric forcing and complex topographic responses. This study proposes a novel hybrid deep learning architecture, MetNet-2-SwinTransformer, designed to integrate the temporal feature extraction capabilities of MetNet-2 with the multi-scale spatial hierarchical modeling of the Swin Transformer. The model’s performance was evaluated using discharge data from the Karaj watershed to assess its performance in capturing complex flow dynamics and extreme hydrological events. The performance of the proposed hybrid model was rigorously compared against individual MetNet-2 and Swin Transformer architectures using multiple statistical metrics, including Nash-Sutcliffe Efficiency (NSE), Coefficient of Determination (R), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The results indicate that the hybrid model exhibits absolute superiority over the baseline models, attaining exceptional predictive accuracy with NSE and RRR values surpassing 0. 95. While individual models exhibited significant limitations in reconstructing peak flows and managing high-magnitude fluctuations particularly in stations with extreme discharge variability the hybrid model effectively mitigated these errors through a synergistic integration of temporal and spatial features. The hybrid model was able to overcome the peak smoothing phenomenon common in standard transformer models and reconstruct flood hydrograph curves with very high accuracy. The findings indicate that the integration of sliding-window mechanisms with temporal forecasting enables the model to capture small-scale spatial effects and complex topographic influences with high precision, yielding correlations and NSE values exceeding 0.95, while reducing error magnitudes to below 0.1 at stations with lower peak discharges. Although the error amount was higher at other stations than at 3 stations, this is acceptable due to the nature of hydrological data. This research confirms that the proposed hybrid approach offers a robust and highly accurate framework for flood forecasting in complex basins, providing a reliable foundation for developing advanced early warning systems and enhancing water resource decision-making processes.