To enhance the efficient utilization of precious water resources in the source region of the Yellow River and improve flood forecasting accuracy by leveraging historical flood information, this study proposes a flood classification and parameter optimization method integrating Principal Component Analysis (PCA) and K-Means clustering, focusing on floods at the Tangnaihai Station (the inflow control station of the Longyangxia Reservoir). A multi-dimensional flood characteristic index system was constructed based on long-term hydrological data from 1956 to 2023. PCA extracted four principal components with cumulative variance contributions exceeding 90%, and K-Means clustering categorized 77 historical floods into three types: short-duration gradual-rise floods, uniform wide-peak floods, and prolonged high-peak floods. These classified floods were simulated using the Vertical Mixed Runoff Generation (VMRG) model and the Xinanjiang model. Results demonstrate that classified floods achieved higher simulation accuracy than unclassified ones. During the calibration period, the VMRG model improved flood peak and volume accuracy by 1.45% and 0.68%, respectively, while the Xinanjiang model showed enhancements of 1.58% and 0.34%. In the validation period, optimized parameters constrained peak flow errors within 10%, achieved 100% accuracy in peak occurrence timing, and reduced maximum flood volume errors by up to 12.78%. This study confirms that the PCA-K-Means integrated flood classification and parameter optimization method significantly enhances model forecasting precision, providing scientific support for flood control safety and efficient water resource management in the Yellow River Basin.