To improve the accuracy of flood forecasting in semi-humid and semi-arid regions, the runoff generation characteristics of these regions were analyzed, as well as indexes such as hydrometeorology, soil and topography, and vegetation landcover. The grid hierarchical-clustering runoff generation pattern identification method (abbreviated as GRGI) was established through the optimization of identification indexes, hierarchical weight evaluation, and rolling correction clustering. Case verification results of the Qianyang Watershed in the middle reaches of the Yellow River show that the GRGI method can accurately identify the spatial and temporal distribution of the saturation-excess and infiltration-excess runoff generation modes in the Qianyang Watershed, with the saturation-excess pattern grid dominating, accounting for more than 70% of the total. Compared with those of the Grid-XAJ and Grid-GA models, the qualified rates of runoff depth error, peak flow error, and peak time difference of the Grid-XAJ-SIDE model improved based on runoff generation identification rules are increased by about 12%, 35%, and 12%, respectively.