Accurate estimation of irrigation return flow plays an important role in the effective management of groundwater, especially in arid and semiarid irrigation regions. However, there is a lack of sufficient research to clarify hydrological process dynamics associated with irrigation return flow. In this study, first, a two-dimensional/three-dimensional model, HYDRUS-2D/3D, was adopted to analyze two different irrigation types in the Delingha Depression, which is located at the northeastern margin of the Qaidam Basin, China. Then, a 3D saturated flow model was established. This study determined the effect of agricultural water application on the dynamics of irrigation return flow. A large difference in the irrigation return-flow coefficient (IRFC) was seen during the growing season; an IRFC of 0.3 was obtained using flood irrigation, whereas ditch irrigation resulted in an IRFC of only 0.1. The lag time of recharge was approximately 150 days. It was necessary to consider the lag time for the 3D numerical model to obtain satisfactory results. Flood irrigation led to a groundwater recharge rate of 90 mm/year. These results indicate that the lag time should be considered when groundwater recharge is estimated or modeled.