To address the limited joint quantification of flow structures and water purification performance in shallow rivers and lakes and in constructed wetlands, as well as the difficulty of validating models at different scales under matched scenarios, two hydrodynamic models were developed: a graphics processing unit (GPU)-accelerated two-dimensional model based on the shallow water equations and a three-dimensional Reynolds-averaged Navier–Stokes (RANS) model with porous-medium parameterization. Annular-flume experiments were conducted under still-water, current, and wave conditions using Pontederia cordata, Thalia dealbata, and Acorus calamus as monocultures, together with an equal-number mixed-species treatment. Benchmark tests showed that the two-dimensional model reliably reproduced mixed flow with a hydraulic jump, steady subcritical flow, wetting and drying, and coupled runoff generation and routing, with a maximum relative error not exceeding 8.28%. Using experimental data from a rigid-vegetation channel as a benchmark, the two-dimensional model was more suitable for assessing overall flow resistance and longitudinal water-level trends at the regional scale, whereas the three-dimensional model was better suited to resolving localized nonlinear flow structures under conditions of pronounced local disturbance. Water-quality responses were both indicator- and species-dependent: current and wave conditions enhanced total nitrogen (TN) removal by P. cordata, whereas still-water conditions generally favored total phosphorus (TP) and ammonium nitrogen (〖"NH" 〗_"4" ^"+" "-N" ) removal in several plant treatments. Chemical oxygen demand (COD) showed no consistent response to hydrodynamic regime, and the equal-number mixed-species treatment did not exhibit a consistent synergistic purification benefit. These findings provide a scientific basis for model selection, plant configuration, and hydrodynamic regulation in the ecological restoration of rivers, lakes, and wetlands.