Performance comparison of green roof hydrological models for full-scale field sites

Green roofs can be valuable components in sustainable urban drainage systems, and hydrological models may provide useful information about the runoff from green roofs for planning purposes. Various models have been proposed in the literature, but so far no papers have compared the performance of multiple models across multiple full-size green roofs. This paper compared 4 models: the conceptual models Urbis and SWMM and the physically-based models Hydrus-1D and Mike SHE, across two field sites (Lyon, France and Umea, Sweden) and two calibration periods for each site. The uncertainty and accuracy of model predictions were dependent on the selected calibration site and period. Overall model predictions from the simple conceptual model Urbis were least accurate and most uncertain; predictions from SWMM and Mike SHE were jointly the best in terms of raw percentage observations covered by their flow prediction intervals, but the uncertainty in the predictions in SWMM was smaller. However, predictions from Hydrus were more accurate in terms of how well the observations conformed to probabilistic flow predictions. Mike SHE performed best in terms of total runoff volume. In Urbis, SWMM and Hydrus uncertainty in model predictions was almost completely driven by random uncertainty, while parametric uncertainty played a significant role in Mike SHE. Parameter identifiability and most likely parameter values determined with the DREAM Bayesian algorithm were found to be inconsistent across calibration periods in all models, raising questions about the generalizability of model applications. Calibration periods where rainfall retention was highly variable between events were more informative for parameter values in all models.