| description abstract | Stormwater best management practices (BMPs) are commonly emplaced to remove contaminants to reduce environmental impacts of runoff, in addition to their flood control function. The selection of BMPs for urban catchments is ideally targeted to local contaminants of concern; however, these decisions often fail to consider observed treatment performance. The International Stormwater BMP Database is a large repository of monitoring data that can inform performance estimates. A major goal of this work is to maximize the usefulness of the BMP database to inform data-driven decision making for BMP selection. However, the database currently lacks sufficient BMP design and construction data to parameterize complex models. In lieu of more complex models, this work uses water quality data from the BMP database to develop and evaluate easy-to-use linear models that, given influent concentration, predict effluent concentrations for a selection of BMPs and stormwater contaminants. Four linear models (percent removal, ordinary least squares regression, Theil–Sen robust line, and a linear decay model with an irreducible concentration parameter) are evaluated using percent bias (PBIAS), Nash–Sutcliffe efficiency (NSE), and root mean square error–standard deviation ratio (RMSE RSR) criteria. Results identify ordinary least squares as the best model and expose patterns in BMP categories, with acceptable model criteria for detention basins, grass swales, media filters, retention ponds, and wetland basins. The ordinary least squares model yielded acceptable model criteria for estimates of total suspended solids removal, as well as total phosphorus, nitrate plus nitrite, and total nitrogen. None of the models met “good” criteria thresholds (PBIAS<10%; NSE>0.6; RSR<25th percentile) for more than 40% of BMP-contaminant combinations, but the ordinary least squares model performed best overall. Ultimately, there is a need for more monitoring projects to report BMP design information to facilitate the use of these data in more complex treatment models. | |
