A Stormwater Management Framework for Predicting First Flush Intensity and Quantifying its Influential Factors

Despite numerous applications of Random Forest (RF) techniques in the water-quality field, its use to detect first-flush (FF) events is limited. In this study, we developed a stormwater management framework based on RF algorithms and two different FF definitions (30/80 and M(V) curve). This framework can predict the FF intensity of a single rainfall event for three of the most detected pollutants in urban areas (TSS, TN, and TP), yielding satisfactory results (30/80: accuracyaverage\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$accuracy_{average}$$\end{document} = 0.87; M(V) curve: accuracyaverage\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$accuracy_{average}$$\end{document} = 0.75). Furthermore, the framework can quantify and rank the most critical variables based on their level of importance in predicting FF, using a non-model-biased method based on game theory. Compared to the classical physically-based models that require catchment and drainage information apart from meteorological data, our framework inputs only include rainfall-runoff variables. Furthermore, it is generic and independent from the data adopted in this study, and it can be applied to any other geographical region with a complete rainfall-runoff dataset. Therefore, the framework developed in this study is expected to contribute to accurate FF prediction, which can be exploited for the design of treatment systems aimed to store and treat the FF-runoff volume.