Rethinking modelling of particulate pollutants in combined sewer overflows (CSOs): A focus on model structure

Abstract

The persistent challenge of combined sewer overflows (CSOs) in urban drainage systems is exacerbated by climate change and urban growth, with increased attention on water quality historically overshadowed by water quantity monitoring. Modelling CSO water quality challenges is affected by several known challenges, especially for particulate pollutants (i.e., data uncertainties, overparameterization, and non-transferability). This study assesses the impacts of model structure and output resolution (aggregated yearly, inter-event and intra-event basis) on model performance when predicting particulate pollutants levels during CSO events. Four model structures are compared for their ability to simulate the TSS discharge load profile at the inlet of a CSO chamber in Graz, Austria, using Mean Absolute Percentage Error (MAPE) and Dynamic Time Warping (DTW) to assess accuracy and profile similarity with observed data. The model structures include two physics-based (detailed hydrodynamic, conceptual) and two data-driven approaches (hybrid machine learning, empirical). Alternative models are proposed to improve model performance, considering a multi-model, a stochastic approach, and an event-based clustering. We showed that data-driven models captured in-sewer processes that are unexplained and not incorporated in physical process-based models. Our results underline the high inter-event variability of CSO pollutant dynamics, showing how a uniform deterministic modelling approach for all wet-weather events leads to poor performance. Intra-event assessment shows significant deficiencies across all models. The use of stochastic approaches and event clustering techniques did not improve to better model performance notably, advocating for a new generation of modelling approaches that explicitly consider the highly spatial and temporal heterogeneity of in-sewer processes.