Evaluation of two statistical approaches for estimating pollutant loads at adjacent combined sewer overflow structures

Quantifying pollutant loads from combined sewer overflows (CSOs) is necessary for assessing impacts of urban drainage on receiving water bodies. Based on data obtained at three adjacent CSO structures in the Louis Fargue catchment in Bordeaux, France, this study implements multiple linear regression (MLR) and random forest regression (RFR) approaches to develop statistical models for estimating emitted loads of total suspended solids (TSS). Comparison between hierarchical clustering selection and random selection of CSO events for model calibration is included in model development. The results indicate that selection of the model's explanatory variables depends on both the type of approach and the CSO structure. By using the cluster technique to select representative events for model calibration, model predictability is generally improved. For the available dataset, MLR may have advantages over RFR in terms of verification performance and lower range of error due to splitting events for calibration and verification. But RFR model uncertainty bands are considerably narrower than the MLR ones.