Deterioration models can be successfully deployed only if decision-makers trust the modelling outcomes and are aware of model uncertainties. Our study aims to address this issue by developing a set of clearly understandable metrics to assess the performance of sewer deterioration models from an end-user perspective. The developed metrics are used to benchmark the performance of a statistical model, namely, GompitZ based on survival analysis and Markov-chains, and a machine learning model, namely, Random Forest, an ensemble learning method based on decision trees. The models have been trained with the extensive CCTV dataset of the sewer network of Berlin, Germany (115,258 inspections). At network level, both models give satisfactory outcomes with deviations between predicted and inspected condition distributions below 5%. At pipe level, the statistical model does not perform better than a simple random model, which attributes randomly a condition class to each inspected pipe, whereas the machine learning model provides satisfying performance. 66.7% of the pipes inspected in bad condition have been predicted correctly. The machine learning approach shows a strong potential for supporting operators in the identification of pipes in critical condition for inspection programs whereas the statistical approach is more adapted to support strategic rehabilitation planning.

As in most of the cities around the world, in the last 30 years Latin-American ones have focused on investing in building infrastructure to provide sewer and water services to the communities. However, these infrastructures are going aging day to day. The municipalities need to extend management activities by the development of support tools such as deterioration models to face the aging problem. In the literature of sewer asset management, SVM has been a useful tool to predict and forecast the structural condition of pipes. In this work, the use of differential evolution method as optimization tool was implemented to find the optimal hyper-parameters for SVM models. The SVM models were applied in the main cities of Colombia (Bogota and Medellin) given as a result that the optimized SVM model provides less than 5% of deviation in the prediction of structural conditions in both cities.

UV/Vis spectrophotometers have been used for one decade to monitor water quality in various locations: sewers, rivers, wastewater treatment plants (WWTPs), tap water networks, etc. Resulting equivalent concentrations of interest can be estimated by three ways: i) by manufacturer global calibration; ii) by local calibration based on the provided global calibration and grab sampling; iii) by advanced calibration looking for relations between UV/Vis spectra and corresponding concentrations from grab sampling. However, no study has compared the applied methods so far. This collaborative work presents a comparison between five different methods. A Linear Regression (LR), Support Vector Machine (SVM), EVOlutionary algorithm method (EVO) and Partial Least Squares (PLS) have been applied on various data sets (sewers, rivers, WWTPs under dry, wet and all weather conditions) and for three water quality parameters: TSS, COD total and dissolved. Two criteria (r2 and Root Mean Square Error RMSE) have been calculated - on calibration and verification data subsets - to evaluate accuracy and robustness of the applied methods. Values of criteria have then been statistically analysed for all and separated data sets. Non-consistent outcomes come through this study. According to the Kruskal-Wallis test and RMSEs, PLS and SVM seem to be the best methods. According to uncertainties in laboratory analysis and ranking of methods, LR and EVO appear more robust and sustainable for concentration estimations. Conclusions are mostly independent of water matrices, weather conditions or concentrations investigated.

This paper reports about experiences gathered from five on-line monitoring campaigns in the sewer systems of Berlin (Germany), Graz (Austria), Lyon (France) and Bogota (Colombia) using UV-VIS spectrometers and turbidimeters. The influence of local calibration on the quality of on-line COD measurements of wet weather discharges has been assessed. Results underline the need to establish local calibration functions for both UV-VIS spectrometers and turbidimeters. It is suggested to practitioners to calibrate locally their probes using at least 15-20 samples. However, these samples should be collected over several events and cover most of the natural variability of the measured concentration. For this reason, the use of automatic peristaltic samplers in parallel to on-line monitoring is recommended with short representative sampling campaigns during wet weather discharges. Using reliable calibration functions, COD loads of CSO and storm events can be estimated with a relative uncertainty of approximately 20 %. If no local calibration is established, concentrations and loads are estimated with strong errors questioning the reliability and meaning of the on-line measurement. Similar results have been obtained for TSS measurements.

The present study aims to explore the relationship between rainfall variables and water quality/quantity characteristics of combined sewer overflows (CSO), by the use of multivariate statistical methods and online measurements at a principal CSO outlet in Berlin (Germany). Canonical correlation results showed that the maximum and average rainfall intensities are the most influential variables to describe CSO water quantity and pollutant loads whereas the duration of the rainfall event and the rain depth seem to be the most influential variables to describe CSO pollutant concentrations. The analysis of Partial Least Squares (PLS) regression models confirms the findings of the canonical correlation and highlights three main influences of rainfall on CSO characteristics: (i) CSO water quantity characteristics are mainly influenced by the maximal rainfall intensities, (ii) CSO pollutants concentrations were found to be mostly associated with duration of the rainfall and (iii) pollutants loads seemed to be principally influenced by dry weather duration before the rainfall event. The prediction quality of PLS models is rather low (R² < 0.6) but results can be useful to explore qualitatively the influence of rainfall on CSO characteristics.