Performance-evaluation of urban drainage models

By introducing digital twins (Pedersen et al., 2021) in the work with the urban drainage system there is an expectancy that the simulation models included in the digital twin perform sufficiently all the time. But the water industry currently has no good and generally agreed way to validate and quantify this. The simulation models have been used in the community for decades and results from hydrodynamic urban drainage models form the basis of many decisions in redesign and upgrading of the current urban drainage systems. They are expected to be able to simulate the current situation when they are well calibrated. However, what is often happening in the utilities is only heuristic calibration of parameters in the model in cases where sufficient new observations are available. With the introduction of digital twins, the simulation models need to perform well all the time with many objectives in mind, or at least give an insight in the model performance for the current situation in order to make good, informed decisions. We therefore need to understand the model performance across many operational conditions and across multiple locations by learning from the locations where monitoring occurs. We need to get an overview of under which criteria the model performs acceptably and which not. By extracting so called signatures (Gupta et al., 2008) that summarize certain characteristics of a timeseries, we are able to identify and diagnose the model performance for specific objectives, instead of applying average distance measures such as RMSE on entire time series.

For this paper we focus on the objective of quantifying combined sewer overflow and use signatures relevant to this objective when evaluating the model performance in a simple way.