Reducing uncertainties in urban drainage models by explicitly accounting for timing errors in objective functions

Abstract

Traditional hydrological objective functions may penalize models that reproduce hydrograph shapes well, but with some shift in time; especially for urban catchments with a fast hydrological response. Hydrograph timing is not always critical, so this paper investigates alternative objective functions (based on the Hydrograph Matching Algorithm) that try to mimic visual hydrograph comparison. A modified version of the Generalized Likelihood Uncertainty Estimation is proposed to compare regular objective functions with those that account for timing errors. This is applied to 2-year calibration and validation data sets from an urban catchment. Results show that such objective functions provide equally reliable model predictions (they envelop the same fraction of observations), but with more precision, i.e. smaller estimated uncertainty of model predictions. Additionally, identifiability of some model parameters improved. Therefore objective functions based on the Hydrograph Matching Algorithm can be useful to reduce uncertainties in urban drainage modelling.

All models are wrong, but are they useful? Assessing reliability across multiple sites to build trust in urban drainage modelling

“…Methods for handling unrealistic anomalies can be found in the literature (e.g. Clemens-Meyer et al, 2021). For this study, the simple data-cleaning approach by Pedersen et al (2021b) was applied, using five techniques (low data quality determined by the supervisory control and data acquisition (SCADA) system manufacturer, manually removed data, out of physical bounds considering the specific sensor, frozen sensor signal, and outlier data as assessed by an operator).…”

“…Determining a best model parameter set can be accommodated by, for example, the post-processing of errors from historical data (Ehlers et al, 2019) or a signature-based evaluation (Gupta et al, 2008). Research has focused on quantifying the uncertainty in model output (Deletic et al, 2012), for example, by using the generalized likelihood uncertainty estimator (GLUE) method (Beven and Binley, 1992), which estimates the error term (del Giudice et al, 2013) or looks into the timing errors related to uncertainty procedures (Broekhuizen et al, 2021). Recently, urban drainage modelling research has focused increasingly on model calibration, i.e.…”

“…Monitoring is not always easy, as urban drainage systems are rough environments, for example, with particles that can settle and clog equipment, and flow meters that have difficulties measuring flow correctly when the pipe is partly filled -a problem that could be solved in the future with image-based flow monitoring (Meier et al, 2022). Water level meters are acknowledged to be fairly accurate, but they can suffer from missing data values, which can be tackled using several methods, as described in Clemens-Meyer et al (2021). In the present paper, a model evaluation is conducted for sites with only level meters installed because, in the future, low-cost level meters are expected to be installed at an increasing number of sites in urban drainage systems (Eggimann et al, 2017;Kerkez et al, 2016;Shi et al, 2021).…”

“…The urban drainage community has only recently started developing tools for anomaly detection in the urban drainage system caused by physical events, e.g. pipe blockage, pumps that did not start as planned, and other sudden activities in the system (Palmitessa et al, 2021;Clemens-Meyer et al, 2021). Although many studies have shown that rainfall varies spatially at scales significant to urban drainage modelling (Gregersen et al, 2013;Thomassen et al, 2022), a method that accounts for unrealistic representation of rainfall spatial variability when using rainfall data from point gauges in model assessment has not yet, to the authors' awareness, been investigated.…”