Keywords
Human intervention, river engineering, hydraulic modelling, uncertainty analysis, computational efficiency
Start Date
16-9-2020 10:00 AM
End Date
16-9-2020 10:20 AM
Abstract
Computer models play a major role in the design of the Dutch river system. Human intervention in the river system is strictly managed – any design must either lead to a lowering of flood levels or mitigation measures must ensure flood risk does not increase. Major engineering works to lower flood risk have been carried out in the past 20 years. The costs of these works exceeded 13 million Euro per centimetre of flood level lowering. Computer models are expected to predict effects in the order of centimetres with high accuracy, but limited data at flood discharges prevents empirical validation of such claims. At the same time, model output uncertainty during floods is estimated to be in the order of decimetres – an order of magnitude higher than the predicted effect of most river engineering works. Here we present two approaches and major findings from our research on the model uncertainty for human intervention in rivers. First, we developed a method based on Bayesian (non-) parametric inference to significantly decrease the computational costs that are typical for the high-resolution physics-based models required in river engineering. Using this, we found that the uncertainty in predicted effects is much smaller than those of the absolute water levels, but significant enough to affect design decisions. Second, we applied a Bayesian approach to study non-stationarity in the observational record in support of model output and look for empirical evidence of water level reduction due to human intervention. While the observational record holds few floods of high discharge, we anticipated that the major engineering works performed over the past two decades would be discernible in the data. However, it was not, which is why the search for data to validate model predictions of water level lowering continues. Meanwhile, in lieu of limited empirical evidence, we argue that quantification of model uncertainty should be a key ingredient in river engineering.
How two Bayesian methods pave the way for uncertainty analysis in river engineering
Computer models play a major role in the design of the Dutch river system. Human intervention in the river system is strictly managed – any design must either lead to a lowering of flood levels or mitigation measures must ensure flood risk does not increase. Major engineering works to lower flood risk have been carried out in the past 20 years. The costs of these works exceeded 13 million Euro per centimetre of flood level lowering. Computer models are expected to predict effects in the order of centimetres with high accuracy, but limited data at flood discharges prevents empirical validation of such claims. At the same time, model output uncertainty during floods is estimated to be in the order of decimetres – an order of magnitude higher than the predicted effect of most river engineering works. Here we present two approaches and major findings from our research on the model uncertainty for human intervention in rivers. First, we developed a method based on Bayesian (non-) parametric inference to significantly decrease the computational costs that are typical for the high-resolution physics-based models required in river engineering. Using this, we found that the uncertainty in predicted effects is much smaller than those of the absolute water levels, but significant enough to affect design decisions. Second, we applied a Bayesian approach to study non-stationarity in the observational record in support of model output and look for empirical evidence of water level reduction due to human intervention. While the observational record holds few floods of high discharge, we anticipated that the major engineering works performed over the past two decades would be discernible in the data. However, it was not, which is why the search for data to validate model predictions of water level lowering continues. Meanwhile, in lieu of limited empirical evidence, we argue that quantification of model uncertainty should be a key ingredient in river engineering.
Stream and Session
false