Uncertainty analysis in urban drainage modelling: should we break our back for normally distributed residuals?

This study presents results on the assessment of the application of a Bayesian approach to evaluate the sensitivity and uncertainty associated with urban rainfall-runoff models. The software MICA was adopted, in which the prior information about the parameters is updated to generate the parameter posterior distribution. The likelihood function adopted in MICA assumes that the residuals between the measured and modelled values have a normal distribution. This is a trait of many uncertainty/sensitivity procedures. This study compares the results from three different scenarios: (i) when normality of the residuals was checked but if they were not normal then nothing was done (unverified); (ii) normality assumption was checked, verified (using data transformations) and a weighting strategy was used that gives more importance to high flows; and (iii) normality assumption was checked and verified, but no weights were applied. The modelling implications of such scenarios were analysed in terms of model efficiency, sensitivity and uncertainty assessment. The overall results indicated that verifying the normality assumption required the models to fit a wider portion of the hydrograph, allowing a more detailed inspection of parameters and processes simulated in both models. Such an outcome provided important information about the advantages and limitations of the models' structure.

Stormwater quality models: performance and sensitivity analysis.

The complex nature of pollutant accumulation and washoff, along with high temporal and spatial variations, pose challenges for the development and establishment of accurate and reliable models of the pollution generation process in urban environments. Therefore, the search for reliable stormwater quality models remains an important area of research. Model calibration and sensitivity analysis of such models are essential in order to evaluate model performance; it is very unlikely that non-calibrated models will lead to reasonable results. This paper reports on the testing of three models which aim to represent pollutant generation from urban catchments. Assessment of the models was undertaken using a simplified Monte Carlo Markov Chain (MCMC) method. Results are presented in terms of performance, sensitivity to the parameters and correlation between these parameters. In general, it was suggested that the tested models poorly represent reality and result in a high level of uncertainty. The conclusions provide useful information for the improvement of existing models and insights for the development of new model formulations.

Analysis of parameter uncertainty of a flow and quality stormwater model.

Uncertainty is intrinsic to all monitoring programs and all models. It cannot realistically be eliminated, but it is necessary to understand the sources of uncertainty, and their consequences on models and decisions. The aim of this paper is to evaluate uncertainty in a flow and water quality stormwater model, due to the model parameters and the availability of data for calibration and validation of the flow model. The MUSIC model, widely used in Australian stormwater practice, has been investigated. Frequentist and Bayesian methods were used for calibration and sensitivity analysis, respectively. It was found that out of 13 calibration parameters of the rainfall/runoff model, only two matter (the model results were not sensitive to the other 11). This suggests that the model can be simplified without losing its accuracy. The evaluation of the water quality models proved to be much more difficult. For the specific catchment and model tested, we argue that for rainfall/runoff, 6 months of data for calibration and 6 months of data for validation are required to produce reliable predictions. Further work is needed to make similar recommendations for modelling water quality.