Structural uncertainty in a river water quality modelling system

There are many sources of uncertainty in modelling systems including uncertainty due to parameter estimations, input data and structure of the system. In this paper, modelling system structure is understood to be the algorithms and equations used to describe and calculate processes. Specific attention is given to the uncertainty in the equations, which link models interactively in a modelling system and algorithms, which pass information between the models. This paper places emphasis on structural uncertainty in modelling systems since this topic has received very little attention in research compared to the wealth of literature on parameter and input data uncertainty This imbalance is partly due to the difficulty in quantifying structural uncertainty and an example is given showing how this may be done. The modelling system used for this study incorporates the coupling of three models from the WASPS package (Water quality Analysis Simulation Program) developed by the US Environmental Protection Agency: DYNHYD (hydrodynamics), EUTRO (dissolved oxygen, nutrient and phytoplankton dynamics) and TOXI (transport and transformation of sediments and micro-pollutants). The three models were coupled in the high level architecture (HLA) platform to enable interactions between the models during simulations. This platform allowed ease of implementing Monte Carlo simulation runs, which were used for the uncertainty analyses. he modelling system structure focuses on the exchange of data between EUTRO and TOXI in simulating two processes: (i) sorption of heavy metals to suspended solids as a function of the fraction of organic carbon constituting the solids and (ii) phytoplankton growth limitation by light extinction through the water column as a function of the concentration of suspended solids and chlorophyll a in the water. Both processes are described by empirical relationships using regression curves derived from field data. The variance in the y-axis interceptions of the curves (also called regression uncertainty) is taken to be the uncertainty bounds for the Monte Carlo simulations used for the uncertainty analyses. The results show that structural uncertainty can be more significant than parameter and input data uncertainty especially if the processes are very sensitive to the state variables. (C) 2007 Elsevier B.V. All rights reserved.