In this work, we target a distributed conceptual hydrological model Larsim, used for operational flood forecasting. The Larsim model is designed to estimate the water balance in different regions, in high temporal resolution, by conceptualizing the complex, spatially distributed, and highly inter-related hydrological processes. A direct consequence of process conceptualization is that the model parameters exhibit extensive uncertainties, leading to significant uncertainty in hydrologic predictions. However, there has been no proper uncertainty study of the sizeable stochastic parameter set up to now. Runtimes of these hydrologic simulations drastically limit the application of standard, sample-based uncertainty quantification (UQ) algorithms. Moreover, the large number of stochastic parameters may cause the typical "curse of dimensionality" problem, which renders the UQ challenges. We build a surrogate model based on the generalized polynomial chaos expansion (gPCE) to quantify the model parameters' uncertainty on the model predictions. To delay the curse of dimensionality, we rely on the spatially-adaptive sparse grids technique for estimating the coefficients of the gPCE. Finally, we easily estimate the time-varying global sensitivity indices, giving us further insight into the parameters' stochastic importance. Our work bridges the gap between earlier theoretical work on UQ applied to relatively simple simulation models and more complex real-world problems.
«
In this work, we target a distributed conceptual hydrological model Larsim, used for operational flood forecasting. The Larsim model is designed to estimate the water balance in different regions, in high temporal resolution, by conceptualizing the complex, spatially distributed, and highly inter-related hydrological processes. A direct consequence of process conceptualization is that the model parameters exhibit extensive uncertainties, leading to significant uncertainty in hydrologic predictio...
»