On the Predictive Uncertainty of a Distributed Hydrologic Model

We use models to simulate the real world mainly for prediction purposes. However, since any model is a simplification of reality, there remains a great deal of uncertainty even after the calibration of model parameters. The model?s identifiability of realistic model parameters becomes questionable when the watershed of interest is small, and its time of concentration is shorter than the computational time step of the model. To improve the discovery of more reliable and more realistic sets of model parameters instead of mathematical solutions, a new algorithm is needed. This algorithm should be able to identify mathematically inferior but more robust solutions as well as to take samples uniformly from high-dimensional search spaces for the purpose of uncertainty analysis. Various watershed configurations were considered to test the Soil and Water Assessment Tool (SWAT) model?s identifiability of the realistic spatial distribution of land use, soil type, and precipitation data. The spatial variability in small watersheds did not significantly affect the hydrographs at the watershed outlet, and the SWAT model was not able to identify more realistic sets of spatial data. A new populationbased heuristic called the Isolated Speciation-based Particle Swarm Optimization (ISPSO) was developed to enhance the explorability and the uniformity of samples in high-dimensional problems. The algorithm was tested on seven mathematical functions and outperformed other similar algorithms in terms of computational cost, consistency, and scalability. One of the test functions was the Griewank function, whose number of minima is not well defined although the function serves as the basis for evaluating multi-modal optimization algorithms. Numerical and analytical methods were proposed to count the exact number of minima of the Griewank function within a hyperrectangle. The ISPSO algorithm was applied to the SWAT model to evaluate the performance consistency of optimal solutions and perform uncertainty analysis in the Generalized Likelihood Uncertainty Estimation (GLUE) framework without assuming a statistical structure of modeling errors. The algorithm successfully found hundreds of acceptable sets of model parameters, which were used to estimate their prediction limits. The uncertainty bounds of this approach were comparable to those of the typical GLUE approach.