Comparison of Uncertainty Analysis for Community Based Watershed Models

Policy-makers rely on predictions from complex watershed models for resource allocations and development decisions. However, most managers do not have access to effective and practical approaches that represent the uncertainty within these systems and within the complex models; nor can they readily incorporate uncertainty into the decision process. Ensemble approaches to uncertainty analysis are a means to link environmental risks to the modeling process. They have been shown to be an improvement over an arbitrary margin-of-safety allocation for Total Maximum Daily Load studies. Although it is demonstrated that this approach greatly increases the reliability of load allocations, as compared to a margin-of-safety approach, uncertainty analysis is computationally intensive and research intensive, which may cause many managers and community-based modelers to rely on simpler, less accurate approaches. This study focuses on reducing the computational requirements for model calibration within an ensemble approach to uncertainty analysis. Latin Hypercube Sampling (LHS), a Monte Carlo approach, and Dynamically Dimensioned Search (DDS), an optimization algorithm, were integrated with the Moore's Creek Hydrologic Simulation Program - Fortran model case study to evaluate their ii performance as model calibration methods. A spectrum of model fit, including 11 hydrologic measures and 3 water quality measures, were used for model calibration. The DDS algorithm was successfully applied to the Moore's Creek case study using a weighted composite objective function, which combined the 14 calibration criteria into a single value for optimization. The resulting sets of parameter values were shown to include less error relative to the calibration criteria for DDS than the LHS method, while reducing computational requirements by half. Comparison measures used in the research, an integral of the prediction bounds and the number of data points captured by the bounds, link the resulting ensemble of model predictions to empirical data. The results show how these measures may be used to guide arbitrary decisions in an uncertainty analysis study.

Note: Abstract extracted from PDF text