Citation:

HANTUSH, M. M. AND L. KALIN. A METHODOLOGY FOR ESTIMATING UNCERTAINTY OF A DISTRIBUTED HYDROLOGIC MODEL: APPLICATION TO POCONO CREEK WATERSHED. Presented at AWRA 2006 Annual Water Resources Conference, Baltimore, MD, November 06 - 09, 2006.

Impact/Purpose:

To present a hybrid approach combining time series analysis and nonparametric probabilistic method to model and synthesize prediction errors of a calibrated watershed model.

Description:

Utility of distributed hydrologic and water quality models for watershed management and sustainability studies should be accompanied by rigorous model uncertainty analysis. However, the use of complex watershed models primarily follows the traditional {calibrate/validate/predict} procedure often without proper characterization of model uncertainty and post-validation forecast evaluation. This study presents a hybrid approach combining time series analysis and nonparametric probabilistic method to model and synthesize prediction errors of a calibrated watershed model. The site of investigation is the Pocono Creek watershed which is located in Eastern Pennsylvania and drains into one of the main tributaries of the Delaware River. The watershed is threatened by high population growth and rapid urbanization and is targeted for a sustainable management planning. To that end, a calibrated and validated Soil Water Assessment Tool (SWAT) is employed to address the potential alterations in the hydrologic response of the watershed to projected land-use changes. Using three-day moving average of streamflows as surrogates to model computed daily flows, ARIMA time series model is fit to SWAT three-day moving average flow prediction errors (observations minus model estimates). The fitted stochastic model is used to construct uncertainty bounds for daily (i.e., three-day moving average) model predictions and generate a forecast (SWAT predictions plus simulated errors) which is tested against newly obtained streamflow data. Monte Carlo simulations over 20-year periods show remarkably low uncertainty (Coefficient of Variation, CV < 0.1) of important watershed response characteristics, such as average daily flow (i.e., averaged over the entire simulation period), average monthly median daily flow, and average monthly maximum daily flow. Low flows (i.e., lower than 5th percentile) showed much larger uncertainty than high flows (greater than 95th percentile) as measured by their respective CVs, whereas intermediate flow rates have CVs almost less than 0.1. The rigorous approach for estimating predictive uncertainty and the consistency of the forecasts point toward a reliable application of the calibrated model for informed watershed planning and management.

Record Details: