@article {Cocca:2003:1938-6478:762,
author = "Cocca, Paul A. and Doherty, John",
title = "Reducing Predictive Uncertainty in HSPF Watershed Model Applications: Impact of Calibration Data and Model Input Errors",
journal = "Proceedings of the Water Environment Federation",
volume = "2003",
number = "4",
year = "2003,
publication date ="2003-01-01T00:00:00",
abstract = "WinHSPF is the interface to the HSPF lumped parameter watershed model in the EPA BASINS modeling system. WinHSPF includes a driver for the Parameter Estimation (PEST) program. PEST calibrates models by minimizing an objective function quantifying the misfit between model
outputs and corresponding field measurements. Where flow gage, precipitation, and other model data are error free, the model structure perfectly represents actual watersheds, and model parameters are not redundant, minimizing an objective function should result in a single unique input parameter
set and highly certain model predictions. In the real world, errors in model input, calibration data, and model structure result in equivalent measures of goodness of fit for wholly distinct input parameter sets. PEST, run in predictive mode, identifies those input parameter sets that maximize
and minimize a given prediction while satisfying the condition that the model satisfactorily fits historical data. This range of uncertainty in model prediction results from parameter non-uniqueness, and from data and model errors.The impact of model input and calibration data errors on
predictive uncertainty is examined. First, the ability of PEST to identify correct input parameter sets under ideal conditions is tested. PEST is used to calibrate HSPF against a pseudo observed time series, i.e. output from a pre-calibrated HSPF model. PEST is then run in predictive
mode to identify minimum and maximum predictions of surface runoff rate. Artificial measurement errors are then imposed on the flow gage and precipitation time series, and other key model inputs. For each alteration of the model input or calibration time series, PEST is run in predictive mode
to identify the effect on predictive uncertainty. The effects from the following error types are examined: precipitation time series errors including altered storm volumes, overall percent change in precipitation, and a random error function; streamflow gage errors including systematic bias
and a random error function; non-representative precipitation and other meteorological gage data (switched with another nearby gage); incorrect fixed input parameters such as stage/discharge relationship and overland flow plane variables; and incorrect parameter range limits.Methods
of insulating model applications from these potential errors are examined. Specifically, the protective effect of different objective function component types against certain error types is demonstrated. Objective function components included in the study are simulated versus measured daily
streamflow, cumulative seasonal flow volumes, flow exceedence curves, and other corrective measures. Specific objective function components, found effective in protecting against common errors in model input and/or calibration time series, are identified.",
pages = "762-780",
itemtype = "ARTICLE",
parent_itemid = "infobike://wef/wefproc",
issn = "1938-6478",
publishercode ="wef",
url = "http://www.ingentaconnect.com/content/wef/wefproc/2003/00002003/00000004/art00043",
doi = "doi:10.2175/193864703784828903"
}