Verification of River Stage Forecasts

Abstract

Little verification of hydrologic forecasts has been conducted to date, and therefore little is known about the skill of hydrologic forecasts. This dissertation presents a verification study of river stage forecasts with lead-times up to three days for sixteen locations in the United States for a period spanning the past decade. The verification metrics from this limited sample indicate that the below flood stage forecasts are skillful, and so are the day 1 above flood stage forecasts. However, by day 3, the longer lead-time, above flood stage forecasts appear to have little skill (when compared with simple persistence). Further, they have not improved during the period of record despite a number of forecast process improvements. A path to improving the forecasts is suggested, via a new approach to selecting enhancements to the hydrologic forecast process. In support of this method, two fundamental building blocks of a robust verification program are presented: a method to pinpoint sources of skill in forecasts, and a standardized process for verifying forecasts.One element of a complete verification system is a process to determine why forecasts behave as they do. Forecasters need to be able to determine what causes a forecast to be good and what causes it to be bad. Therefore, an operationally implementable method for conducting this type of verification analysis is described and demonstrated. The method is used to evaluate the influence of model calibration, model initial conditions, and precipitation forecasts on the skill of single-valued (deterministic) river forecasts.A second important element of any forecast process, is a well defined, standard verification methodology. This dissertation proposes a standard verification system for deterministic river forecasts as a foundation for future discussions and for development of a well accepted set of verification practices for hydrologic forecasts. The proposed standards account for the needs of users, forecasters, scientists and administrators and are designed to be easily implemented within the constraints of an operational system.