83rd Annual

Tuesday, 11 February 2003
Evaluation of skill and error characteristics for alternative seasonal streamflow forecast methods
Alan F. Hamlet, University of Washington, Seattle, WA; and D. P. Lettenmaier
Most water resources systems in the western U.S. whose hydrologic response is dominated by snowmelt make use of statistical streamflow forecasting systems that relate spring and summer runoff to snow observations. Recently, experimental long-lead streamflow forecasting systems that use climate teleconnection information, like ENSO, have been developed. In retrospective evaluations, the potential utility of these long-lead (several seasons to a year or more) forecasts for water resources management has been demonstrated, largely because they can provide information about future hydrologic conditions prior to the winter snow accumulation season. Aside from the obvious differences in useful lead-time in the two types of forecasts (which has value in itself in many applications), the error characteristics of the two kinds of forecasting systems are also different. Statistical forecasts, for example, generally do not have substantial skill until mid winter, but are robust to large forecast errors, because much of the precipitation available for runoff is stored in the snowpack at the time the forecast is made. To better understand the implications of these differences in skill and error characteristics, two sets of retrospective ensemble forecasts for the Columbia River at The Dalles are produced and evaluated. The first set of forecasts are simulated using an assumption of perfect advance knowledge of ENSO, but are made prior to accumulation of the winter snowpack, so in situ observations are used only to specify the soil moisture initial condition. The second set of forecasts represents a forecast that could be made on January 1 if perfect information were available about the snowpack accumulated to that time, but with no forecast of future climate. The relative skill in comparison with climatology (using various skill scores), and sources of error in these forecasts are compared quantitatively for a 50-year period. In the case of the long-lead forecasts, the influence of errors in ENSO forecasts is evaluated, as is the influence of snowpack estimating error for the snowpack-based forecasts.

Supplementary URL: