Predictability in the Extended Range

Abstract

A two-level spherical quasi-geostrophic model is formulated for predictability experiments. The stationary external forcing for this model is calculated from observations. Both barotropic and baroclinic forcings are required in order to achieve a realistic model climatology. Realistic transient behavior is also present in the model. The most notable difference is that the observed transient kinetic energy has more energy in the smallestscales. Predictability experiments have an initial rms doubling time of approximately two days. This growth rate along with an initial error of about l/2 the initial error of present operational models produces an rms error equal to the climatological rms error and a correlation of 0.5 on about day 12 of the forecast. At the largest scales, this limiting point is reached shortly thereafter. The error continues to grow at a decreasing rate until at about 30 days the forecast skill is extremely small and comparable to the skill of a persistence forecast. Various time averages at various lags were examined for skill in the extended range. Filters that weighted most strongly the initial forecast days.were shown to provide increased skill. At the furthest limits (60-day time averages), filters improve the skill of prediction by an amount comparable to that which a numerical forecast is an improvement over a persistence forecast. A window filter improves forecasts of time averages by simply eliminating forecast days beyond about day 15. Besides the overall limit, no stable geographical or spectralvariations in the cutoff time could be determined from the limited sample of forecasts described in this paper.