The breeding method has been used to generate perturbations for ensemb
le forecasting at the National Centers for Environmental Prediction (f
ormerly known as the National Meteorological Center) since December 19
92. At that time a single breeding cycle with a pair of bred forecasts
was implemented. In March 1994, the ensemble was expanded to seven in
dependent breeding cycles on the Gray C90 supercomputer, and the forec
asts were extended to 16 days. This provides 17 independent global for
ecasts valid for two weeks every day. For efficient ensemble forecasti
ng, the initial perturbations to the control analysis should adequatel
y sample the space of possible analysis errors. It is shown that the a
nalysis cycle is like a breeding cycle: it acts as a nonlinear perturb
ation model upon the evolution of the real atmosphere. The perturbatio
n (i.e., the analysis error), carried forward in the first-guess forec
asts, is ''scaled down'' at regular intervals by the use of observatio
ns. Because of this, growing errors associated with the evolving state
of the atmosphere develop within the analysis cycle and dominate subs
equent forecast error growth. The breeding method simulates the develo
pment of growing errors in the analysis cycle. A difference field betw
een two nonlinear forecasts is carried forward (and scaled down at reg
ular intervals) upon the evolving atmospheric analysis fields. By cons
truction, the bred vectors are superpositions of the leading local (ti
me-dependent) Lyapunov vectors (LLVs) of the atmosphere. An important
property is that all random perturbations assume the structure of the
leading LLVs after a transient period, which for large-scale atmospher
ic processes is about 3 days. When several independent breeding cycles
are performed, the phases and amplitudes of individual (and regional)
leading LLVs are random, which ensures quasi-orthogonality among the
global bred vectors from independent breeding cycles. Experimental run
s with a 10-member ensemble (five independent breeding cycles) show th
at the ensemble mean is superior to an optimally smoothed control and
to randomly generated ensemble forecasts, and compares favorably with
the medium-range double horizontal resolution control. Moreover, a pot
entially useful relationship between ensemble spread and forecast erro
r is also found both in the spatial and time domain. The improvement i
n skill of 0.04-0.11 in pattern anomaly correlation for forecasts at a
nd beyond 7 days, together with the potential for estimation of the sk
ill, indicate that this system is a useful operational forecast tool.
The two methods used so far to produce operational ensemble forecasts-
that is, breeding and the adjoint (or ''optimal perturbations'') techn
ique applied at the European Centre for Medium-Range Weather Forecasts
-have several significant differences, but they both attempt to estima
te the subspace of fast growing perturbations. The bred vectors provid
e estimates of fastest sustainable growth and thus represent probable
growing analysis errors. The optimal perturbations, on the other hand,
estimate vectors with fastest transient growth in the future. A pract
ical difference between the two methods for ensemble forecasting is th
at breeding is simpler and less expensive than the adjoint technique.