6 Forecasting with a numeric model

Is it possible to make a real forecast with our model? To answer this question it is mandatory to define what "forecast'' means.

In a forecast process three different concepts of time have to be considered: an adaptation time of the model to the orography $t_{\rm A}$, a prevision time t_P and a time of the production of prevision $t_{\rm PP}$.

$t_{\rm A}$ is the time that the model needs to reach a thermodynamic and kinetic equilibrium already discussed in Sect. 5.2. Generally, it will depend on the geographical characteristics and from the initialization method. If the model is initialized with an unperturbed flow (this is the case for our simulations), the adaptation time is longer than that obtained with a flow in a quasi-equilibrium state. This last initialization mode consists on taking, over the domain surface, the (P, T, $\vec{V}$) fields adapted to the orography at a synoptic scales, that is at scales greater than the horizontal resolution used. The prevision time $t_{\rm P}$ is the time at which the model forecasts the atmospherical parameters. The $t_{\rm PP}$ time is a "lost technical time'' for a customer. It represents the time necessary to provide a forecast depending on many different elements such as the computer memory capacity, the computation speed, the gestion of the computer memory. We think that $t_{\rm PP}$ will still decrease in our case in the near future, because the replacement of the conventional vector-processing computer Cray with a parallel processor that will allow simultaneous calculations for many grid points at an higher speed. The value of a forecast degrades rapidly as $t_{\rm PP}$ approaches $t_{\rm P}$. Clearly the $t_{\rm PP}/t_{\rm P}$ ratio must be less than one and $t_{\rm PP}$ must be optimized for each kind of forecast (24, 18, 12 and 6 hours). Hence $t_{\rm PP}$ does not constitute a limitation parameter even for the shortest forecasting (6 hours).

In this paper we limit our analysis to the $t_{\rm A}$ time. We showed in the previous paragraph (Fig. 7 and Fig. 9) that $t_{\rm A}$, using such a geographic domain and initialization method, is as long as at least 30 minutes. Moreover, one can observe that in these simulations, the important turbulent layer near the ground is generated by the model only after 3 - 4 hours. We can hope to decrease this time by optimizing the initialization procedure. Knowing (Racine 1996; Munoz-Tuñón et al. 1997; Sarazin 1997) that the seeing characteristic time (decorrelation time) for good sites is of the order of 20-30 minute, we could take a minimum value of $t_{\rm A}$ of the same order of magnitude.

The results that we show in this paper check the model ability to reproduce a 3D atmospheric turbulence distribution but we cannot give a precise temporal location to our simulations, that is we did not yet provide a real "forecasting''. Only using the forecasting products of a meteorologic center such as the ECMWF we can estimate the real forecasting performances of this technique and give an estimation of $t_{\rm P}$. In the ECMWF/WCRP Level III-A Global Atmospheric Data Archive for example, the data are classed in different sets; we should be able to provide 6-hour forecast values from the ECMWF/TOGA Supplementary Fields Data Set and 24-hour forecast values from the Extension Data Set.