METHODOLOGYObservations and Evaluation Procedure
This study is conducted for the Andasol Solar Thermal Power Plant (37.228º N, 3.069º W; 1100m.a.s.l.), Fig. 1. Ten-minute DNI measurements arecollected with an RSR2 radiometer. This instrumentis well maintained and calibrated. Data for the 12-month period 01/12/2009 to 30/11/2010 were firstcorrected for spectral effects, and finally filteredwith a series of quality control tests. For this study,irradiance values corresponding to solar zenith an-gles above 85º were filtered out to avoid the highmeasurement uncertainties associated with low-sunconditions. The original 10-minute data were alsoaveraged to obtain hourly values. From a climato-logical standpoint, 2010 was an exceptionallyrainy—and therefore cloudy—year.Two forecast horizons are studied separatelyhere: hours 1–24 (day 1, or “day ahead”), and hours25–48 (day 2). Sky conditions are characterized bythe clearness index (k
) to separate clear-sky (0.65 <k
), cloudy (0.4
0.65) and complete overcast(k
< 0.4) conditions. The forecast reliability is ob- jectively evaluated in terms of mean bias error (MBE) and root mean squared error (RMSE), andtheir relative values (in %), obtained by normaliza-tion to the mean of the ground measurement irradi-ance for the considered period. The forecast errors(residuals) are calculated as the difference betweenforecasted values and observations. A positive MBEis thus indicative of an overestimation of the mod-eled DNI. Finally, the trivial persistence model isused as the skill reference model.
The model’s domain configuration is representedin Fig. 1. The dynamical downscaling is driven bythe use of four nested domains with progressivelydecreasing horizontal resolutions of 27, 9, 3, and 1km for the outermost to innermost domains. The at-mospheric column is decomposed into 28 verticallevels. The ECMWF/IFS weather forecasts are usedas initial and boundary conditions. For each day of the evaluation period a WRF (ARW, version 3) sim-ulation of 60 hours is run. The first 12 forecastedhours are considered as model spin up, and discard-ed. The next 48 hours are evaluated independentlyfor the first and second 24-hour periods. The WRF parameterizations are selected based on . In par-ticular, Dudhia’s scheme is used for the shortwaveradiation parameterization.
NWP models (e.g., WRF) do not usually provideDNI as an output variable. Therefore, DNI needs to be derived in a post-processing step based on WRF’scomprehensive forecasted information and an exter-nal radiative model [4, 7]. Recent studies have rig-orously analyzed the performance of a large set of different radiative models [8, 9]. Results showed thatmeteorological radiative models achieve a very high performance in DNI estimation under clear-sky con-ditions. In contrast, statistical/empirical modelsshow lower performance but more simplicity. Sincethis work focuses on the WRF aspects related tocloud modeling, the simplest way to derive DNI is preferred; nevertheless better results can be expected by using the first kind of radiative models mentionedabove . An empirical statistical model  issimply applied here to obtain DNI from the WRFglobal horizontal irradiance (GHI) forecasts.
Configuration of the WRF domains. The spa-tial resolutions are 27 km, 9 km, 3 km and 1 km for do-mains D01, D02, D03 and D04, respectively. The radio-metric station is located at the center of all domains.
RESULTSDependence On Horizontal Resolution
Figures 2 and 3 show the performance results for the day-ahead forecast horizon over the whole 12-month period. Extremely large errors are obvious for complete overcast conditions (Fig. 2). Since DNI isvery sensitive to the presence of clouds, any misrep-resentation of cloudiness—in either space or time— in the model’s predictions may cause significant er-rors. The high variability of cloud type and cloudamount enhances this effect. The most important re-sult is that the errors (RMSE and MBE) increasewith spatial resolution, contrarily to what wouldhave been expected. An interesting exception is thecase of clear-sky conditions, under which the RMSEdecreases (Fig. 3). These results mean that cloudsare not better resolved at higher spatial resolution bythe model, at least in terms of their effect on solar radiation. In contrast, under cloudless conditions thetopographic effects, which are better resolved at fin-er resolutions, become more relevant. This is of par-ticular interest to CPV, since this technology cannormally be installed on uneven terrain.