Solar forecast

Use Case: Site specific solar forecast assessment with the help of cloud physical parameters

Within the preparations for the GMES/Copernicus atmospheric service (MACC-2 project) historical datasets of cloud physical parameters as derived from the Meteosat Second Generation satellite are provided. The ENDORSE project makes use of them for the assessment of solar energy sites with respect to the quality of irradiance forecasts at these locations. This includes

a)   the assessment of solar irradiance forecast capabilities as provided by a standard numerical weather prediction . done here with respect to concentrating solar technologies and the day ahead forecasts needed for the electricity grid integration of solar power

b)   the use of cloud fields in a nowcasting scheme enhancing the forecast capabilities further

  
  
  
  
  
  
  
  
  
  
  
  
  

Fig. 1: Plataforma Solar de Almeria test site

This web site is dedicated to report on these use cases of the cloud physical parameters derived from satellites. All results are given for the year 2006 and the location of the Plataforma Solar de Almeria (PSA), a solar energy test facility operated by CIEMAT in southeast Spain. Pyrheliometer measurements (OBS) are made at DLR/PSA facilities by our colleagues from the DLR Institute for Solar Research.

Global irradiance forecasts as provided by the ECMWF (European Centre for Medium-Range Weather Forecasts) have been obtained and transferred into direct normal irradiance (DNI) forecasts in an hourly resolution for the 48 hours forecast length (FC). Statistical evaluations are provided for all daytime hours available and passing the quality standards as described in Beyer et al. (2008). Statistics are generally given for the 2nd forecast day as this is the relevant period for the day ahead electricity markets.

First of all, a histogram of all values in FC vs. OBS (Fig. 2) shows good agreement at DNI values above 800 W/m2, but also overestimation in the 500 to 800 W/m2 and underestimation in the 0 to 150 W/m2 range. This is complemented by a histogram of all differences FC-OBS (Fig 3) having a non-symetric shape with a higher number of large positive differences.

Fig. 2: Number distribution forecast (FC) and observations (OBS)

 

Fig. 3: Number distribution FC - OBS

 

Fig. 4: Two-dimensional histogram FC vs.OBS

A scatter plot with the frequency of occurrence as color bar (Two-dimensional histogram) provides information about the FC vs. OBS distribution (Fig. 4). A broad scatter can be seen-frequently, the FC is overestimating OBS values in the 0 to 100 W/m2 range. On the other hand, OBS values around 400 to 600 W/m2 are underestimated by the FC and also in the maximum value range above 800 W/m2 the FC tends to underestimate. The following figures 5 to 8 provide statistical parameters as a function of forecast lag up to 48 hours.

 

Fig. 5: Skill scores based on the RMSE and 2d-persistence

Fig. 6: Bias over forecast lag

 

Fig. 7: RMSE FC vs. OBS

 

Fig. 8: Pearson correlation coefficients FC vs. OBS

 

Skill scores are defined as the difference in RMSE of the forecast and a persistence relative to the persistence-based RMSE. Please note that the persistence is defined as a 2-day persistence following the rules of the day-ahead electricity market. Day ahead forecasts have to be delivered in the late morning hours, therefore only yesterday's OBS can be used as persistence FC for the day-ahead market. Also, the perfect RMSE is set to zero. The other statistics (bias, standard deviation, root mean square deviation, relative values) follow the MESOR guidelines (Beyer et al., 2008).

Interesting for direct normal irradiances is the fact that typically statistical parameters of the day ahead (forecast hours 25 to 48) show comparable values to those of the first forecast day. This is different from most other meteorological parameters, where typically a decrease in accuracy is observed over the forecast lag. The FC in this case shows a good RMSE performance compared to the 2 day persistence, but also a remarkable bias structure during the day with underestimation in low sun conditions and overestimation in high sun conditions.

Fig. 9: Statistics of occurrence of cloud types (low level, medium level, high level mixed phase/water clouds and high thin ice clouds) during the time of the day

 

Based on a 2008 to 2012 time series of cloud physical parameters as derived from MSG APOLLO, cloud statistics at the PSA location are derived. Fig. 9 shows the importance of thin ice clouds at this location-overall and especially with an emphasis on afternoon hours. Thin ice clouds or cirrus-like clouds still allow a non-zero direct irradiance being observed at the ground despite of the cloud's extinction. Therefore, it is meaningful to distinguish in FC assessments into thin ice clouds, other clouds and cloud-free cases (being affected by aerosols mainly) in order to assess FC improvement potentials. Based on the different cloud regimes a clear difference in biases of the FC can be observed between the thin ice and the other cloud types, while clear sky cases are slightly underestimated by the FC (Fig. 10). Also the CRMSE (being the standard deviation of differences FC-OBS) shows an increase for thin ice clouds if compared to other clouds and the cloud free cases.

Fig. 10: Target plot for FC vs OBS, but separated in thin ice clouds (+), cloud free (diamond) and other clouds (*) at the PSA location. All units are W/m.

 

In the following plots we show example results for a 10 UTC satellite-based nowcasting run. Based on the shown difference in the cloud type regimes at the PSA and the general knowledge of different cloud movements in different vertical regions, the nowcasting algorithm treats thin ice clouds separately from the other clouds. For 10 UTC at the day ahead day a nowcasting of DNI is performed based on the 9:30, 9:45, and 10:00 UTC MSG images using a cloud tracking procedure. DNI is forecasted for each hour from 10:00 to 18:00 UTC. Skill scores and the RMSE are improved over up to 4-6 hours, while especially biases are reduced up to 8 hours (Fig. 11 and 12).

Fig. 11: Impact of nowcasting at 1000 UTC (solid) on biases of the 2nd forecast day

 

Fig. 12: Impact of nowcasting at 1000 UTC (solid) on RMSE of the 2nd forecast day

 

Overall, the use case illustrates that cloud physical parameters allow a deeper insight into direct irradiance forecast assessments and open the floor for a more specific improvement of forecasts.

References

Beyer H. G., Martinez J. P., Suri M., Torres J. L., Lorenz E., Hoyer-Klick C., Ineichen P., 2008. D 1.1.1 Handbook on Benchmarking, Management and Exploitation of Solar Resource Knowledge, CA . Contract No. 038665

Acknowledgements

This work has been supported by European Commission.s ENDORSE project (GA 262892) under the Seventh Framework Programme Theme .Stimulating the development of downstream GMES services'.