Meteorological Forecasting of DNI, clouds and aerosols

Transcription

1 Meteorological Forecasting of DNI, clouds and aerosols DNICast 1st End-User Workshop, Madrid, Heiner Körnich (SMHI), Jan Remund (Meteotest), Marion Schroedter-Homscheidt (DLR)

2 Overview What do we want to forecast? Local and short forecasts of DNI How to forecast? Different methods based on observations and numerical modeling Why is it difficult? Atmospheric phenomena have limited predictability. Combining different methods for the best forecast

3 What do we want to forecast? Direct normal irradiance Clouds Aerosols Local forecast for a solar power plant, about 1 square kilometre Forecast length: 0 to 240 minutes

4 How to forecast Ground-based observations by all-sky imagers: Calibrated images for cloud detection and classification Cloud velocities from two consecutive images Forecast with extrapolation Forecast length: ~ 30 minutes Spatial resolution: ~ 100 m Domain: local, ~ 1 km

5 Forecast with all-sky imagers Cloud Motion vector Forecast time t + 30 min

6 How to forecast Satellite-based observations: Utilize satellite imagers Optimize with rapid scan and highresolution observations Cloud motion vectors from optical flow techniques Forecast with extrapolation Forecast length: ~ 5-60 minutes, possibly up to 240 minutes Spatial resolution: ~ 3 km Domain: regional

7 Satellite-based observations Cloud fields from Meteosat Second Generation High Resolution Visible, centered over Greece (source DLR-IPA)

8 How to forecast Numerical weather prediction: Estimating initial state using all available information Solving the forecast equations as realistically as possible Time-critical production limits computational size Forecast length: up to 240 minutes, and longer Spatial resolution: ~ 3-20 km Domain: regional to global

9 Numerical weather prediction WRF (Meteotest), Harmonie (SMHI), WRF-Ensemble (DLR)

10 Difficulty with models: Parameterisation of unresolved physical processes Aerosol chemistry Turbulent mixing

11 Forecasting of DNI with Numerical weather prediction models Direct normal irradiance (DNI) is normally not a direct output of NWP models. DNI can be calculated with an external radiation code on basis of NWP input. NWP input: Clouds, water vapour (, precipitation) Winds, water vapour and temperature for aerosol chemistry and transport Winds also for advection of observed clouds

12 How to forecast aerosols Collect available observational information, e.g. from the WMO Sand and Dust Storm Warning Assessment and Advisory System s (SDSWAS) activity

13 How to forecast aerosols Global data assimilation of aerosols in the project Monitoring Atmospheric Composition and Climate (MACC) MACC delivers for Europe / Northern Africa generally good estimations Easily (and freely) available AOD@550nm - MBE Positive Bias Negative bias MBE of MACC AOD 550 nm Source: Mines Paristech / IEA SHC 46 MACC homepage

14 How to forecast aerosols High resolution aerosol modelling with COSMO- MUSCAT (TROPOS) Source: TROPOS

15 Targeted spatial and temporal scales Target Sky imagers Satellite Limited area NWP Global NWP 100 km Spatial resolution (log scale) 10 km 1 km 100 m 10 m 0.1 min 1 min 5 min 45 min 240 min 1 day Forecast time (log scale)

16 Why is forecasting difficult? Theoretical growth of forecast error

17 Error growth for different length scales From Lorenz-96 model: Error growth for two different length scales. The smaller length scale meso-gamma saturates faster.

18 Is higher resolution useful? Small-scale features in clouds have only a short predictability, if initialized correctly. Use of neighborhood method Take forecast over certain surrounding area, e.g. 20km x 20km and time interval. Calculate statistical measure for area: min, max, mean, median Result: Probability forecast for area Useful information for energy production?

19 Combining methods for the best forecast 3 Forecast error Forecast length

20 Combining methods for the best forecast Ground based methods Satellite based methods NWP based methods Images: SMHI, DLR Machine learning techniques (soft computing) combine the different forecasts in an optimal way. For example: FIS (Fuzzy inference systems), SVM (Support vector machine) KNN (K nearest neirbourgs) NN (Neural networks) ESTIMATIONS MEASUREMENTS

21 Meteotest: Shortest term forecast model satellite data cloud index cloud mask IR nighttime update: 15 min Numerical weather model WRF wind vectors update: 2x per day Meteotest WRF / different sources Aerosol data Meteotest / diff. sources Calculation of cloud index trajectories Prediction cloud position steps 15 min MACC / diff. so. ~16 min Post processing to reduce uncertainty Clearsky model prediction global & direct irradiance

22 Satellite & NWP based wind vectors Comparison of cloud motion vectors (CMV) based on satellite or NWP: quality is comparable but NWP based vectors much faster Comparison of Univ. Oldenburg (IEA SHC 46) Sat: satellite estimation NWP: NWP forecast CMV org: satellite based CMV 3500 m: NWP wind vectors at 3500 m

23 Cloud index & wind trajectories Source: Meteotest

24 Post processing Post processing: Kalman filter depending on cloud situation (only used for clear or totally cloudy situation but not for mixed) (Source: Meteotest)

25 Results: Forecast error for Global horizontal irradiance Opt. Modell (ECMWF/MOS)

26 Conclusions Methods for DNI forecasts include ground- and satellitebased observations, post-processing as well as numerical weather prediction. Different spatial and temporal scales are adressed by the methods. Predictability of atmospheric phenomena is scaledependent, extending from hours to weeks. Best forecast can be achieved by a combination of the different methods. Wind vectors from NWP and CMV are equally good. RMSE for 0-6 hours / all stations, GHI: W/m 2 (20-60%)

27 Thank you for your attention!