Short-term Solar Forecasting

Short-term Solar Forecasting Presented by Jan Kleissl, Dept of Mechanical and Aerospace Engineering, University of California, San Diego 2

Agenda Value of Solar Forecasting Total Sky Imagery for Cloud Tracking Solar Forecasting Results Integration with OSIsoft Products Copyright 2011 OSIsoft, LLC. 3

About UC San Diego s Microgrid 42 MW p costumer microgrid 12 million ft 2 of buildings Self generate 82% of annual demand 30 MW natural gas Cogen plant 2.8 MW of Fuel Cells contracted 1.2 MW of Solar PV installed, additional 2 MW planned Copyright 2011 OSIsoft, LLC. 4

Business Challenge/Problem Addressed NREL / GE Western Wind and Solar Integration Study: Use of state-of-art wind and solar forecasts reduces WECC operating costs by up to 14%, or $5 billion/yr, as compared to not using wind or solar forecasts for day ahead unit commitment ($12 20/MWh of wind and solar generation). WECC operating costs could be reduced by an additional $500 million/yr in the 30% case if wind and solar forecasts were perfect. Copyright 2011 OSIsoft, LLC. 5

Background Solar photovoltaic output nearly linear with irradiance Day-ahead solar forecasting using numerical weather prediction Hour-ahead: satellites Copyright 2011 OSIsoft, LLC. 6

Challenge/Problem Detail Most clouds are unresolved in numerical weather prediction or satellite models Time interval for satellite imagery is 15 to 30 minutes Copyright 2011 OSIsoft, LLC. 7

Solution: Total Sky Imager Copyright 2011 OSIsoft, LLC. 8

Automated Cloud Decision Algorithm: True color - Clear sky scatters shorter wavelengths (strong signal on blue channel, weak on red channel) - Clouds scatter the visible wavelengths more evenly - Ratio of red and blue channel provides information for cloud detection red red channel blue channel

sunshine parameter Final Cloud Detection clear sky library simple threshold red blue ratio final decision

8 km Cloud Mapping Cloud Map Hemispheric Sky Image cloud height Site Ground Map Figure 1 Cartoon showing how the shadow value at a ground point is computed. If the location in the cloud map where the sun vector intersects contains a cloud, then the ground location is considered shaded. The altitude of each pixel in the ground map is considered when computing the location of Sky intersection. Coordinates UCSD Sky Imager October 4, 2009 11:44:30 Cloud height ~ 1050 m

(a) Cross-correlation method Using 2 images taken 30 sec apart, a region of pixels from (a) is correlated to (b) within a search distance. The location of the highest correlation is found and a motion vector is defined. 2009-10-04 16:18:30.000 Loop over the entire image Cloud Tracking (b) Motion Vector Box with max correlation Search distance subset of pixels to correlate Original subset

Cloud Motion Vectors Apply cross-correlation method Retain only vectors for which high correlation is obtained Assume homogeneous cloud velocity crop image

GHI [W GHI m -2 ] Nowcasting results Movie Problems in the morning when shadowband covers sky over site Very good agreement in the afternoon 1100 1000 900 MOCC 04-Oct-2009 now-cast MOCC GHI 800 700 600 500 400 300 200 100 0 06:00 09:00 12:00 15:00 Time Time (1 day)

Global Horizontal Irradiance [W/m 2 ] 1200 1000 800 600 400 200 (c) Nowcast results MOCC 0 12:40 12:50 13:00 13:10 13:20 13:30 13:40 13:50 14:00 Time Table 5 Percentage co-occurrence of clear and cloudy conditions for measured/nowcast. CLR/CLR CLR/CLD CLD/CLR CLD/CLD September 14, 2009 56.1 20.6 8.1 15.2 October 4, 2009 55.2 9.9 3.2 31.7 March 4, 2010 59.2 18.3 7.8 14.6 March 10, 2010 54.2 12.8 4.2 28.9 Total 56.1 17.3 6.7 19.9

cap error [%] matching error [%] cloud fraction [-] cloud speed [ms -1 ] 30 sec Forecast actual t o forecast t o+30s actual t o+30s error a) b) c) d) 5 a) E-W N-S 0-5 1 0.75 0.5 0.25 15 7.5 b) c) 0 75 50 25 d) 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 time [HH:MM]

1 to 5 minute forecast Mean total matching error and total cap error for 30-sec to 5-min ahead forecast. Since errors during overcast and clear conditions are zero, the errors in the table are biased high. 30 sec 1 min 2 min 3 min 4 min 5 min Time until advection out of scene [min] Sept 14, 2009 45.0 51.5 63.9 70.0 76.5 123 4-27 Oct 4, 2009 47.2 49.8 55.6 61.5 66.6 70.3 8-18 Mar 4, 2010 54.6 55.3 59.3 63.4 67.7 71.8 9-24 Mar 10, 2010 48.8 53.9 62.3 68.8 75.1 78.0 9-15 Error increases with forecast horizon, but 25% better than persistence after 5 minutes. After 10 to 25 minutes the scene is advected out of the field of view.

Conclusions Solar variability analysis tools developed Quantify variability for different array sizes and geographic layout Quantify ramp rates Demonstrated sky imagery in testbed at UCSD

Future Plans and Next Steps Conduct solar forecasting at 48 MW Henderson, NV solar power plant Inverter-level power output and meteorological data acquired through PI Server Copyright 2011 OSIsoft, LLC. 19

48 MW Data from each inverter and 10 meteorological stations Statistical forecast methods (ARIMA) Sky Imager Forecast Ramp forecast for CAISO Nevada Lat: 35 o 46 52 Lon: 114 o 59 35 Optimum Tilt: 35o GI at fixed tilt Arizon a Copyright 2011 OSIsoft, LLC. 20

Copyright 2011 OSIsoft, LLC. 21

Questions Jan Kleissl, jkleissl@ucsd.edu, http://solar.ucsd.edu Copyright 2011 OSIsoft, LLC. 22

Thank you Copyright 2011 OSIsoft, LLC.