Intra-hour forecasting with a total sky imager at the UC San Diego solar energy testbed

TL;DR: In this paper, a method for intra-hour, sub-kilometer cloud forecasting and irradiance nowcasting using a ground-based sky imager at the University of California, San Diego is presented.

About: This article is published in Solar Energy.The article was published on 2011-11-01 and is currently open access. It has received 544 citations till now. The article focuses on the topics: Cloud cover.

Summary

1. Introduction 1

• Integration of large amounts of photovoltaics (PV) into the electricity grid poses technical challenges 2 due to the variable nature of the solar resource.
• The spatial 8 resoluton of NWP is coarse at about 100 km2, but there is active research on high-resolution rapid 9 refresh models with grid cell areas of less than 10km2 (Benjamin et al. 2010, Lara-Fanego et al. 2011).
• In Section 2, 23 the authors describe their experimental setup, present methods to obtain sky cover, cloud motion, irradiance, and 24 to forecast cloud locations and irradiance.

2.1 Experimental Setup 5

• Typical components of a sky imager are a charge coupled device 4 (CCD) camera, a fisheye lens, an environmental housing, and possibly a solar occultor depending on 5 the choice of CCD sensor and application.
• Since clouds scatter the visible wavelengths more evenly, 12 the red signal is similar to the blue signal.
• By adding the clouds detected using the 8 CSL (Fig. 3e) to the clouds detected using SP (Fig. 3d), the overall cloud decision image is obtained 9 (Fig. 3f).

2.3 Irradiance Estimation 10

• Prior techniques for estimating global horizontal irradiance (GHI) over an extended area consisted 11 of satellite-derived cloud cover coupled with an empirical clear sky model (Cano et al.
• On the other hand, ground based instruments are typically point sensors and do not provide 14 spatial irradiance information.
• This value is chosen because the clouds that 11 occurred on the days selected for validation were optically thick and reduced the irradiance to 12 approximately this level.

2.4 Cloud Motion 16

• Cloud velocity and direction of motion is determined through the cross-correlation method (CCM) 17 applied to two consecutive sky images (Hamill & Nehrkorn 1993).
• The CCM is performed on the 9 red channel image which has a higher contrast between clear sky and cloud than the blue and 10 green channels (the cloud decision image would have even higher contrast because it is a binary 11 image but no texture information).
• Future work will use the RBR image because this has higher 12 contrast than the red image.
• Finally, vectors that are more than one standard deviation from the mean in either the or 5 component of the velocity are eliminated.
• The authors note that this method needs further development for the case 7 of several cloud layers with different velocity vectors.

2.5 Cloud Forecasting 9

• To forecast cloud cover, the cloud map at time is advected at the speed and direction of the 10 global vector determined from cross-correlating the images at time and ( 30 seconds).
• 11 To determine accuracy, the actual cloud map at time (Fig. 5b) is overlaid onto the advected 12 cloud map (Fig. 5c) to determine the pixel-by-pixel forecast error (Fig. 5d).
• Since the circumsolar region 13 has a large potential for erroneous cloud decision, the sky image is divided into the circumsolar region 14 (within 35° of the solar azimuth angle) and the outer region (Fig. 3a).
• 18 , 100% (6) describes the forecast error obtained by cloud advection (Fig. 5d) divided by the error obtained if the 19 image at was assumed to persist until (no advection), also known as The cloud-advection-versus-persistence (cap) error.
• An 1 implies that the cloud 20 advection improves the forecast compared to persistence.

3.1 Selection of Days 2

• To assess the accuracy of sky imager forecasts, the authors selected relevant scenarios with typical sky 3 conditions and high image quality.
• In their binary (cloud / no cloud) system, entirely clear or overcast 4 days will always result in perfect forecast and these days are eliminated.
• Furthermore days with 5 multiple cloud layers moving in different directions, rapid cloud deformation, formation, or evaporation 6 were removed.
• Lastly the heating system on the imager failed for part of the year leading to excessive 7 dirt aggregation on the mirror because of dew that made cloud detection impossible.

3.2 Cloud Decision 2

• Since cloud decision images can only be validated visually, the authors describe here qualitative 3 experiences of applying their method.
• 5 the standard deviation of the RBR in the circumsolar region is found to be largest when the sun is near 6 the horizon (low sun-pixel-angle and high zenith angle).
• Anecdotal observations confirm that clear 7 conditions in the circumsolar region at low sun elevation are often misclassified as cloudy while the 8 horizon and circumsolar region appear white.
• If the RBR threshold is lowered to detect thin clouds it will increase false cloudy 15 (cloud detected but no cloud exists) detections, especially in the circumsolar region.
• When the 23 sun is obscured, dark clouds are correctly classified due to the SP algorithm (Fig. 7).

3.3 Cloud Motion 8

• Figure 8 illustrates the computed motion vector field at several stages in the quality control process.
• A reduction in the CCC usually implies smaller accuracy of motion vectors.
• 13 After removing vectors in the circumsolar and clear sky region, and vectors with CCCs less than 0.8, 14 most remaining vectors are uniform (Fig. 8b).
• If the sky contains clouds with sharp cloud boundaries, CCM generally 4 performs well.
• Due to the general persistence of cloud velocities (usually within 2 m s-1 over several 5 minutes), erroneous decision vectors can be identified since they are associated with large cloud 6 velocity fluctuations (Fig. 9a).

3.4 Nowcast of Binary Cloudy Conditions 1

• Using the binary irradiance technique described in section 2.3, GHI time series were produced for 2 the ground stations listed in Table 2 Location and status information for w with daily availability on the 3 four days listed in section 3.1.
• The 1 Hz GHI data collection is faster than the image capture frequency, so validation GHI time series 8 were constructed for each station by using only the data points gathered at the time when the sky 9 image was taken (no averaging is performed).
• Once the projected sky position moved into the outer 3 region the TSI nowcasts the sky conditions between the sun and MOCC accurately, correctly predicting 4 the sky condition 68.3% in clear conditions and 80.4% in cloudy conditions, where clear is defined as a 5 clear sky index ( ) greater than 0.7.
• CLRm and CLDm are the number of clear and cloudy measurements, respectively, in the outer region.
• The negative bias in the circumsolar region is primarily due to false clouds decreasing the estimated 3 GHI.

3.5 Minutes Ahead Forecast 9

• The performance of a minutes ahead forecast is discussed in this section for four days following the 10 metrics established in section 2.5 and Fig.
• The matching error is larger than on other days, but the large cloud speed causes large 6 persistence errors which reduces the cap error.
• Note that while the outer 18 region matching error was considered for 30 second forecasts, here the total (circumsolar and outer 19 region) errors are considered as clouds are frequently advected between the circumsolar region and 20 the outer region (depending upon cloud speed and direction).
• The perspective error may also create the illusion that there is one cloud while 18 there are two nearby but separate cloud.

Figures

Figure 7: Cloud decision comparison between clear sky library (CSL) method (b) and integrated CSL and sunshine parameter method (c) on October 4, 2009 12:43:30 PST.

Figure 12: Time series of (a) westward and southward cloud speed, (b) cloud fraction, (c) matching error, and (d) cloud-advection-versus-persistence (cap) error obtained every 30 seconds on September 14, 2009. The red line in c) and d) indicates the error in the circumsolar region, while the blue line indicates the error in the outer region. In clear conditions (e.g. 1145 – 1350 PST) no cloud speed is determined and the errors are zero.

Figure 3: Processing chain of a sky image on October 4, 2009 15:45:30 PST (a) to obtain the cloud decision image. The sunshine parameter is 0.85 and is evaluated around the sun position indicated by the blue cross. The dotted black lines show the borders of the circumsolar region defined as solar azimuth ±35° and the solid black line shows ZA at 65°. (b) Clear sky RBR (colorbar) background image plus the threshold. (c) RBR (colorbar) image. (d) Pixels in (c) with RBR > SP or (e) RBR > CSL are assumed to be cloudy. (f) Shows the final cloud decision image. Green areas are clouds and blue areas are clear skies

Table 4 Mean and standard deviation of matching and cap error (Eq. 5 and 6) for the outer region on four different days. The errors are averaged only during conditions that were not clear or overcast. Since errors during overcast and clear conditions are zero, the errors in the table are biased high.

Figure 4: Illustration of the cross-correlation method using two images on October 4, 2009 at 16:18:30 (a) and 16:19:00 PST (b). Each 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 for each box. The cloud velocity determined from this pair of images was 5.2 m s-1. While the method is illustrated here on a full color sky image, the method is actually applied to the coordinate transformed (Eq. 1) red channel image.

Figure 10: Time series (a) of MOCC station for October 4, 2009 showing measured GHI (blue) versus the TSI nowcast. The TSI shadow map is separated into circumsolar (red) and outer-region (green). Nowcasts of the TSI after 1100 PST (b),(c) and (d) match the trends in the measured data well.

Table 5 Mean total matching error and total cap error for a 30 second to 5 minute ahead forecast. The errors are averaged only during conditions that were not clear or overcast and when the entire scene had not been advected out of the field-of-view. Since errors during overcast and clear conditions are zero, the errors in the table are biased high.

Figure 6: (a) Average RBR (colorbar) as a function of zenith angle and sun-pixel-angle based on TSI imagery on a clear day (September 24, 2009). This image is used to generate the clear sky library in Fig. 3b. (b) Standard deviation (colorbar) of clear sky RBR and (c) Number of occurrence (colorbar) for each zenith angle and sun-pixel-angle pair.

Figure 11: Time series (a) of HUBB station for March 10, 2010 showing measured GHI (blue) versus TSI nowcast, with the latter separated into circumsolar (red) and outer-region (green). Zooming in (b), (c) and (d) shows the ability of the TSI to produce nowcasts for a distant site.

Table 2 Location and status information for weather stations used to validate TSI cloud shadow maps

Figure 9: Time series plot of eastward and northward cloud velocities (a) and cloud fraction (b) on March 10, 2010. The velocities are persistent under partial cloud cover (0800-1030 PST and 1530-1730 PST), but fluctuate largely in overcast and almost clear sky conditions.

Table 1 Sky condition occurrence frequency from a ceilometer at Miramar Naval Air Station (KNKX) in 2010. Partly cloudy conditions are defined as few, scattered, or broken clouds.

Figure 5: The forecast cloud map (b) is produced by advecting the cloud map at time (a) in the direction of the motion vector (here XX degrees). To determine the forecast error (d), the future cloud map at time 30 seconds (c) is compared to the forecast (b). Blue and red colors in (d) show forecast errors (blue: pixel forecast cloudy and but actually clear; red: pixel forecast clear but actually cloudy) and white indicates accurate forecasts.

Figure 13: Total matching error for different forecast horizons on October 4, 2009. (a) 1 minute, (b) 2 minutes, (c) 3 minutes, (d) 4 minutes, (e) 5 minutes. Missing data occur because of missing images, clear or overcast skies, and when the entire scene had been advected out of the field-of-view.

Figure 8: Illustration of procedures for obtaining a motion vector field on a sky image in sky coordinates on March 4, 10:21:30 PST. (a) Original vector field resulting from the cross-correlation method, (b) after removing vectors near the sun and vectors with a correlation of less than 0.8, and (c) after removing outliers. The average cloud velocity was 4.0 m s-1.

Frequently asked questions

Q1. What are the contributions in this paper?

13 A method for intra-hour, sub-kilometer cloud forecasting and irradiance nowcasting using a ground14 based sky imager at the University of California, San Diego is presented. 

Q2. What are the future works in this paper?

16 Future work will include several elements to improve forecast accuracy. Sky imagers are also currently being installed at inland sites to study 3 cloud conditions likely to be experience by solar power plants. A CCD camera with large well-depth will be deployed to minimize 19 saturation in direct sunlight. A fisheye lens with an equisolid angle projection will improve the image 20 resolution at low elevation angles to increase forecast horizon and accuracy at long forecast horizons. 

Q3. What is the scenario to test the potential of the method?

Single layer 8 broken clouds without rapid deformation moving across the sky are thought to be the best scenario to 9test the potential of the method. 

Q4. How does the cloud deformation affect the forecast?

The rapid condensation and evaporation and associated deformation of clouds introduce 12 complexity to deterministic cloud forecasting. 

Q5. What causes the inaccuracies in the forecast sky?

Inaccuracies in the forecast sky conditions are also due to cloud deformation, evaporation, and 8 condensation, as well as uncertainty in cloud base height. 

Q6. What is the common method of forecasting?

physically based forecasting is primarily conducted using numerical 5 weather prediction (NWP) and satellite cloud observations. 

Q7. What is the forecast horizon for the area surrounding the imager?

The footprint of the cloud shadows on the 5 ground varies with sun angle and cloud height and the forecast horizon is a function of cloud height and 6 cloud speed. 

Q8. How are the forecasts for the weather in the NWP?

13 To achieve high temporal and spatial resolution for intra-hour forecasts, NWP and Satellite 14 forecasts are currently inadequate. 

Q9. What is the pixel coordinate of the projected sky image?

The projected sky image is 1 partitioned into subsets of pixels of equal size such that each subset is about 1% of the sky image 2 area. 

Q10. What is the role of the UCSD microgrid?

The UCSD microgrid provides a globally unique testbed 20 for a customer side of the meter smartgrid with renewable generation, thermal, and electricity energy 21 storage, demand side management, and demand response. 

Q11. What is the effect of cloudy clouds on the TSI?

These effects cause clouds to be dark in color at the 11 cloud-base and results in a smaller RBR than the CSL leading to incorrect clear classification. 

Q12. How many sky imagers will be installed at the UC San Diego energy testbed?

Additional 26sky imagers will be installed at the UC San Diego energy testbed to increase the coverage area and 1 forecast horizon. 

Q13. How many vectors are uniform in the cloud?

13 After removing vectors in the circumsolar and clear sky region, and vectors with CCCs less than 0.8, 14 most remaining vectors are uniform (Fig. 8b). 

Q14. What is the turbidity factor used in the Ineichen model?

This model requires the Linke turbidity factor (Linke 1922) as input and has a reported mean bias error 9of -6 W m-2, and a root mean square error of 19 W m-2 (Ineichen 2006). 

Q15. What is the way to assess the solar resource?

The use 8 of sky imagery to assess the solar resource for solar energy applications shows much potential for 9augmenting the spatial and temporal resolution provided by satellite and numerical forecasting methods. 

Q16. What is the pixel coordinate of the intersection of the 1 solar vector with the cloud map?

The pixel coordinate of the intersection of the 1 solar vector with the cloud map for the , element of the ground map is: 23 Actual spatial coverage of the GHI estimates within the 100 km2 region considered varies with the sun’s 4 position, cloud height and topography, e.g. when the clouds are low, the horizontal component of the 5 distance to the cloud is smaller and thus the coverage is smaller. 

Q17. How many measurements per km2 are there?

17 The advanced smart 42 MWP microgrid of the University of California, San Diego (UCSD) is one of 18 the world’s most densely monitored environments with over 18,000 measurement points per km2, 19 including a hemispherical sky imager (Fig. 1). 

Q18. how many days did the tsi estimate the condition of the sky?

For the four days chosen and for all available stations (Table 2 Location and status 9information for w), the TSI correctly estimated the condition of the sky 69.7% of the time in the outer 10 region.