Authors: Alex Kubiniec, Phil Gruenhagen, Adam Kankiewicz
Affiliation: Clean Power Research
Motivation: Clean Power Research currently produces behind-the-meter (BTM) PV fleet power forecasts every 30 minutes for five load regions in the territories of California’s three investor-owned utilities (IOUs), as identified by the California ISO. These forecasts are produced using satellite and numerical weather prediction (NWP)-derived irradiance values from SolarAnywhere® at 1 km x 1 km spatial resolution. System specifications such as latitude, longitude, tilt and azimuth, PV module and inverter efficiency ratings, obtained from IOUs, the California Energy Commission (CEC) and the California Solar Initiative (CSI), are used to model power output from 200,000 + systems. The power output from these individual systems is then aggregated to provide regional fleet power output. In addition to the increased computing horsepower required to model such large numbers of systems, specifications for the systems in the publicly available data are inexact or missing altogether. For example, locations are anonymized by providing only the systems’ zip code and system orientation (tilt and azimuth) is only available for 30% of the systems. Rather than aggregating system specifications and creating capacity weighted averages or a “top down approach”, we currently scale the modelled PV fleet power output on the assumption that the locations and system orientations of new systems will have a distribution similar to that of the current fleet. In a way, this is one of the simplest methods for creating a bottom-up representative PV fleet. This work was done by Clean Power Research and funded by a CEC-EPIC award.
Methodology: To investigate methods for reducing the computational resources required for modelling large PV fleets and to better quantify the margin of error that might be introduced by generalizing the locations and orientation of systems in such fleets, we created various representative fleets using the CSI systems in the PG&E Non-Bay Area load region behind-the-meter fleet as a baseline for comparison. This baseline fleet, consisting of the 35,562 PV systems mapped in Figure 1, is spread across a large portion of California and has a wide variety of system orientations. The 30-minute power output from each of these systems was simulated for a one-year period from January 1, 2014 through December 31, 2014 and the results were aggregated to produce the baseline PV fleet output every 30 minutes during the period. “Bottom-up” representative fleets were created by binning the capacity in a baseline fleet with known system specifications. The relative Mean Absolute Error (rMAE) for each of the representative (test) fleets was calculated compared to the output of this baseline (reference) fleet.
Results: The amount of error introduced by using bottom-up representative PV fleets with regular geographic dispersion, rather than fleets consisting of individual systems with exact system specifications varied from 4.2% for the coarsest spatial resolution and smallest number of orientation bins, to 1.1% for the fleet with 10 km x km spatial resolution and the largest number of orientation bins. As shown in Figure 2, the greatest impact on error was due to spatial resolution, rather than the number of orientations considered. However, further increases in spatial resolution would likely have proportionally less impact as we approach the native resolution of the satellite data. The relative Mean Absolute Error for the representative fleets with a single location had significantly higher error than the other fleets, ranging from 10.1% to 10.7%. The representative fleet based on multiple orientations at each zip code fared reasonably well with rMAE ranging from 2.2% to 2.4%. However, the zip code based fleet that used a single orientation at each zip code had a much higher rMAE at 6.6%. While a multi-orientation zip code based fleet may be appropriate when exact system locations are unknown, performance is only slightly better than the representative fleets with the highest number of orientation bins and spatial resolution and error is approximately double.
Summary and Conclusion: Bottom-up representative PV fleets, created by generalizing the location and orientation of a set of individual systems with known specifications, can be used in modelling fleet output to reduce computing resource requirements by more than 80%, while introducing as little as 1.2% rMAE on an annual basis. Zip code based representative fleets, which make use of known individual system orientation data can reduce computing resource requirements by more than 65%, while introducing as little as 2.2% rMAE on an annual basis. Although they can be simulated very quickly, representative fleets that make use of a single location exhibit more than 10% rMAE and have a fairly inaccurate power production curve on a daily basis. At 6.6% rMAE, zip code fleets that use a single system orientation have less error than single location fleets, but typically exhibit a narrower daily production curve with a higher peak. Further details, including an analysis of cloudy sky BTM fleet error statistics will be reported on in this paper.
References:
Thomas E. Hoff, J. K. (2012). REPORTING OF IRRADIANCE MODEL RELATIVE ERRORS. Proc. ASES Annual Conference. Raleigh, NC: American Solar Energy Society.
![](images/flag.png)
![](images/blue_ribbon.png)