In the previous two posts in our “Chasing Waterfalls” blog series, we introduced Locus Energy’s Waterfall as a tool that we developed to model the causes of PV system underperformance to give a more complete understanding of what is occurring in large fleets of residential PV systems. We also discussed some of the trends that we were able to find in this data and what that showed us about typical real world performance with respect to snow, soiling, and shading. In this final post of the series, we turn our attention to the question of specifically addressing how this data is useful in a fleet management context.
Modeling specific causes of system underperformance allows us to achieve significantly higher levels of accuracy when modeling system performance, compared to previous approaches employed at the fleet level. The illustration below shows the contrasts between the distributions of monthly ratios of measured system performance normalized by energy expectations (EER), GHI-adjusted energy expectations (GEER), and modeled system performance inclusive of modeled system underperformance.
Locus uses the term Expected Energy Ratio (EER) to describe a non-weather-adjusted performance ratio commonly used to assess site performance. The EER performance metric is defined as follows:
EER = measured energy
expected energy based on long-term weather data
Locus uses the term GHI-adjusted Expected Energy Ratio (GEER) to describe a type of weather-adjusted performance ratio where GHI (Global Horizontal Irradiance) is used for the weather adjustment factor. This ratio is an improvement over EER because it accounts for how much sunlight energy the PV site had as input. In the context of this blog, it is also notable that GEER is significantly more accurate than relative performance as a metric. The GEER performance metric is defined as follows:
GEER = measured energy ÷ actual GHI = EER ÷ actual GHI
long term expected energy long term expected GHI long term expected GHI
There are two particularly important takeaways about the plot above. The first is that, as would be expected for increasingly complex PV performance models, the modeling errors after specific loss factors have been accounted for have a much narrower distribution than the corresponding errors using a GEER metric, which is in turn an improvement on the EER metric.
The second is that the peak of the errors for the improved model is centered much closer to the 1.0, since it represents only modeling errors and system under-performance that we were not able to categorize. Errors such as the soiling, shading, and snow losses that shift the peak of the EER and GEER curves to the left have been successfully removed.
The dashed line, at a ratio of 0.75, illustrates the difference this makes when using these different metrics to attempt to identify underperforming systems. When looking at sites with unexplained monthly underperformance in excess of 25 percent, we are squarely in the tail of the blue curve, i.e. when we are modeling drivers of system underperformance. Systems that fell into this part of the distribution had a high likelihood of significant issues during the QA process for the performance metrics. They typically arose from gradual or sporadic inverter failure, incorrect metadata, or underperformance of an uncertain cause.
In contrast, using the same monthly threshold for the GEER or EER metrics will capture much of the shoulder of the distribution. If your goal is to find all the sites that are underperforming by 25 percent for an unknown reason, at a monthly level, you will have to look at roughly 100 times as many sites if you are using a GEER metric (and the overwhelming majority of this effort will require investigating sites that are shaded or soiled).
Using an EER metric to identify the underperforming sites will be an even greater challenge due to the variations in solar irradiance that compound with modeling errors to further obscure what is occurring at individual sites. Compared to using a GEER, this would mean looking at three to five times as many sites and seeing a similar drop in the fraction of sites investigated that have an identifiable problem.
Conversely, instead of asking what fraction of sites will need to be investigated to consider all the sites below a specific performance threshold, e.g. 75 percent of expected, you could ask the question: “If I only have the time to investigate the most poorly performing one percent of the sites in my fleet every month, what performance threshold does this correspond to?” For the Waterfall, this represents the sites that are underperforming by approximately 12 percent or more, whereas for the GEER metric, this represents the sites that are underperforming by approximately 72 percent.
In practice, investigating sites that are underperforming by 12 percent for a single month may not be a particularly valuable use of time. Nonetheless, this alternate view of the problem highlights just how much this tool shifts the trade-off between aggressive performance thresholds and the number of systems that require investigation. Going for a pithier description, utilizing a tool like the Waterfall transforms the task of efficiently identifying underperforming systems from a proverbial ‘find the needle in the haystack’ type of task into a task more like, “find the needles in a small pile of hay and needles,” of which a fairly large portion appears to be needles.
That wraps up Locus Energy’s series of blog posts about the Waterfall. We are super excited about getting this project released and into production where our clients can access this data through the SolarNOC portal or our API! To learn more about Locus Energy’s VI Performance Waterfall, please click here.