March 31, 2020 — Industry Trends
Could machine learning triple the productivity of your solar farm?
Machine learning based on real-world data gives a far better chance of measuring the performance of solar panels than the theoretical models or industry standards that operators currently rely on. So will machine learning replace those models?
Head of Research, Greenbyte
Allow me to calm your nerves at the outset: robots are not coming to take over the world.
They could however help us make the world a better place and that, in a nutshell, is my team’s mission. We research, we question, we test and, above all, we refuse to accept that the status quo is perfect.
The advent of machine learning represents a monumental opportunity to solve problems and improve processes in all walks of life. We’ve been looking at what this might mean for the owners and operators of solar PV assets and the early signs are extremely encouraging.
Our starting thesis was that the solar industry has long accepted second best when it comes to assessing the performance of its assets.
Whereas for the wind industry, it is relatively straightforward to calculate the potential power of each individual turbine by cross-comparing weather data with the appropriate manufacturer power curve, things are not so simple in the solar space.
Typically, solar asset owners and managers will consider localised irradiance and temperature data but with no guaranteed power curves from the manufacturers, they instead use standardised industry equations that are applied across all panel types.
Because it’s widely accepted that these industry standards tend to over-estimate performance, many asset managers tell us they’ve learnt not to stress about asset performance ratio unless it dips below 80% of the standard.
All of which means that arbitrary solar performance KPIs can become stuck in time (‘it’s the way we have always done it”) and no solar asset owner can consider the actual versus potential performance of their assets in any great detail (since they don’t really know what the potential performance really is). Nor therefore can they consider the amount of money being left on the table through sub-par productivity.
Enter machine learning
Because machine learning employs real world data, as opposed to generic industry standards, it has a better chance of providing an accurate estimation of the potential power – or ‘Modeled Energy’ - of each individual inverter in each individual location.
With access to sufficient historical data from a site, it is possible to create a profile for each inverter based on how that inverter performed previously in similar meteorological conditions. A power curve custom-tailored to each inverter’s unique characteristics.
This is a solution that Greenbyte continues to work on in partnership with a few of its major customers and the early field test results make for compelling reading.
For example, in the case of one Independent Power Producer in the USA, over the course of January 2020, our machine-learning approach identified more than three times the level of lost production as compared with the traditional performance KPI. That’s triple the potential to enhance asset productivity!
So, move over industry standards, the robots are coming and they’re bringing with them a superior KPI for PV site performance assessment, right?
Wrong. In our view, advances in machine learning like this can enrich the user experience, furnishing asset owners with actionable data that can ultimately be turned into productivity gains. But the approach is only ever as reliable as the data that goes in.
Data quality and consistency are prerequisites to success in the field of machine learning and it is unrealistic to expect that these ingredients will be readily available in all instances.
In any case, asset owners should be wary of replacing apples with pears.
The machine-learning approach considers resource availability and will therefore give a more accurate assessment of lost output and thus lost revenue. Whereas the traditional benchmarking calculation does not consider the resource available but the performance of each inverter against its peer group.
Both the calculations have their benefits and drawbacks as shown below:
We’d sooner adopt a ‘stronger together’ mindset whereby the pros of each approach are embraced and the cons mitigated. In this fashion, 1+1 can really = 3!
My team continues to trial its data-driven potential power calculation for PV assets. If you’re a Greenbyte customer and would like to talk to the team or get involved, please contact us. For all others, watch this space for further blog posts on the theme!