Researchers in Cyprus evaluated six different models that were used to predict power losses caused by the accumulation of dust, dirt and other substances on the surface of solar panels in the island’s dry climate. The results of the various models were compared with data on the fouling of a “test bench” installation at the University of Cyprus in Nicosia, revealing a potential advantage in machine learning methods supported by satellite data.
Dealing with losses caused by dust accumulating on the surface of a module is big business for the PV industry, as these losses can quickly lose significant amounts of revenue. Cleaning modules too often or placing them in the wrong type of cleaning equipment can also hurt the project’s economics. And so the ability to accurately predict fouling losses in both the long and short term is something that is highly valued by PV project developers and system operators.
There are different approaches that use different combinations of in-situ sensors, historical climate data, local weather data, satellite imagery, and more. A team of researchers led by the University of Cyprus attempted to compare the accuracy of some of these by comparing modeled predictions of fouling with data from a test installation on the campus of the University of Cyprus in Nicosia.
Machine learning
The soiling of the test site was calculated by comparing the cleaned and uncleaned module side by side. Six different models – three using physical modeling and three based on machine learning – were evaluated for accuracy with the site’s data.
The three “physical” models are established methods for dirt modeling, while the machine learning methods are open-source programs applied for the first time to soil measurement. Full details of the models used and their evaluation can be found in the article “Characterizing fouling for photovoltaic systems in an arid climate: a case study in Cyprus”, published Solar energy.
The evaluation showed that physical models fed with data observed in the field achieved the highest accuracy. The error rates (root mean square error) were 1.16% for daily fouling and 0.83% for monthly fouling in the most effective machine learning model. , called CatBoost.
However, machine learning approaches were not far behind: the error was 1.55% for daily contaminations and 1.18% for monthly contaminations. The researchers point out that given the shortcomings in the availability of field observations covering the entire region over a sufficiently long period of time, machine learning models based on environmental data collected via satellite could also be a useful approach.
“Modeling pollution with this kind of satellite-derived environmental data can help design O&M strategies and actions to minimize pollution losses throughout the year, especially in dusty and dry regions where there can be sudden changes in aerosol loads and rainfall is much less frequent,” the researchers explained.
