US researchers say a self-guided machine learning tool could identify long-term physical defects in solar power systems weeks or years before conventional inspections, potentially reducing operating and maintenance costs.
Researchers from Stony Brook University, in collaboration with Ecosuite and Ecogy Energy, have done just that developed a self-supervised machine learning algorithm designed to identify physical anomalies in solar energy systems. The project aims to reduce operation and maintenance (O&M) costs, which remain a major hurdle to project economics as the industry scales.
Using historical datasets from Ecogy Energy, researchers Yue Zhao and Kang Pu trained the anomaly detectors using a holistic pipeline that integrates inverter performance with weather data. The approach avoids non-standard measures, opting instead for publicly available generation and environmental data to ensure the tool remains operational in diverse data environments.
The research places particular emphasis on long-term anomalies, including the underlying physical issues that often escape the attention of asset managers until they result in significant downtime or hardware failure. When applied to updating solar energy generation data and weather data, the detectors are designed to predict and diagnose these problems weeks or even years in advance.
“A more accurate and timely awareness can significantly improve the efficiency and effectiveness of O&M practices,” the researchers said.
The study outlines three key impact areas:
-
Efficient Scheduling: Site visits for maintenance personnel can be optimized to address specific detected issues, reducing unnecessary truck rolls.
-
Asset Life: Timely hardware maintenance can significantly extend equipment life and reduce the need for expensive, premature replacements.
-
Revenue Protection: Addressing system issues before they lead to disruptions can minimize loss of energy production.
As the solar portfolio grows, the ability to automate the detection of underperformance becomes critical. Recent industry data shows that U.S. solar installations lost an average of $5,720 per MW due to equipment problems in 2024 alone. These “software superpowers,” as researcher John Gorman calls them, provide advanced alerts directly from already paid for edge computing hardware.
“Adding these software superpowers creates immediate value, but as part of a flexible ecosystem, our machine learning algorithms can also evolve,” Gorman said. “Being able to translate the learnings from one system to another is our next goal, unlocking value across a portfolio.”
The researchers believe that as these algorithms evolve, they will enable asset managers to move from reactive maintenance to a proactive, predictive model that maximizes the financial and operational health of renewables.
This content is copyrighted and may not be reused. If you would like to collaborate with us and reuse some of our content, please contact: editors@pv-magazine.com.
