Researchers from the United States’ Stony Brook University, in collaboration with software solutions provider Ecosuite and renewables developer Ecogy Energy, have developed a self-supervised machine-learning algorithm designed to identify physical anomalies in solar energy systems. The project aims to reduce operations and maintenance (O&M) costs, which remain a significant hurdle for project economics as the industry scales.
By leveraging historical datasets provided by 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, instead opting for widely available generation and environmental data to ensure the tool remains operational across diverse data environments.
The study places a particular focus on long-term anomalies including the underlying physical issues that often evade the notice of asset managers until they result in significant downtime or hardware failure. When applied to updating solar generation and weather data, the detectors are designed to predict and diagnose these issues weeks or even years in advance.
“More accurate and timely awareness can greatly improve the efficiency and effectiveness of O&M practices,” the researchers said.
The study outlines three primary areas of impact:
Efficient scheduling: Site visits for maintenance personnel can be optimised to address specific detected issues, reducing unnecessary truck rolls.
Asset longevity: On-time hardware maintenance can significantly prolong equipment lifetime and reduce the need for expensive, premature replacements.
Revenue protection: By addressing system issues before they lead to failure, the loss of energy production can be minimised.
As solar portfolios grow, the ability to automate the detection of underperformance becomes critical. Recent industry data shows that U.S. solar facilities lost an average of $8,566 (USD 5,720) per MW to equipment issues in 2024 alone. These “software superpowers,” as researcher John Gorman calls them, provide advanced warnings directly from already-paid-for edge-compute hardware.
“Adding these software superpowers immediately creates value, but as part of a flexible ecosystem, our machine learning algorithms can also evolve,” Gorman said. “Being able to translate learnings from one system to the next is our next goal, unlocking value across a portfolio.”
The researchers believe that as these algorithms evolve, they will allow asset managers to transition from reactive maintenance to a proactive, predictive model that maximises the financial and operational health of renewable energy assets.
