Abstract

The increasing worldwide adoption of solar energy motivates the need for innovative solutions to optimize the operations and maintenance (O&M) of solar energy systems. Soiling, i.e., the accumulation of material such as dust on solar photovoltaic (PV) arrays, poses a key challenge for solar operators as it can compromise the performance and reliability of solar energy systems. To address this challenge, we develop a machine-learning-(ML)-based in-situ monitoring solution in order to timely detect the occurrence of soiling events, thereby alerting operators to take necessary mitigation measures to ensure the cost-effective management of their assets. First, an experimental testbed has been set up at the Energy Lab at Rutgers University–New Brunswick, wherein a new optical sensing technology is installed on an operational PV system in order to collect sensor measurements about the soiling ratio under normal and abnormal conditions. Then, a cluster of ML classifiers are trained on those sensor data in order to effectively dis-entangle the random variations in soiling ratio versus systematic drops due to ongoing soiling, and finally output a probabilistic decision on the soling status of the PV system. Our experiments on three artificially generated soiling events in 2023 suggest that kernel-based models such as support vector machines–when supplemented with a newly introduced feature that describes the rate of change in soiling ratio—can produce up to 99% accuracy in predicting soling events, while maintaining a sensible balance between false and missed alarms (up to 85.4% in F1-scores).