Anomaly Detection Workflow Using Random Forest Regressor in Large-Scale Photovoltaic Power Plants

In large-scale photovoltaic (PV) plants, anomalies can affect performance, and maintenance schedules, and compromise long-term reliability, reducing the financial returns from the PV plant. Although the importance of anomaly detection in this context is clear, it poses a significant challenge due to the vast amount of data generated by the plant. Therefore, tools that assist in the anomaly detection process, minimizing human intervention, are essential. This paper proposes a workflow using a random forest regressor, with the application of a dynamically modeled mathematical threshold for anomaly detection. The model predicts the target power using features such as plane-of-array (POA) irradiance and temperature. To this end, Random Forest was selected after testing with other presented models, followed by the implementation of a mathematical model to establish a dynamic threshold that classifies the data as normal, alert, or anomaly. The threshold is calculated by comparing the mean absolute difference between predicted and actual power and adjusting it based on a dynamic multiplier to establish alert and anomaly limits. Subsequent tests were conducted on different inverters, applying various dataset splits. The workflow achieved excellent accuracy in detecting anomalies and alerts, with a total accuracy of 99.69%. The model successfully predicted anomalies in different inverters within the same plant. Thus, it is expected that the application of prediction methods, combined with the dynamic threshold, will serve as a support tool for anomaly detection in large-scale PV power plants.