Evaluation of Machine Learning Algorithms for Supervised Anomaly Detection and Comparison between Static and Dynamic Thresholds in Photovoltaic Systems

Abstract

The use of photovoltaic systems has increased in recent years due to their decreasing costs and improved performance. However, these systems can be susceptible to faults that can reduce efficiency and energy yield. To prevent and reduce these problems, preventive or predictive maintenance and effective monitoring are necessary. PV health monitoring systems and automatic fault detection and diagnosis methods are critical for ensuring PV plants’ reliability, high-efficiency operation, and safety. This paper presents a new framework for developing fault detection in photovoltaic (PV) systems. The proposed approach uses machine learning algorithms to predict energy power production and detect anomalies in PV plants by comparing the predicted power from a model and the measured power from sensors. The framework utilizes historical data to train the prediction model, and live data is compared with predicted values to analyze residuals and detect abnormal scenarios. The proposed approach has been shown to accurately distinguish anomalies using constructed thresholding, either static or dynamic thresholds. The paper also reports experimental results using the Matthews correlation coefficient, a more reliable statistical rate for an imbalanced dataset. The proposed approach leads to a reasonable anomaly detection rate, with an MCC of 0.736 and a balanced ACC of 0.863.