Adaptive Thresholding for Production Imbalance Detection in Renewable Energy Systems Using Statistical and Machine Learning Models

Abstract

The exponential growth of renewable energy installations worldwide has created an urgent need for sophisticated monitoring and anomaly detection systems that can ensure optimal performance, prevent equipment failures, and maintain energy security. This comprehensive study presents a novel multialgorithm anomaly detection framework specifically designed for renewable energy facilities, with a particular emphasis on detecting production imbalances across multiple energy generation units. The proposed system integrates four distinct machine learning approaches: Isolation Forest for outlier detection, One-Class Support Vector Machine for boundarybased anomaly identification, a novel, optimized Statistical Threshold method with adaptive mechanisms, and Autoencoder networks for reconstructionbased anomaly detection. The evaluation on a realworld dataset from a solar power facility in Jordan demonstrated that the Adaptive Statistical Thresholding (AST) model achieved the most balanced performance and the highest score. Specifically, AST achieved an F1-Score of 0.924, outperforming Isolation Forest (F1-Score: 0.772) and the Autoencoder model (F1-Score: 0.486). This framework features a novel adaptive algorithm that calculates the threshold, dynamically adjusting it based on the 90th percentile of reconstruction errors over a 7-day rolling window. This is crucial in overcoming the natural variability and seasonal fluctuations inherent to renewable energy systems. A thorough feature engineering operation was applied to identify all-important indicators of operation, such as production ratios, performance measures, time trends, and statistical indicators that portray the sophisticated nature of multi-unit renewable energy plants. The system was extensively evaluated using real-world operational data from Jordan's renewable energy infrastructure, comprising 4,464 high-resolution records from three energy production meters. Experimental assessment of the performance indicates an F1-score of 0.924, precision of 0.878, and recall of 0.975 of the Statistical Threshold method, confirming its superior operational stability and performance in this specific context.