Comparative Analysis of Data Representation Methods for PV Anomaly Detection: Raw Time-Series Data vs. Diverse Binning Strategies

As the global cumulative capacity of photovoltaic (PV) systems continues its rapid expansion, reaching approximately 2.2 TW by the end of 2024, the frequent occurrence of various facility failures has become a critical challenge, leading to generation losses and increased maintenance costs. While previous studies have predominantly focused on binary anomaly detection, they often lack the capacity for root cause analysis and maintenance prioritization. To address these limitations, this study defines nine major anomaly types based on actual PV operational data and evaluates the performance of binary classification for each specific anomaly type to detect its individual occurrence. To optimize classification performance, we conducted a comparative analysis of various machine learning and deep learning models—including tree-based, distance-based, and Transformer-based algorithms—across four data representation methods: raw data and three distinct binning strategies (domain knowledge-based, K-means clustering-based, and decision tree-based). Experimental results demonstrate that deep learning models capable of processing raw time-series data, specifically CNN and Transformer models, achieved the highest performance. These findings provide a more robust framework for multi-anomaly diagnosis in large-scale PV plants and suggest a strategic direction for future research in machine learning models.