Uso de machine learning para análise de subperformance de usinas solares

Abstract

The expansion of large-scale solar power generation imposes increasing challenges for operation and maintenance (O&M), as traditional monitoring methods based on fixed thresholds, such as Supervisory Control and Data Acquisition (SCADA) systems, often fail to identify subtle anomalies. This work presents an automated fault detection system for stringboxes in utility-scale solar plants using machine learning techniques. The approach proposes a One-vs-All architecture based on the XGBoost algorithm to identify three distinct anomaly classes: zero-output stringboxes (total loss of communication or power), underperformance (continuous low efficiency), and shading (temporal distortions in the generation curve). The methodology used real data from a plant in Northeast Brazil, applying feature engineering that extracted 28 statistical and morphological features from daily time series. The results validated the system’s effectiveness: the zero-output detector achieved ideal performance (100% F1-Score); the underperformance detector obtained 95.22% accuracy and an F1-Score of 0.7059, demonstrating robustness against false positives; and the shading detector, validated via stratified k-fold, reached an F1-Score of 0.6368, proving the ability to capture temporal patterns even in imbalanced data scenarios. The system proves to be a viable tool to support O&M decision-making, overcoming the limitations of purely deterministic approaches.