A Physics-Informed Digital Twin Framework for Explainable Operational Health Monitoring of Utility-Scale Photovoltaic Systems Using SCADA Measurements

The increasing deployment of utility-scale photovoltaic (PV) systems necessitates advanced monitoring frameworks capable of providing interpretable operational assessments beyond conventional SCADA visualization and threshold-based supervision. This study proposes a physics-informed analytical digital twin framework for monitoring the operational behavior of a grid-connected PV system using synchronized three-phase electrical measurements. The framework integrates SCADA data acquisition and preprocessing, physics-informed electrical feature engineering, Composite Health Index (CHI) construction, feature redundancy analysis, dimensionality reduction, HDBSCAN-based operational-state discovery, and interpretable surrogate machine learning modeling. The framework was developed using 17,568 half-hourly SCADA observations from a 110 kW PV plant in Ankara, Türkiye, in 2024. Six measured electrical variables—including three-phase voltages, total DC power, total AC active power, and daily energy production—were transformed into physically interpretable indicators representing voltage balance, phase symmetry, conversion efficiency, energy utilization, and power stability. These indicators were integrated into a normalized CHI for unified operational assessment. The resulting analytical workflow identified three data-driven operational states, characterized as healthy, transitional, and degraded, using principal component analysis and HDBSCAN clustering. Cluster-quality assessment, CHI-weight sensitivity analysis, and surrogate-label reproducibility evaluation were conducted to examine the consistency and interpretability of the framework. The final Random Forest surrogate model achieved 97.3% accuracy, 97.4% precision, 97.3% recall, and a 97.1% F1-score in reproducing the HDBSCAN-derived operational-state labels. Permutation feature importance analysis was used to examine the contribution of the engineered electrical indicators to state-label reproducibility. The 2025 SCADA dataset was subsequently used as a same-site, out-of-time temporal evaluation of the fixed analytical workflow rather than as an independent external validation dataset. The results demonstrate the potential of combining physics-informed electrical indicators, transparent health indexing, unsupervised operational-state discovery, and interpretable machine learning to transform routinely available SCADA measurements into operational health information. The proposed framework supports continuous monitoring, early indication of abnormal operational patterns, and maintenance-related decision-making. Its current implementation does not include bidirectional cyber–physical synchronization, autonomous actuation, closed-loop control, or independently verified component-level fault diagnosis.