Hierarchical modeling of grid-connected photovoltaic systems for fault detection and localization: Case study of the Ziga PV plant in Burkina Faso

• Full-chain hierarchical physics-based modeling of grid-connected PV plants under data-scarce conditions • Multi-variable residual analysis on power, voltage, current, and temperature with adaptive thresholds. • Validated on the 1.1 MWp Ziga real PV plant with a MAPE of 2.7%, NRMSE of 3%, and R 2 of 88%. • Fault injection validation yields a global Detection Rate of 95%, reaching 100% for individual fault types. • Competitive performance against AI-based methods with no labelled fault data required. Grid-connected photovoltaic (PV) power plants are a key technology for advancing Sustainable Development Goals. This is particularly relevant in tropical regions, where high solar irradiance is very often accompanied by high occurrence of faults. Existing diagnostic methods either address the PV array or the inverter alone, using output power as the only diagnostic variable. This single-variable approach is fundamentally limited. Indeed, distinct fault types can produce identical power losses, and faults in string combiner box and connection cables that remain undetectable. Recent AI-based methods achieve accuracy rates of up to 100%, but require large labeled datasets and do not cover all subsystems. Few existing methods allow for simultaneous fault monitoring across the full energy production chain. This study addresses this gap through three main contributions. First, a hierarchical physics-based model is proposed, covering the PV array, the string combiner box, connection cables, and the grid-tied inverter along the energy flow direction. Next, the residuals between measured and simulated values of power, voltage, current, and temperature are simultaneously computed at each subsystem boundary. The residuals are then compared against dynamic thresholds to detect, locate and identify the faults. Finally, a site-adapted meteorological correction procedure is applied to data collected 14 km from the plant, ensuring full grounding in real operational data. The entire methodology was implemented using MATLAB/Simulink 2021b. The simulation results are validated on a 1.1 MWp grid-connected PV plant. Energy yield modeling achieves a mean absolute percentage error (MAPE) of 2.7%, a normalized root mean square error (NRMSE) of 3%, and an R 2 of 88%. Fault detection yields a global detection rate of 95%, reaching 100% for individual fault types, with a false alarm rate of 0%. For comparison, ANN-based methods report detection rates of 82.2%-92.6% [1] , while BA-XGBoost achieves 87.56% [2] . Unlike these data-driven approaches, the proposed method requires no labeled fault data and no training phase.