Datagedreven Foutdiagnose voor Fotovoltaïsche Systemen

Despite the rapid growth of photovoltaic (PV) energy, hidden yield losses from gradual degradation, soiling, and abrupt faults continue to erode the economic viability of PV installations worldwide. Manual inspection remains the dominant practice, yet it is labour-intensive (up to 8 h/MW), error-prone, and impractical for large-scale or remote plants. This thesis introduces a comprehensive, fully automated, data-driven monitoring and fault-handling framework that operates exclusively with measurements that are already available in the vast majority of modern PV systems: plane-of-array irradiance (G_PoA), ambient conditions, and DC-side electrical parameters (I_MPP, V_MPP, P_MPP) at inverter and, when present, array/combiner-box level. The core of this framework is a high-fidelity degradation-informed digital shadow that predicts the expected electrical output of a healthy, progressively ageing PV system. The shadow is built in two main stages: Accurate PV output estimation (thermal and electrical modelling) Accurate prediction of electrical output requires first an accurate estimate of module temperature (T_PV), which strongly influences voltage and power. An enhanced version of the Faiman thermal model is developed with a multi-step calibration procedure incorporating exponential weighted moving averages, linear correction, and the first-ever calibratable wind-direction-dependent cooling weights (f(WD)). This yields MAE = 0.96-1.33 °C and RMSE = 1.31-1.86 °C across diverse mounting configurations (free-field, rooftop with wind deflectors, ground-mounted), representing 16-40 % improvements over the current state-of-the-art models. A novel 'blind' calibration method is introduced that extracts accurate coefficients without any on-site T_PV sensor, using only inverter V_MPP and weather data, enabling application to legacy or sensor-limited plants. Using the resulting high-quality T_PV (measured or estimated) together with G_PoA, multilayer perceptron (MLP) regressors are trained to predict P_MPP, I_MPP, and V_MPP. Systematic comparison with KNN and XGBoost, including simulated temperature-ramp tests, demonstrates that regularised MLP provides the smoothest response and lowest uncertainty. The static model is transformed into a dynamic, degradation-adaptive version via a new daily Baseline Efficiency (BE) metric that captures non-linear degradation and soiling without restrictive assumptions. By feeding the product of BE and G_PoA into the static model, long-term prediction error remains stable at nRMSE lower than 3 % (typically 1.6-2.2 %) over 7+ years on fixed-tilt, dual-tilt, and bifacial systems, a substantial improvement over linear degradation rate approaches, which can exceed 10 % error in non-linear cases. Year-on-Year analysis of BE yields degradation rates matching IEA-PVPS Task 13 reference values within ±0.2 %/year. Fault detection with dynamic irradiance-dependent thresholds Residuals between measured and degradation-corrected predictions are compared against thresholds derived from the 1st/99th percentiles within 25 W/m² irradiance bins on high-quality reference days (filtered by R² higher than 0.9 and EPI). Faults are declared only when deviations persist for more than 2 hours, reliably separating persistent DC-side faults (line-line faults, open circuits, bypass-diode failures, partial disconnection, etc.) from transient shading or clouds. The method successfully labels real multi-day faults on real world datasets. Unsupervised fault diagnosis via signature clustering Faults are characterised by degradation-corrected performance indices (CPI*, VPI*) at inverter and (when available) string level. Extensive tests on simulated faulted datasets (2×12 and 5×5 topologies) show that Hierarchical Density-Based Spatial Clustering (HDBSCAN) with separate string-wise clustering and composite label creation achieves completeness higher than 0.97 and homogeneity higher than 0.69 even under realistic ±2 % measurement noise, better than K-means, Fuzzy C-Means, and standard Density-Based Spatial Clustering (DBSCAN). With string currents, the algorithm consistently groups faults by severity and affected string(s), while merging fault types of identical electrical impact. On real-world data from the composite clustering label immediately reveals implicated arrays during partial reconnection events after downtime, dramatically reducing diagnostic time compared to conventional alarm lists. The proposed pipeline is automated, runs on standard monitoring hardware, and scales from residential to multi-MW plants. It eliminates the need for IV-curve tracers, reference modules, or labelled fault databases, making it immediately deployable on the existing global PV fleet. Beyond photovoltaics, the same anomaly-detection to impact quantification to signature-clustering paradigm is shown to be directly transferable to wind-turbine SCADA data and electric-vehicle efficiency monitoring. In conclusion, this thesis delivers a practical, highly accurate, and industrially viable solution that transforms raw time-series data into actionable maintenance intelligence, significantly reducing operation & maintenance costs and approaching the vision of fully autonomous PV asset management.