ED Sciences Physiques et de l'Ingénieur
Advanced Strategies for Online Monitoring, Fault Detection, and Identification in Photovoltaic Systems: A Hybrid Model-Based and Data-Driven Approaches
by Yehya ALRIFAI (ESTIA-Recherche)
The defense will take place at 14h00 - 211 Ecole d'ingénieurs ESTIA, Technopole Izarbel 90 Allée Fauste d'Elhuyar, 64210 Bidart
in front of the jury composed of
- Ionel VECHIU - Professeur - Université de Bordeaux - Directeur de these
- Jérôme BOSCHE - Professeur des universités - Université de Picardie - Rapporteur
- Carlos Manuel ASTORGA ZARAGOZA - Professor - Instituto Tecnológico de México - Rapporteur
- Delphine RIU - Professeure des universités - Grenoble INP - Examinateur
- Ascension ZAFRA CABEZA - Associate Professor - Universidad de Sevilla - Examinateur
- Yann Eric BOUFFARD VERCELLI - - Schneider Electric -
The increasing integration of photovoltaic (PV) systems as sustainable energy sources introduces critical reliability concerns, particularly regarding fault detection on the DC side, where conventional protection mechanisms often fall short—especially under low-irradiation or incipient fault conditions. This doctoral research addresses these challenges by developing two distinct fault detection and diagnosis (FDD) strategies based on global Maximum Power Point (MPP) measurements, offering a low-cost and scalable solution. The first strategy adopts a hybrid approach, combining a statistical regression model with a Kalman Filter to account for noise and uncertainty, followed by a rule-based diagnostic layer utilizing adaptive thresholds to identify fault signatures. This method is especially effective for early-stage fault detection and provides a structured mechanism to handle measurement variability across diverse operating conditions. The second strategy follows a data-driven paradigm, leveraging artificial neural networks (ANNs) with a simplified architecture based on PV resistance and logarithmic transformation. It is subdivided into two schemes: The first incorporates classical supervised classifiers (Decision Tree, SVM, Naïve Bayes, K-NN) and ensemble methods (Random Forest, Bagging, AdaBoost, RUBoost, Subspace K-NN), each evaluated with various kernels and structural configurations. The second builds upon this by employing a Stacking ensemble framework, integrating the most performant base classifiers via a meta-learner. A novel weighted recall strategy is proposed to enhance class diversity and improve fault identification accuracy. Extensive simulation studies across multiple weather scenarios and fault types—such as partial shading, open/short circuits, cross-string faults, and degradation modes—demonstrate the robustness of both approaches. The hybrid method achieves a detection rate of 90.12%, although its scope is limited to array-level diagnostics and struggles with class overlap in certain fault types. The data-driven approaches exhibit greater adaptability, with the Stacking-based model reaching over 95% accuracy, especially when supported by appropriate data preprocessing (e.g., outlier filtering, class balancing, normalization). In sum, this thesis advances the state-of-the-art in PV system fault diagnosis by proposing robust, scalable, and interpretable strategies tailored for real-time deployment in noisy and nonlinear environments.