Implementierung von KI-Algorithmen zur Anomaliedetektion in PV-Anlagen und Aufbau einer Leitwartenapplikation

Maximising energy yield and ensuring reliable operation are key challenges for photovoltaic (PV) systems. Using artificial intelligence (AI) for the early detection of anomalies such as shading can significantly minimise losses. This thesis analyses the suitability of Support Vector Machines (SVMs) and Convolutional Neural Networks (CNNs) for detecting shading faults in a real PV system at the Fraunhofer IFF. As part of this process, a supporting monitoring dashboard is being developed. Four scenarios involving different levels of feature engineering were compared using data from inverters and weather stations. The SVM models consistently showed better performance and were significantly more computationally efficient (47 seconds vs. nearly six hours) than the CNNs. The results of the study show that a targeted configuration of specific technical features leads to a significant increase in performance. The analysis revealed that an SVM with minimal feature processing achieved optimal results, with an accuracy of 82,06% on the test data. The best CNN model achieved an accuracy of 80,62%. The developed dashboard successfully integrates the best SVM model and improves the transparency of the results with the help of explainable artificial intelligence (XAI). The work demonstrates the potential of AI algorithms for efficient shading detection and provides a functional dashboard application.