Photovoltaik-Monitoring mit Methoden des maschinellen Lernens

The objective of this dissertation is the development and evaluation of machine learning models for fault detection and energy yield forecasting in photovoltaic (PV) systems based on real-world monitoring and weather data. Given the increasing share of solar energy generation and the limited efficiency of current PV technologies, reliable detection of performance losses and operational faults is essential for the efficiency, lifetime, and economic viability of PV installations. A particular focus is placed on the processing and preparation of large, heterogeneous field data generated in real operational environments. A central challenge lies in transforming noisy, incomplete, and inconsistently structured data into representations suitable for machine learning methods. The study investigates multiple algorithms, including long short-term memory (LSTM) networks, variational autoencoders (VAE), ridge regression, and multilayer perceptrons (MLP). The models are evaluated with respect to accuracy, robustness, and generalizability across a diverse set of PV systems. To support the experimental workflow, a modular software system was developed, covering the entire pipeline from data acquisition and preprocessing to model training and result visualization. The system was iteratively extended to integrate new data sources, model types, and application scenarios. The results demonstrate that, with appropriate preprocessing and model selection, reliable detection of typical faults such as shading, snow coverage, soiling, and defective diodes is achievable. The best forecasting performance for energy production was obtained using LSTM networks. In addition, unsupervised approaches enable the identification of previously unknown fault patterns. An integrated feedback mechanism allows the generation of labeled data, which is used to further improve supervised learning models. Overall, this work contributes to data-driven optimization of PV systems and demonstrates practical strategies for applying machine learning under real-world data constraints.