Thesis title: Modeling and Forecasting of Photovoltaic Energy Yield

Abstract: The global energy landscape is undergoing a profound transformation, with solar photovoltaic (PV) power emerging as a central pillar of decarbonization efforts. However, the large-scale integration of this inherently variable energy source presents significant challenges to grid stability and reliability. The unpredictable nature of solar generation, primarily driven by dynamic cloud cover, necessitates the development of sophisticated tools to accurately model PV system performance and forecast its power output. This thesis addresses these critical challenges by developing and validating a cohesive suite of novel methodologies for high-fidelity solar power modeling and high-accuracy short-term forecasting. The first part of this work introduces a novel, dynamic electro-thermal PV model. Implemented in Verilog-AMS, this model transcends the limitations of conventional steady-state approaches by incorporating the thermal inertia of PV modules. Rigorous validation against high-resolution, real-world data demonstrates its superior accuracy in capturing system behavior, particularly during rapid irradiance fluctuations. To bridge the gap between physical principles and data-driven techniques, this thesis further proposes a physics-guided machine learning paradigm. By using the detailed electro-thermal model to generate physically consistent features, this hybrid approach enhances the accuracy and data efficiency of machine learning models, providing a robust tool for performance analysis and energy yield assessment. The second part of the thesis focuses on the critical task of very short-term solar power forecasting, or nowcasting. Leveraging ground-based sky imagery, this research details the development of an end-to-end deep learning framework based on a Nonlinear Autoregressive with Exogenous Inputs (NARX) architecture. This approach evolves beyond simpler multi-stage models by directly learning the complex spatio-temporal relationship between cloud patterns and PV power output. Key innovations are introduced to enhance operational relevance, including a cloud-condition-based model specialization strategy that adapts the forecasting algorithm to specific weather regimes, and a novel paradigm for forecasting cumulative energy yield over short horizons. Extensive validation demonstrates that the proposed framework achieves state-of-the-art accuracy, consistently outperforming established persistence and machine learning benchmarks. Collectively, the contributions of this thesis provide a significant advancement in the ability to understand, simulate, and predict the behavior of PV systems. By delivering validated, high-performance tools for both detailed modeling and real-time forecasting, this work supports the optimization of solar assets and facilitates their seamless and reliable integration into the power grids of the future.

Supervisor: Dimitrios Soudris

PhD Student: Dimitrios Anagnostos