Forecasting to Welcome the Solar Era: Supporting the integration of solar photovoltaics in the electricity grid

Abstract

Solar Photovoltaic (PV) systems form a key technology to cut greenhouse gas emissions and decarbonize the energy system. In recent years, an unprecedented growth of the installed capacity of solar PV systems has been observed and is expected to continue in the coming decades. The intermittent and variable nature of PV power generation affects the operation of the electricity system, posing challenges in integrating substantial amounts of solar PV. The adoption of solar power forecasting models provides a cost-effective solution to address these challenges and facilitate successful integration of substantial solar PV capacity. This thesis develops day-ahead and intraday solar power forecasting models, investigates their application and explores potential measures to enhance their performance. After giving an overview of the state-of-the-art of solar forecasting, this thesis ponders on ensuring high-quality data, which is essential to develop and operate (forecasting) models. Hence, a methodology is proposed for the quality control of power measurements from PV systems. The methodology is tested on an extensive dataset containing power measurements from 175 PV systems and is made publicly available as a Python package, along with the dataset. Next, fifteen solar power forecasting models are successfully developed and tested for their intended application. These models include physical, regression and artificial intelligence based models including deep learning. The performance of the forecasting models is examined considering technical and economic performance metrics. The results show the overall superiority of probabilistic over point forecasting models. Additionally, tree-based models are found to outperform alternatives including physical, regression and deep learning models in light of their technical performance. The economic results depend on the case study and are, therefore, less unambiguous. The findings in this thesis demonstrate that forecasts can be improved through incorporating several physical models used to generate expert variables in pre-processing steps. These physical models comprise a decomposition, transposition and PV model, where the adoption of the former two lead to the most substantial performance gains. This thesis also assesses the effect of aggregating PV systems on the performance of the forecasting models. This positive effect is attributed to the number of systems considered and the geographical region they cover. Next, the value of extending day-ahead solar power forecasting models with intraday models is assessed, which gives an indication of the performance improvement as a result of a diminishing forecast horizon. The findings present a technical and economic performance gain of 46 and 21% as a result of the extension. Last, the research in this thesis explores the relation between the technical and economic performance of solar power forecasting models, revealing a non-linear relationship between these metrics. In summary, the results presented in this thesis show the capability of forecasting models to predict the PV power generation. This thesis furthermore demonstrates that solar forecasts can be improved through incorporating pre-processing steps, aggregating PV systems and reducing the forecast horizon. Finally, a non-linear relation between the technical accuracy and economic value of solar forecasting models is exposed.