Master's thesis: Development of advanced cloud classification and segmentation models for solar energy applications using deep learning and synthetic data augmentation

The Institute of Solar Research is committed to advancing the application of solar energy. Reliable information on solar resources is essential. The Institute's Energy Meteorology Group collects meteorological data and identifies parameters essential for the design, operation and qualification of solar power plants. Clouds are the main cause of shortwave solar irradiance variability, which affects the performance of solar power plants and electricity grids. The Institute of Solar Research is one of the leading groups in the development of short-term solar irradiance forecasts. The automatic detection and classification of clouds based on high-resolution sky images is one of the most important steps in the generation of such forecasts. The use of deep learning (DL) in computer vision has led to massive advances in this area. However, due to the high variability and complexity of the conditions, there is still a need for improvement, especially under multi-layer cloud conditions. One challenge in training such models is the lack of sufficient reference data. The tedious task of manually labeling images is very time consuming and error prone.

The aim of this thesis is to create and evaluate a large reference dataset for deep learning cloud detection and classification models. The dataset will be created by combining synthetic data with weakly labeled and labeled data. The benefits of optical flow techniques for further data augmentation will be evaluated, especially under complex multi-layer conditions.

In order to achieve this goal, the following tasks will be carried out:

Location: Almería, Spain
Minimum Period: 6 months (extendable)