Semantic Cloud Segmentation for All Sky Images using Deep Learning and LWIR Data

Kurzfassung

The transition toward sustainable and renewable energy sources represents a significant challenge. Among different alternatives, solar power is the most abundant, however, its integration into the global energy mix is complicated, due to the variability of solar irradiance. While seasonal and diurnal changes are predictable and well-researched, intra-hour fluctuations, which primarily come from cloud cover, introduce instability into the power generation grid. The development of reliable short-term Nowcasting systems is important to maintain grid stability, prevent damages through power ramps in photovoltaic parks and optimize the thermal efficiency of concentrated solar power plants. A state-of-the-art approach to Nowcasting utilizes ground-based all-sky imagers (ASI) to capture hemispherical sky images. Clouds are detected and tracked in real-time to generate future irradiance maps. Furthermore, the utility of ASI systems extends to optical satellite communications, where a clear line of sight is mandatory. By leveraging thermal infrared sensors, which remain operational during night, cloud-free windows can be detected to optimize transmission schedules. The fidelity of these nowcasts is highly dependent on the accuracy of the cloud detection phase. In recent years, deep learning methods have emerged as dominant, outperforming conventional detection techniques. However, those models introduce the critical dependency on large amounts of high-quality manually labeled ground truth data. Since manual pixel-wise annotation is time-intensive and costly obtaining such datasets remains a primary bottleneck in the field. This thesis aims to remove that bottleneck by using the recently acquired long-wave infrared data to generate an automatically labeled ground truth height map dataset. A combination of infrared imagery and reanalysis data is used to generate height maps as pixel-wise ground truth to train a convolutional neural network for cloud detection and pixel-wise height estimation. The current state-of-the-art cloud detection model is used to generate binary ground truth masks for cloud detection in infrared images. Evaluation on a predefined test set containing different cloud formations, shows accurate cloud detection with a mean IoU of 0.777. The cloud base height is predicted with a mean absolute error of 1046.09 m and a bias of -399.76 m from a single visible light input image. The conversion capability from the obtained height predictions to cloud height layer maps is limited with a mean IoU over all 4 classes of 0.372. The cloud detection in infrared images performs strongly on the test set with a mean IoU of 0.835 and F1 score of 0.908.