Semi-supervised Learning for Probabilistic Cloud Detection in Ground-based Imagery

Kurzfassung

One of the most significant challenges of our era is the transition to clean and renewable energy sources. The most abundant energy source on Earth is the sun, yet it is not consistently available due to the variability in solar irradiance. The most significant impact on local fluctuations in solar irradiance is the presence of clouds. Nowcasting systems provide intra-hour forecasts to anticipate this short-term variability. These forecasts are beneficial for grid control, the reduction of ramps in large-scale photovoltaic parks, and the operation of concentrating solar power (CSP) plants. A common approach to nowcasting is to use ground-based sky images from All-Sky Imagers to detect clouds, which are then further tracked in a physical nowcasting system to estimate the future irradiance. The quality of these nowcasts depends highly on the quality of the cloud detection, which is commonly performed on the pixel level. Recent years have demonstrated that deep learning-based methods outperform all existing conventional approaches on these kinds of tasks. The challenge of deep learning-based methods is that they require a significant amount of high-quality, human-annotated allsky images with annotations at the pixel level, which are in practice costly to obtain. Furthermore, the uncertainty in the predictions of systems based on artificial intelligence is often difficult to predict. This thesis proposes an approach based on semi-supervised learning that fuses measurements from a ceilometer, which is a LiDAR sensor used to measure the cloud base height, into the learning process of a camera-based cloud detection model to improve the detection of three different cloud layers. A dataset of all-sky images with over 47000 weakly annotated images is created based on heuristics. This dataset is employed in conjunction with 770 human annotated all-sky images to train a cloud detection model, utilising semi-supervised learning techniques such as pseudo-labeling and consistency regularisation. Furthermore, a probabilistic calibration method is applied as a post-processing step to calibrate the uncertainty estimates of the developed cloud segmentation model for predictions on unseen all-sky images. Evaluation on a benchmark of 36 all-sky images from three cameras shows the effectiveness of the two methods. The semi-supervised learning method demonstrated superior accuracy and IoU, respectively, by 2.4% and 3.5% compared to the current state-of-theart method for semantic cloud segmentation. Calibration error metrics, such as ECE and MCE, are reduced significantly from 0.176 to 0.037 and from 0.276 to 0.111 through the probabilistic calibration, indicating a notable improvement in uncertainty estimation.