CC BY-NC 4.0Lorenz, ElkePastewka, LarsStraub, NilsNilsStraub2023-12-012023-12-012020https://publica.fraunhofer.de/handle/publica/457409https://doi.org/10.24406/publica-222610.24406/publica-2226Solar resource forecasting contributes to solving the challenges that come with the variable power production of solar power plants. All sky imager (ASI) based forecasting systems represent a relatively new class of very short-term forecasting systems which can, unlike conventional forecasting systems such as numerical weather prediction models or satellite image-based forecasting, resolve small-scale changes in cloud cover and thereby predict sudden changes in irradiance, usually referred to as ramps.These very short-term irradiation forecasts can contribute to an increased grid-stability, make management of electrical storage capacities more e cient, facilitate short-term trading and save fuel in o grid PV-Diesel hybrid systems. In this thesis a novel method to retrieve irradiance from sky images is presented, which can be used as part of an ASI irradiance forecasting system or to create maps of current irradiance. This method exploits image features, irradiance measurements from a network of eight measuring stations and cloud base heigh data to train a machine learning algorithm on making areal estimations of GHI. The combination of machine learning algorithms with spatially distributed irradiance measurements is a novel approach that has not yet been explored in previous work. Cloud shadows on the ground are computed by geolocating clouds in the image and projecting their shadows to their real world positions, which enables areal estimations of irradiance. The focus of this thesis is the development, optimization and testing of the irradiance modelling algorithm. Applying the shadow projection to the sky image local features such as pixel values around in the position of interest can be extracted. The machine learning algorithm is trained to learn the relationship between these local feateres as well as external features like cloud base height and sun position and the measured irradiance levels. After training of the algorithm has nished it can model irradiance in new situations. The final model is tested on two di erent image datasets: One consists of preselected single cloud layer images that promote an accurate shadow projection (machine learning dataset), the other one covers one entire month of sky images and therefore also comprises more complex situations (sample month). Both datasets are split into 75% training and 25% testing data respectively and the nal model is trained and tested on both datasets separately. Our model could outperform the irradiance algorithm by Dittmann et al. [1], which was used as well established reference model. It proves to be robust even in cloud situations where the cloud shadow mapping is inaccurate and is able to partly compensate these projection inaccuracies.enmaster thesis