Verification of the process reliability and scalability of an airborne measuring system for the heliostat orientation in solar power tower plants

The efficiency of a solar tower power plant (STPP) is oticeably impaired by incorrect orientation of the heliostats. An error in tracking the heliostat orientations causes a partial loss of incident irradiance during reflection. This reduces the yield of an STPP. The conventional flux density measurement (FDM) method, where the heliostat orientations are calibrated by sequentially pointing the heliostats onto a target, is a very slow process. To date, in fact, no commercially available solution for orientation monitoring of STPPs has been established that is both fast and accurate, especially for large STPPs. Therefore, German Aerospace Center (Deutsches Zentrum für Luft- und Raumfahrt e.V.) (DLR) has developed “QFly”, a measuring system for quality control of utility scale solar fields1 (QFly), which utilizes an unmanned aerial vehicle (UAV) to measure Parabolic Trough Power Plants (PTPP) within a few hours. DLR has extended the functionality of QFly for the application on STPPs, which is still to be fully approved. The goal of this thesis is to verify the measurement accuracy and scalability of QFly for STPPs. Measurement campaigns including QFly and the reference measurement FDM took place on the CESA-12 (CESA-1) at the Plataforma Solar de Almería3 (PSA), and at the STPP in Jülich, Germany4 (Jülich). Based on the measured values, different models for the quality assessment and the performance of QFly are created and finally compared with the reference data. \nThe first approach is to analyze an early intermediate result of QFly, called “QFly Zero Assumption”5 (ZA), which gives a rough estimation of the orientation of the heliostats. If a \ncertain heliostat is visible on an image, in a best-case-scenario an orientation is calculated for this heliostat, judging by its outer corners. On average 69 (CESA-1-2018) or 207 (CESA-1-2020) ZA orientations per heliostat are calculated. For the CESA-1-2020 data-set, the normal vectors are deviating with a standard deviation of 34 mrad around the respective heliostat’s mean normal vector, which is quite a lot. In this thesis a novel parameter-based optimization process was \ndeveloped, which filters the calculated alignments, independently from the reference result. In the context of the data basis, parameters like the distance of the heliostat to the camera orthe angle, which is calculated from the heliostat mirror normal vector and the connecting line between camera and heliostat center (αcam2Helio3D), which can also be viewed in a differentiated way when projected onto the XY plane, play a decisive role for the optimization. It was shown \nin the case of CESA-1-2020, that for αcam2Helio3D = [58‘,70‘] an improvement of 38.2% of the deviation from the reference measurement is achieved based on an unfiltered data set. The mean deviation between FDM and ZA for all available FDM heliostats is 8.9 mrad (standard deviance \nof 7.0 mrad), and therefore lies within the required error spectrum of 10 mrad compared to the reference measurement. The ZA results are thus as accurate as the PG result, which has a mean deviation in normal vector of 8.9 mrad for the same FDM heliostats. The process scalability of QFly can be tremendously improved by aiming at the photogrammetric calculation at the end of the QFly process in “Aicon 3D Studio”6 (Aicon). In fact over 70% of the QFly processing duration is spent with the QFly photogrammetry in Aicon (PG) in Aicon. The currently forbidding long processing duration could be improved by reducing the number of images given into Aicon as an input. To this end, a program has been written to transfer only high-quality images and detection data to the photogrammetric software, while maintaining the information content and interconnectivity of the images. This is done by eliminating redundant images and detections of features. Redundancy in images is present, the more \nsimilar the camera position, camera viewing direction and the detected features are. In addition, image-specific parameters are developed, such as the fill level and coverage of detections per image. Another important aspect is the filtering of detections based on expected weak results – this can be defined by a ZA calculation with all detected features instead of the outer corners. Further improvement is done by reducing the quantity of detections per feature to a certain fixed \nvalue, while deciding to keep those detections, for which the camera positions are most distinct to each other. It could be approved, that filtering images having similar EOR (<0.5m), similar viewing direction (<0.5°) and similar detections (>90%) calculate 8% faster than filtering out the same quantity of images randomly. The accuracy of the measurement regarding the heliostat orientation was not scope of the investigations.