Enhancing Heliostat Calibration in Solar Tower Power Plants: A Novel Dataset Evaluation Metric and Hybrid Kinematic Modelling Technique

Abstract

The efficiency of heliostat fields is mostly affected by the heliostat’s alignment accuracies. A novel approach for optimizing these accuracies is introduced within this thesis, that includes dynamic heliostat behavior and reduces the amount of required training data. The current state of the art approach of using a rigid-body kinematic model is extended by exchanging the model’s geometry parameters by a neural network. Using this method allows the model to account for various dynamic impacts, such as structural bending. By excluding most of the ideal kinematic behavior from the neural network, the problem-complexity is reduced to computing deviations from the ideal behavior and thus applying simpler neural network architectures. In constrast to current neural network approaches, no pretraining on modelled data must thus be applied, which reduces the required amount of training data. The combination of using a pre-optimized kinematic model with dynamic geometry parameter adaptation by a neural network exceeds the current industry standards accuracies in more than 95% of the given test cases. By introducing a distance-metric between calibration data points’ solar positions, a new method for evaluating a dataset’s coverage of the entire heliostat behavior is achieved. Following this principle, an algorithm for splitting datasets into training-, validation- and testdata is introduced, that heuristically optimizes the data point distribution and heliostat behavior coverage. By means of three datasets that were collected at the German Aerospace Center (DLR)’s solar tower in Jülich between May 2021 and October 2022, the introduced metrics benefits to model training are verified. For two out of three datasets the trained model achieves average heliostat alignment accuracies below 1mrad for the entire year starting from as little as 20 training- and 20 validation data points. The obtained scientific insights in future can be applied to improve the evaluation of heliostat alignment dataset distribution as well as the performance of different heliostat models.