Utilizing Neural Networks for Generalizing Heliostat Calibration in Solar Tower Plants

Abstract

Heliostats play a crucial role in solar tower plants, where they reflect sunlight onto a central receiver to generate thermal energy. Ensuring precise heliostat alignment is essential for maximizing energy capture and system efficiency. However, traditional calibration methods, such as rigid body models, suffer from limited accuracy due to mechanical wear, environmental changes, and simplifying assumptions. Neural networks have shown promise in heliostat calibration by learning complex relationships in tracking data, but their application has been limited by data requirements and a lack of generalization across heliostats. This thesis introduces Helioformer, a Transformer-based neural network designed for heliostat calibration, capable of predicting both normal vectors and direct motor positions while maintaining high accuracy. The model was trained and evaluated using datasets from the Solar Tower Jülich (STJ), which is operated by the German Aerospace Center (DLR). The dataset includes heliostat positions, motor positions, and Sun positions. Results demonstrate that Helioformer significantly outperforms traditional rigid body models, reducing aimpoint offsets from 2.05 mrad (rigid body model) to 1.45 mrad when predicting normal vectors, and achieving even better accuracy of 1.31 mrad when predicting motor positions. In addition, this work introduces a method for comparing the accuracy of the motor position using more tangible aimpoint offsets, making the predictions of the motor position more interpretable. These findings highlight the potential of deep learning to enhance heliostat tracking precision, reduce recalibration efforts, and improve the overall efficiency of solar tower plants.