High accuracy data-driven heliostat calibration and state prediction with pretrained deep neural networks

Abstract The efficiency of solar tower power plants depends strongly on the ability to reflect the sun light onto a defined point on the receiver. Due to the high demands on the heliostats to achieve high accuracy at low costs, a regular calibration is necessary to reduce the tracking error. In this paper a new method for improving existing calibration methods using deep learning is presented. The results are validated by using calibration data recorded at the Solar Tower Jolich with the example of one heliostat. Through a combination of Self-Normalizing Neural Networks and transfer learning it is possible to benefit from the advantages of neural networks already with a training dataset of only 300 measuring points. With that a measured test accuracy of 0.42 mrad was achieved. This was approximately three times more accurate than the best result of the compared state-of-the-art regression algorithm used in Jolich. Furthermore we give recommendations on the structure of the dataset and the neural network (NN) pretraining necessary for these results.