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

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