Inverse deep learning raytracing for scalable heliostat surface prediction

RWTH Publications (RWTH Aachen)

Abstract

Concentrating solar power (CSP) is a key technology for the energy transition due to its ability to deliver dispatchable electricity and high-temperature heat for thermochemical processes. In a tower system, tens of thousands of heliostats are continuously aligned to concentrate sunlight onto a central receiver. For autonomous, safe, and efficient operation, the spatial flux density distribution on the receiver is critical: local peaks must remain below material limits to prevent damage. However, single heliostat focal spots are irregular and vary across heliostats because of surface errors, which necessitates integrating surface characterization and flux prediction into plant control to reduce safety margins and increase receiver efficiency. Classical workflows–manual surface measurements followed by raytracing–are costly, error-prone, and do not scale to commercial field sizes. This thesis introduces inverse Deep Learning Raytracing (iDLR), a data-driven approach that infers a heliostat’s surface directly from target images of its focal spot, a data source readily available during routine operation. Validated on real measurements from the Solar Tower Jülich (STJ), iDLR reconstructs the majority of the heliostat surface with very high similarity to deflectometry, the state of the art for heliostat surface measurement, achieving a median Structural Similarity Index Measure (SSIM) of 0.77 with an Inter-Quartile-Range (IQR) of 0.32 and a median mean absolute error (MAE) of 0.17 mm with IQR of 0.08 mm and when comparing normal vectors a Root Mean Square Error (RMSE) of 0.65 mrad±0.18 mrad. However, due to the ill-posed nature of the inverse problem, degraded predictions can occur in approximately 13% to 20% of the predictions. Additionally, to enable compact representations, a Non-Uniform Rational B-spline (NURBS) model is deployed as a stand alone surface parameterization, reducing parameter count by 99.97% and memory by 99.91% relative to normal-vector fields, with only minimal parameterization error (mean angular deviation of 0.09 mrad±0.03 mrad). When iDLR surface predictions are integrated in a raytracing environment, the resulting single heliostat flux density predictions reach an accuracy of 89% ± 5% compared to deflectometry-based predictions. For the superposed flux density distribution on the receiver, prediction accuracy ranges from 96.5% ± 2.5%to 98.9% ± 2.2%. Power-plant-control-level analyses indicate that integrating iDLR into plant control can yield efficiency gains of 3%–8%. In summary, this work provides a scalable and precise methodology for heliostat surface- and flux density characterization using only data available from routine power plant operation, thereby coming–for inference–at nearly zero additional cost. Incorporating iDLR into standard characterization routines and power plant control can improve CSP plant performance and support the broader deployment of the technology.

Authors 1

1. Jan Pascal Lewen corresponding Aachen

   RWTH Aachen University · Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR)

   Affiliation as printed

   DLR

   RWTH Aachen