Abstract
Concentrating Solar Power (CSP) plants rely on thousands of heliostats to focus sunlight onto a central receiver. Precise characterization of mirror surface deformations and alignment errors is critical for performance but traditionally labor-intensive. This paper presents inverse Deep Learning Raytracing (iDLR), a novel pipeline that reconstructs 3D surface profiles directly from standard calibration images. By bridging the "Sim-to-Real" gap through domain adaptation and randomization, the method achieves a median surface error of 0.17 mm in real-world tests at the Solar Tower Jülich, outperforming standard "ideal mirror" assumptions by 26%.
Introduction
In CSP plants, efficiency is dictated by the precision of solar flux distribution on the receiver. Real-world heliostats deviate from ideal geometries due to "canting" (misalignment of facets) and surface "waviness." While high-precision deflectometry can measure these imperfections, its cost prohibits plant-wide deployment. Consequently, most control systems rely on idealized flat or parabolic models, leading to energy loss and potential receiver damage from "hot spots." The challenge lies in characterizing these surfaces automatically using routine operational data.
Methods: The iDLR Pipeline
The proposed iDLR pipeline processes "target images"—photographs of light spots cast on a calibration target—to infer the underlying mirror geometry.
- Domain Adaptation (UNet): To handle real-world noise, background radiation, and exposure variations, a UNet translates raw photos into clean "flux density maps" consistent with idealized simulations.
- The iDLR Model: The core engine employs a Vision Transformer (ViT) encoder to process flux maps and a StyleGAN2 decoder to generate the corresponding 3D surface profile.
- Domain Randomization: Training data is injected with physical phenomena such as mirror soiling, sunshape variations, and camera artifacts to ensure robustness when transitioning from simulation to reality.
Datasets and Validation
The model was trained on a "semi-artificial" dataset derived from 393 real deflectometry measurements at the Solar Tower Jülich (STJ). Data augmentation techniques, including rotation and weighted averaging, were used to generate approximately 400,000 synthetic surfaces. Validation was conducted on 63 real-world heliostats at the STJ facility under actual operational conditions. High-resolution deflectometry served as the ground truth for accuracy assessment.
Key Results and Insights
- Surface and Flux Accuracy: The model achieved a median Mean Absolute Error (MAE) of 0.17 mm on mirror surfaces. Using these predicted surfaces in raytracing simulations yielded 90% accuracy in predicting actual light distribution.
- Sim-to-Real Efficacy: Bridging the simulation-reality gap proved essential; without domain adaptation and randomization, error rates tripled.
- Distance-Dependent Performance: Accuracy improved with distance from the tower, as alignment errors (canting) become more distinct from surface deformations at greater ranges.
- Robustness: The model demonstrated high accuracy in "double-extrapolation" scenarios, predicting flux for unseen sun positions and receiver geometries, indicating it has learned fundamental optical physics.
Conclusion and Future Work
The iDLR method provides a scalable path toward creating "digital twins" for entire heliostat fields. While the inverse problem remains partially underdetermined—leading to a 16% misprediction rate where different surfaces produce similar spots—the overall 26% improvement in flux prediction accuracy is significant. Future work aims to close the remaining performance gap and integrate these models into autonomous flux control and predictive maintenance systems for next-generation solar plants.
Summary updated