Differentiable aimpoint control for CSP towers using learned optical models: large-scale live demonstration at the Solar Tower Jülich

Abstract

The goal of this work is aimpoint control that is real-time capable and robust in plant operation, and that rests on optical models derived automatically from the beam characterization images a solar tower already produces during routine heliostat calibration. Two coupled bottlenecks have stood in the way. Optical models are either too simplified to represent the true focal spot of an individual heliostat, or they rest on dedicated metrology such as deflectometry that does not scale to a full field. Optimizers, in turn, are too slow for real-time operation and therefore resort to heliostat grouping or to a discretized set of candidate aimpoints. We present an aimpoint control framework that addresses both. Per-heliostat focal spots are predicted by a transformer model trained exclusively on standard camera–target calibration images, so that the optical characterization of the field is obtained from data the plant already produces, without reconstructing mirror surfaces or fitting per-heliostat parameters. Heliostat characterization and aimpoint control thereby become a single automated chain: calibration images enter, aimpoints leave, and no manual step remains between them. This optical model is embedded in a fully differentiable chain from continuous aimpoint coordinates to the receiver flux distribution, so that a coordinated update for all aimpoints is obtained from a single gradient evaluation, without grouping and without discretizing the search space. The flux distribution is evaluated from the current optical model at every iteration rather than looked up in precomputed flux maps, so a change in the heliostat models propagates into the control immediately and no precomputation stage has to be repeated. In a simulation study for 1988 heliostats of the Solar Tower J¨ulich with an open volumetric receiver, the optimized aiming reduces the peak flux density by 24% relative to the aiming configuration used in plant operation while simultaneously raising the intercept factor from 0.949 to 0.959 — a combination that neither the operational reference nor an analytical Vant-Hull strategy achieves. A multi-objective formulation over nested subfields additionally makes one solution valid across the full range of field occupancy, so that the number of on-receiver heliostats can be varied between re-optimizations, which together with the slow drift of the solution against the sun position permits a re-optimization interval of 10 minutes. The controller is integrated into the plant control system, satisfies the real-time requirements of operation, and has been used on-sun over fourteen operating days with up to 1130 heliostats, both by plant operators and within the autonomous plant control loop.

Authors 5

1. Mathias Kuhl

   Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR)

   Affiliation as printed

   Deutsches Zentrum für Luft- und Raumfahrt e.V. (Germany)

2. Inga Miadowicz

   Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR)

   Affiliation as printed

   Deutsches Zentrum für Luft- und Raumfahrt e.V. (Germany)

3. Peter Schwarzbözl

   Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR)

   Affiliation as printed

   Deutsches Zentrum für Luft- und Raumfahrt e.V. (Germany)

4. Daniel Maldonado Quinto

   Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR)

   Affiliation as printed

   Deutsches Zentrum für Luft- und Raumfahrt e.V. (Germany)

5. Robert Pitz‐Paal Aachen

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

   Affiliation as printed

   Deutsches Zentrum für Luft- und Raumfahrt e.V. (Germany)

   RWTH Aachen Univ. (Germany)