Towards a neural network based flux density prediction – Using generative models to enhance CSP raytracing

Each solar tower power plant is designed for a pre-calculated optimal flux density distribution on the receiver. Any deviation from it has a direct impact on the component durability, energy efficiency and all downstream processes. An accurate knowledge of the current and predicted flux density is therefore essential. However, the existing measurements to obtain the flux density are either inaccurate, complicated or expensive. Moreover, to predict upcoming flux density maps, simulations are indispensable. Also, because there is still no cost-efficient way to measure heliostat surfaces on an industrial scale, these simulations neglect the individual surface errors of each heliostat and are therefore too inaccurate for a reliable prediction. We present a novel method based on artificial intelligence (AI) to include real heliostat surface errors into the flux density simulation, only by measuring the heliostats focal spot. The method needs only data generated by the already well-established camera-target heliostat calibration. Using the AI method as a supplement to this calibration, within this measurement, it is possible to simultaneously detect both main heliostat errors, the surface deformations and the misalignment. In this work, different neural network (NN) architectures are trained with artificially generated data and studied for their applicability to the described methodology. The network types used are conditioned and unconditioned generative adversarial networks (GANs) as well as neural radiance fields (NeRFs). The latter archives at best a Peak Signal to Noise Ratio (PSNR) of up to 27.8. Afterwards the network results are compared qualitatively in terms of image quality, controllability and dataset size.