In-Situ Solar Tower Power Plant Optimization by Differentiable Raytracing

Abstract

Solar tower power plants deliver climate-neutral electricity and process heat and can play a key role to facilitate the ongoing energy transition. These plants reflect sunlight with thousands of mirrors (heliostats) to a receiver and can generate temperatures over 1000°C. In practice, a plant must be operated with safety margins as even small surface deformations and heliostat misalignments can locally lead to dangerous temperature peaks. These imperfections are difficult to assess and limit the plant's efficiency, which hinders commercial success in a competitive market. We present a computational technique that predicts the incident power distribution of each heliostat including the inaccuracies based solely on focal spot images that are already acquired in most solar power plants. The method combines differentiable ray tracing with a smooth parametric description of the heliostat and reconstructs flawed mirror surfaces with sub-millimeter precision. Applied at the solar tower plant in Jülich, our approach outperforms all alternatives in accuracy and reliability. The approach can be integrated into the existing infrastructure and plant control at low cost, leading to increased efficiency of existing and decreased expenses for future power plants and supports establishing a new, green energy technology. For other fields, our approach can be a blueprint. We implement a common simulation technique in the Machine Learning framework PyTorch, leveraging automatic differentiation and GPU computation. By combining gradient-based optimization methods and a tunable parametric heliostat model, we overcome the high data requirements of data-centric methods while at the same time maintaining the flexibility required for modeling a complex real-world system.