End-To-End Sensitivity Analysis of a Hybrid Heliostat Calibration Process Involving Artificial Neural Networks

Abstract

Methods of Artificial Intelligence (AI) have permeated research on solar tower power plants following general breakthroughs in the third AI summer. Whether in cloud forecasting, flux density prediction, or heliostat calibration, machine learning methods have shown to outperform state-of-the-art approaches and pave the way to process automatization. Especially neural networks have been on the rise for Concentrated Solar Power (CSP) applications. These novel approaches, while bearing potential for a more efficient power plant operation, also require different ways to assess their safety in critical infrastructure such as solar towers, where high temperatures could damage valuable material. The black box nature of many AI algorithms begs the question of how they will behave on unseen or noisy data and whether certain robustness properties of these algorithms can be guaranteed. So far, this has not been studied in the context of CSP. In this paper, we present a method for end-to-end sensitivity analysis examined on a neural network developed for heliostat calibration. We make assumptions of certain measurement noise and conclude with an estimate of how much that noise can impact the prediction of the overall algorithm. For this, we employed formal neural network verification and coupled it with interval arithmetic to deduct bounds in preprocessing, neural network, and geometric model calculations. Our analysis shows that errors induced by noisy heliostat state information can be guaranteed to be below 0.02 mrad, which is an irrelevant magnitude for heliostat calibration. We can provide this guarantee for points within and outside the training dataset, for networks trained on as little as 30 datapoints. This deepens our understanding of AI-driven heliostat calibration and motivates its adoption in an industrial context. We anticipate our paper to spur further research concerning the safety of promising AI applications in the field of CSP, as they are being developed continuously.