Towards more reliable flux density prediction using uncertainty quantification of neural networks

Solar Energy, vol. 300, pp. 113739

Abstract

Solar tower power plants represent a renewable source of dispatchable energy. By concentrating incoming sunlight via tens of thousands of heliostats, they render large amounts of energy in the form of heat that is usable as a source of electricity or in an industrial context. In this work, we propose an uncertainty-aware prediction framework using Monte-Carlo Dropout and a random forest classifier to flag unreliable flux predictions of said heliostats. Precise knowledge about the shape and position of the heliostat’s reflections is crucial to an efficient and safe operation of the plant. As their facets exhibit surface imperfections and the reflections span distances of hundreds of meters, the flux density distributions generated by a heliostat’s reflection on the receiver differ greatly from an assumed ideal. Occlusion caused by shading or blocking, rare misalignment scenarios or extreme flux distributions represent edge cases for flux density prediction. Therefore, each heliostat bears an individual behavior concerning the produced flux density distribution. To predict this individual behavior, researchers have employed data-driven methods showing an improvement over state-of-the-art approaches by allowing a seamless integration into the calibration process of the camera-target method using the same calibration images as training data. Their approach shows strong performance, with an error analysis yielding an estimate of 4.5%, where the neural network contributes 1.5%. However, the analysis uses the network’s mean training error as a component, leaving edge cases unanswered. As neural networks are uninterpretable models, their integration into safety-critical operations requires thorough uncertainty analysis. In our work, our aim is to equip the data-driven approach with a layer of security by reasoning about network uncertainties, resulting in an algorithm that filters out certain samples to increase the predictive qualities of the overall process. • Flux Density Neural Nets encode more data than they use for prediction. • Uncertainty metrics explain faults, yielding up to 80% F1 scores in fault detection. • The fault detector adapts to new tasks and performs better at higher fault rates.

Authors 3

1. Leon Tim Engelbert Sievers corresponding

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

   Affiliation as printed

   DLR, Institute of Solar Research, Linder Höhe, 51147 Köln, Germany

2. Daniel Maldonado Quinto

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

   Affiliation as printed

   DLR, Institute of Solar Research, Linder Höhe, 51147 Köln, Germany

3. Bernhard Hoffschmidt Aachen Chair of Solar Technology

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

   Affiliation as printed

   DLR, Institute of Solar Research, Linder Höhe, 51147 Köln, Germany

   RWTH Aachen University, Chair of Solar Technology, Linder Höhe, 51147 Köln, Germany