Convolutional neural network surrogate model for real-time flux density prediction in cylindrical solar thermal receivers

Concentrating solar technologies (CST) have shown a wide potential not only for power generation at highly-irradiated locations, but also for thermochemical processes and industrial heat production. In particular, solar power tower (SPT) systems are especially effective for high temperature applications. However, this plant architecture entails severe challenges in operation and maintenance, such as accurate and real-time flux prediction. In order to address these issues, this work presents a novel surrogate model for flux density prediction in cylindrical receivers irradiated by multi-facet heliostats, based on convolutional neural networks (CNN). The study shows the assessment of four different CNN architectures, using as an example the heliostats and receiver of Dunhuang 10 MW e power plant. The best models have shown the ability to learn and predict the effect of sun position and heliostat location on the resulting flux with two short datasets composed of only 311 and 289 flux maps, respectively for time and location effects. Inference has been assessed under 7 time conditions –for a single heliostat– and 10 heliostat positions –at a single timestep– not included in the datasets. The results show solid accuracies above 90%, structural similarity indexes (SSIM) up to 98% and peak signal-to-noise ratio (PSNR) metrics over 30 dB, which proves robustness. These predictions are now obtained in real-time, which boosts large-scale applicability of flux density nowcasting and plant design possibilities.