Solar power tower systems require optimal heliostat aiming strategies to maximize energy capture while ensuring receiver safety under dynamic environmental conditions. Conventional approaches relying on analytic optical models and heuristic methods are limited in computational efficiency and flexibility in handling design constraints for a large-scale heliostat field. This study applies machine learning techniques for developing a fast receiver flux prediction model that is designed for use in a model predictive control (MPC) framework to enhance operational efficiency and safety of the central receiver under cloud shading. A novel two-stage machine learning approach is proposed that uses a conditional variational autoencoder (CVAE) for efficient flux data compression, followed by an augmented neural network for rapid flux prediction (with peak flux errors of 2.91%) under varying DNI levels, sun positions, and cloud patterns. The MPC framework, based on the flux prediction model and with receiver thermal safety constraints enforced, facilitates real-time closed-loop optimization of aiming strategies under environmental variations. Performance assessments using simulated data generated by SolarPILOT for the Crescent Dunes solar power tower plant confirm the model’s robustness under various operational scenarios. The integrated MPC framework effectively balances power maximization and receiver safety constraints, and shows a significant performance improvement over static control methods, with a 10.8% higher power tracking accuracy, a 6.4% increase in power output, an allowable flux density (AFD) violation rate of only 0.91%, and a 22.3% enhancement in flux uniformity.

原文DOI：https://doi.org/10.1016/j.solener.2025.114258