Concentrating solar power (CSP) plant data-driven digital twin: A novel method for flux density prediction

• A universal open-source methodology for a digital twin is proposed in this work. • Near real-time flux density monitoring is achieved, with latencies below 30 s. • A synthetic dataset for flux predictions’ self-corrections is presented. • A U-Net based self-correction module is presented. • Accuracies up to 95 % and biases below 7.5 % are reached through this approach. Concentrating solar power (CSP) is a promising renewable energy source because of its potential contribution to power generation and fuel synthesis (among others). However, several constraints continue to limit the economic and environmental attractiveness of this technology. One of the most important issues to address is measuring receiver efficiency, yet current methods are disruptive, expensive and complex. To overcome this, the present work introduces a data-driven digital-twin model for solar flux density predictions in central receiver systems (solar power towers). It consists of a real-time data agent that integrates signals collected at a CSP plant into flux density prediction tools. This flux prediction is coupled with a self-correction module based on graph neural networks (attention-gated U-Net) designed to replace simulations’ mathematical with deep-learning algorithms trained on realistic flux density measurements. The outcome is a semi-autonomous cyber-physical model that achieves latencies below 30 s for common CSP operations and flux density characterization accuracies up to 95 %, which makes it readily scalable, industrially applicable and more accurate than SOTA models. Its industrial relevance is boosted by a comprehensive training routine based on several heliostats’ superposed fluxes at realistic conditions.