A data-driven digital twin approach for real-time flux density prediction in solar power towers using ray-level deep learning corrections

• Ray directions compressed 99.99% enabling near-real-time inference at 4.6 ms • Power and direction CNN corrections validated at 80.9% cavity flux accuracy • 5-fold cross-validation confirms robustness with fold std below ±0.005 • 13 × 13 grid resolution verified as optimal via Kolmogorov-Smirnov sensitivity • Aggregated training across 931 conditions generalizes without heliostat calibration Concentrating solar technologies offer significant potential for sustainable and dispatchable heat, power and fuel production, especially through solar power tower systems. However, accurate flux density measurement at the receiver remains a critical barrier to the commercial deployment of this technology. Current direct and indirect methods are disruptive, expensive and inapplicable to cavity receivers. In order to overcome these challenges, this work proposes a computational data-driven method that couples four convolutional neural networks operating sequentially with a Monte Carlo ray-tracing simulator to correct both the power and directional components of simulated rays at the receiver aperture plane. Ray direction data are compressed from over 100 × 10 6 parameters to 2028 values through statistical distribution fitting, enabling near-real-time inference on standard computing hardware. The power correction achieves accuracies up to 95.0% at the aperture plane, while direction predictions exceed 80.0% Pearson correlation for the three ray direction components across 63 diverse meteorological conditions spanning direct normal irradiance values from 100.0 to 900.0 W/m². The corrected rays are subsequently projected into the three-dimensional cavity receiver geometry, yielding an integrated flux prediction accuracy of 80.9% compared to experimental camera-based measurements at the Solarturm Jülich facility. An aggregated training strategy across 931 meteorological conditions and multiple heliostat configurations ensures generalization without cost-extensive per-heliostat calibration. These results demonstrate that ray-level deep learning correction provides a universal, non-disruptive, and industrially scalable solution for flux density prediction in both open and cavity solar receivers.