Advect-Spectral Flow: Probabilistic Solar Irradiance Nowcasting with Advection-Aligned Dynamics and Stochastic Spectral Budgets

Regional solar irradiance nowcasting requires uncertainty estimates for both radiation magnitude and spatial structure. Advect-Spectral Flow separates stochastic regional mean and multiscale spectral budgets from spatial shape generation in advection-aligned coordinates. A matched comparison with an endpoint-constrained generator evaluates the effect of the complete generation scheme under shared budgets and equal-sized networks. Using hourly satellite fields over China, the prespecified 2014 in-domain test reduces 3 h empirical CRPS from 95.299 to 94.448 W m−2 (paired difference −0.852 W m−2; seven-day block-bootstrap 95% interval [−0.978, −0.725]). A frozen model exploratory evaluation on 1237 origins across six regions in a different 2023 satellite product retains overall Endpoint-relative CRPS reductions of 1.41%, 0.60%, and 0.26% at 1–3 h. The benefit is therefore modest and lead dependent. Static gating has lower 3 h CRPS. SolarSTEPS has lower external CRPS in every region and lead, and final-field spectral errors do not consistently favor Sphere. The framework and matched evaluations provide evidence on the relation between representation-level constraints and forecast quality, identifying the accuracy, structural fidelity, and computational trade-offs that matter when selecting constrained probabilistic nowcasting methods.