A Physics-Informed Diffusion Model for Surface Solar Irradiance Temporal Downscaling

The integration of solar photovoltaics at scale requires high-frequency irradiance data to quantify minute-scale variability for grid integration, storage sizing, and ramp-rate management. However, satellite-based solar irradiance estimates face a fundamental scale mismatch. Ground-based pyranometers provide accurate high-frequency measurements but are spatially sparse and observe point-scale transients that differ from the area-averaged irradiance a utility-scale plant experiences. Satellite products such as the Copernicus Atmosphere Monitoring Service (CAMS) offer broad spatial coverage but lack temporal resolution (15 min), smoothing away the high-frequency fluctuations driven by cloud passage. Because many distinct sub-pixel cloud configurations can produce the same coarse-resolution mean, recovering minute-scale temporal trajectories from satellite inputs is inherently ill-posed and requires probabilistic treatment.

To address this ambiguity, we present a physics-informed generative framework. We derive a temporal degradation operator grounded in Taylor's frozen turbulence hypothesis, which recasts the spatial averaging of a satellite footprint as a temporal Gaussian convolution — a point spread function (PSF) representing the spatial response of the sensor. To invert this operator, we train a conditional denoising diffusion probabilistic model (DDPM) that generates ensembles of minute-resolution irradiance trajectories conditioned on coarse satellite inputs. A spectral regularization term encourages the generated sequences to follow the power-law energy cascade observed in cloud-driven irradiance fluctuations, preserving realistic high-frequency structure while suppressing spurious artifacts.

The framework is evaluated through a sim-to-real domain transfer experiment: the model is trained on synthetic pairs built from BSRN measurements using the physical forward operator and then applied directly to operational CAMS data, isolating and quantifying the representativity gap between idealized and real-world degradation. The added value of the proposed approach is evaluated for selected operational applications. Results show that the generated sequences reproduce high-frequency variability consistent with ground observations, including the capture of significant ramp events, recovery of curtailed irradiance peaks relevant to inverter clipping estimation, and improved characterization of storage requirements relative to temporally smoothed operational inputs.