The first, a physics-informed Wasserstein generative adversarial network with gradient penalty (PIL-WGAN-GP), fills gaps in the data using self-supervised masking, solar geometry, a clear-sky-based power envelope, and deterministic projection. By anchoring the imputation process to the clear-sky power envelope, the module avoids physically inadmissible reconstructions. The second component, a dual-exploration spatiotemporal graph neural network (DEST-GNN), models the relationships among sites by fusing a static spatial prior with a time-varying dependency graph, allowing the model to represent cross-site dependencies as they change over time rather than treating them as fixed.

The third piece, an arbitrary-quantile recurrent decoder (AQ-RNN), predicts aggregate PV output quantiles directly by enforcing bounded positive spacings and monotone interpolation. This design ensures that the forecast for one quantile always remains below the forecast for the next higher quantile, eliminating quantile crossing by construction rather than correcting it after the fact. The researchers evaluated the system retrospectively using hourly PV generation records from 11 Belgian regional PV zones along with matching meteorological records in 2025.

For imputation, the physics-informed module reduced normalized mean absolute error substantially: for 3-, 6-, and 12-hour consecutive gaps, error fell from 1.450%, 2.473%, and 5.252% under linear interpolation to 0.869%, 1.027%, and 1.300% respectively. Hard projection eliminated nighttime leakage and upper-envelope violations in the reconstructed data. For forecasting, the full DEST-GNN-AQ model reduced the average daylight normalized mean absolute error across 1-, 6-, 12-, and 24-hour horizons from 3.689% for a no-graph version of the same decoder to 3.509%, with the main gains occurring at the longer 6-to-24-hour horizons.

The improvements are modest in absolute terms, but the results indicate that explicitly modeling the dynamic network of dependencies among sites provides measurable value at those horizons. On the probabilistic side, the model generated nested 50%, 70%, 80%, and 90% prediction intervals with a zero quantile-crossing rate. The authors are candid about the remaining limitation: the empirical coverages of these intervals came in at 41.72%, 62.81%, 69.84%, and 85.26%, meaning the intervals were structurally valid—properly ordered and non-overlapping—but fell short of their nominal confidence levels.

Calibration, they note, remains improvable. The broader significance lies in showing that physically admissible data reconstruction and probabilistic forecasting can be bridged within a single framework, with solar geometry and clear-sky limits enforced during imputation and the non-crossing guarantee encoded directly into the decoder architecture. The authors caution that the findings come from a retrospective case study on Belgian data, and the coverage shortfall indicates further refinement is needed.

Subject of Research: Physics-guided probabilistic forecasting of distributed photovoltaic clusters using physics-informed imputation and dynamic graph neural networks

Article Title: Physics-guided non-crossing probabilistic forecasting for distributed photovoltaic clusters using dynamic graph attention

Article References: Yan, Z., Xu, T., Shen, Z., Wei, W., Huang, Y., & Long, J. (2026). Physics-guided non-crossing probabilistic forecasting for distributed photovoltaic clusters using dynamic graph attention. Frontiers in Energy Research, 14, Article 1952891. https://doi.org/10.3389/fenrg.2026.1952891

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1952891

Keywords: photovoltaic power forecasting, probabilistic forecasting, graph neural networks, quantile crossing, physics-informed imputation, prediction intervals, distributed PV clusters, smart grids, Wasserstein GAN, spatiotemporal modeling, Physics-guided, non-crossing