Physics-guided non-crossing probabilistic forecasting for distributed photovoltaic clusters using dynamic graph attention
TL;DR
A physics-guided framework that integrates a physics-informed Wasserstein generative adversarial network with gradient penalty, a dual-exploration spatiotemporal graph neural network, and an arbitrary-quantile recurrent decoder to bridge physically admissible data reconstruction with reliable, non-crossing probabilistic forecasting for PV clusters is evaluated.
Abstract
Accurate probabilistic forecasting of geographically distributed photovoltaic (PV) generation is challenged by incomplete measurements, weather-dependent cross-site correlations, and quantile crossing. This study proposes a physics-guided framework that integrates a physics-informed Wasserstein generative adversarial network with gradient penalty (PIL-WGAN-GP), a dual-exploration spatiotemporal graph neural network (DEST-GNN), and an arbitrary-quantile recurrent decoder (AQ-RNN). The imputation module combines self-supervised masking, solar geometry, a clear-sky-based power envelope, and deterministic projection. The forecasting module fuses a static spatial prior with a time-varying dependency graph, while the decoder predicts aggregate PV quantiles directly through bounded positive spacings and monotone interpolation. The framework was evaluated in a retrospective case study using hourly PV generation records from 11 Belgian regional PV zones and matched meteorological records in 2025. For 3-, 6-, and 12-h consecutive gaps, PIL-WGAN-GP reduced normalized mean absolute error from 1.450%, 2.473%, and 5.252% for linear interpolation to 0.869%, 1.027%, and 1.300%, respectively. Hard projection eliminated nighttime leakage and upper-envelope violations. Across 1-, 6-, 12-, and 24-h horizons, DEST-GNN-AQ reduced the average daylight nMAE from 3.689% for the no-graph AQ-RNN to 3.509%, with the main gains occurring at 6–24 h. The model generated nested 50%, 70%, 80%, and 90% prediction intervals with a zero quantile-crossing rate. Their empirical coverages were 41.72%, 62.81%, 69.84%, and 85.26%, indicating that structural validity was achieved while calibration remains improvable. Collectively, these findings validate the proposed framework’s ability to bridge physically admissible data reconstruction with reliable, non-crossing probabilistic forecasting for PV clusters.