Spatio-Temporal Copula Modeling for Aggregated Day-ahead Photovoltaic Power Forecasting

Abstract

The increasing integration of photovoltaic (PV) power and its inherent variability require reliable aggregated probabilistic forecasts to support grid operation and market participation. However, most existing approaches focus on deterministic or site-level forecasting, with limited attention to coherent system-level uncertainty that captures spatial and temporal dependencies. This paper proposes a multistage framework for day-ahead probabilistic forecasting of aggregated PV power, combining physics-guided modeling and copula-based dependency construction. Power generation of 175 PV systems in Utrecht, the Netherlands, is clustered by physical similarity, and a physical model chain converts numerical weather predictions into power estimates. Cluster-level residuals are modeled probabilistically to generate quantile forecasts, which are aggregated using Gaussian Copulas to represent spatial dependence and enforce temporal coherence in day-ahead trajectories. Results indicate that physics-guided quantile model improves probabilistic accuracy compared to the pure quantile approach. The proposed framework provides reliable and coherent day-ahead scenarios, reduces extreme ramp behavior and physical boundary violations, while providing a scalable and operationally realistic methodology for aggregated PV probabilistic forecasting in hierarchical settings.