SC-CP: spatially-correlated conformal prediction for GNN-based distributed photovoltaic power forecasting
Abstract
High-penetration distributed photovoltaic (DPV) generation exposes distribution grids to rapid, spatially-coupled power swings under passing cloud fronts, so secure dispatch of reserves, storage, and voltage-control devices hinges on prediction intervals with trustworthy coverage, not point forecasts. Two limitations of current practice are particularly costly in this physical setting: graph neural network (GNN) forecasters reproduce inter-site spatial coupling but emit only point estimates, leaving operators without a quantified risk margin; classical conformal prediction (CP) restores a marginal coverage guarantee yet calibrates each site independently, so a single cloud passage that drives correlated errors across neighbouring DPV plants collapses the joint guarantee exactly when reserve and voltage margins are tightest. We propose Spatially-Correlated Conformal Prediction (SC-CP), which reuses the adjacency matrix already learned by the GNN backbone to weight nonconformity scores by inter-site error coupling, so that meteorologically co-driven errors inflate the interval while isolated anomalies do not. As an integral component of SC-CP, a Site-Adaptive Coverage (SAC) rule reallocates the global miscoverage budget non-uniformly across feeders in proportion to their calibration difficulty. Experiments on five NREL stations in northern California confirm that SC-CP restores nominal coverage in spatially-correlated regimes where standard CP systematically undercovers, and a real heterogeneous validation on the four German TSO control zones from the Open Power System Data release shows that SC-CP reduces the per-zone PICP spread of Standard CP by roughly 40%, with SAC supplying the expected non-uniform per-zone budget on the harder zones.