Day-Ahead PV Power Prediction Intervals Based on Physical Sensitivity Analysis and Robust Error Decomposition

As photovoltaic (PV) penetration grows, scheduling depends on a firm measure of irreducible forecast uncertainty. We present a hybrid physics–statistics framework that computes a theoretical predictability limit conditioned on numerical weather prediction (NWP) uncertainty via error propagation. A deterministic irradiance to power chain using solar geometry, Perez transposition, Sandia Array Performance Model (SAPM) cell temperature, and NREL PVWatts provides sensitivity gradients. Ridge debiasing and a minimum covariance determinant (MCD) estimator isolate the stochastic part of NWP errors and yield a time-varying covariance. At Desert Rock in Nevada with the European Centre for Medium-Range Weather Forecasts high-resolution forecast (ECMWF HRES), realised errors of the evaluated forecasting models consistently exceed the limit. Optimisation potential (OP) indicates actionable headroom, with gated recurrent unit (GRU) (Formula presented.) on clear days and multilayer perceptron (MLP) (Formula presented.) on non clear days, showing proximity to the ceiling when weather is variable. The theoretical (Formula presented.) interval is always nested within model (Formula presented.) intervals. The limit offers an objective benchmark for reserve sizing, commitment, and deciding when extra algorithmic effort is worthwhile.