Climate-informed residual learning and conformal calibration for probabilistic photovoltaic power forecasting using open reanalysis data

This paper presents a climate-informed probabilistic benchmark for day-ahead forecasting of aggregated photovoltaic generation using open electricity data and ERA5 reanalysis. The proposed method uses a residual formulation in which reanalysis variables explain deviations from a 24 h persistence reference, rather than predicting photovoltaic power directly. XGBoost estimates the residual component from calendar variables, lagged photovoltaic generation and ERA5 covariates, while split conformal prediction provides prediction intervals. The framework is validated with German aggregated photovoltaic generation from Open Power System Data and ERA5 variables including surface solar radiation, cloud cover, temperature, dew point, wind, pressure and precipitation. Results show that the residual model reduces MAE from 1259.57 MW for persistence to 646.72 MW, representing a 48.66% improvement. Compared with XGBoost without ERA5 and direct ERA5-XGBoost, the MAE reductions are 27.27% and 16.40%, respectively. Daylight evaluation confirms that these gains are not caused by night-time zero-production samples. The conformal residual forecast reaches 87.68% empirical coverage for nominal 90% intervals, but daylight coverage decreases to 76.36%, revealing undercoverage during operationally relevant periods. These results show that reanalysis variables are more useful when they explain deviations from persistence, and that probabilistic photovoltaic forecasts should be evaluated through conditional reliability diagnostics rather than aggregate interval metrics alone.