Climate-Invariant Conformal Prediction Intervals for Multi-Horizon Solar and Wind Forecasting

Reliable integration of renewable energy into power grids requires more than just accurate point forecasts; it demands robust uncertainty quantification. Existing probabilistic methods often struggle with finite-sample validity or require extensive per-site recalibration, making them difficult to deploy across geographically dispersed assets. This paper addresses these challenges by developing a unified, climate-invariant conformal prediction framework for multi-horizon solar and wind forecasting.

The authors introduce a two-stage pipeline. First, they employ a bootstrap-diverse XGBoost ensemble to generate both a point forecast and a measure of local dispersion (uncertainty). Second, they apply a conformal interval layer that refines these outputs into prediction intervals. This layer integrates three key mechanisms: heteroscedastic residual normalization (to adjust width based on local difficulty), asymmetric calibration (to handle skewed error distributions), and Mondrian group-conditional thresholds (to ensure validity across different operating regimes, such as time-of-day for solar or uncertainty-tertiles for wind). Crucially, the entire pipeline uses a single, fixed specification that is not retuned for different sites or forecast horizons.

The framework was evaluated using NASA POWER data across four climatologically distinct sites in both hemispheres. The results demonstrate that the method maintains near-nominal coverage for both solar irradiance and wind speed across all tested horizons (1 to 12 hours). By avoiding site-specific tuning, the authors show that the calibration and sharpness of the intervals are inherent properties of the method. The proposed approach outperformed competitive baselines, achieving up to a 35% reduction in the Interval Score, confirming its effectiveness for robust, large-scale renewable energy management.