Quantile calibrated ensemble deep learning framework for daily global horizontal irradiance forecasting

This study presents a chronology-preserving framework for daily global horizontal irradiance (GHI) forecasting over central India (21.78°N–27.23°N, 73.96°E–83.07°E) using national aeronautics and space administration (NASA) prediction of worldwide energy resource (POWER) reanalysis for 2020–2025. The pipeline uses a combination of targeted feature engineering, such as lagged persistence, multiwindow rolling statistics, sinusoidal seasonality encodings and clear-sky residuals, with heterogeneous base learners, such as Random Forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM) and short-term memory (LSTM). These models are aggregated by simple averaging, validation-weighted linear blending and a stacking ensemble whose meta-learner is a ridge regularized linear model trained on out-of-fold predictions to avoid information leakage. Probabilistic forecasts are computed with the help of quantile regression and are calibrated by conformal prediction to compute statistically valid prediction intervals. The proposed model gives the best results, with mean absolute error (MAE) = 0.085 W m−2, root mean square error (RMSE) = 0.110 W m−2, and coefficient of determination (R²) = 0.94, outperforming the results of the best single learner RF as well as the persistence and climatology baselines (MAE = 0.115 and 0.124). Conformal calibration enhances 90% prediction interval coverage probability (PICP) from 0.7713 to 0.8703, where interval sharpness is preserved as the mean prediction interval width (MPIW) ≈ 0.8790. Residual and binned-error diagnostics demonstrate that the stacking ensemble provides tighter and more symmetric residuals, lower variance and stable performance in monsoon and winter regimes, making it a robust and operationally interpretable model for grid-aware solar forecasting.