Adaptive wavelet-CNN-LSTM forecasting with Bayesian optimization and SHAP interpretability for short-term solar irradiance prediction in a hot arid climate

Short-term global horizontal irradiance (GHI) forecasting is essential for solar grid integration in hot arid sites due to atmospheric variability and dust aerosols. This study proposes a hybrid deep learning (DL) pipeline that combines Discrete Wavelet Transform (DWT), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) layers for 15-min GHI forecasting at a site in Riyadh, Saudi Arabia. Wavelet decomposition and model hyperparameters are jointly optimized using Optuna’s Bayesian search across 20 trials, and SHAP-based feature selection retains only the most informative predictors. Uncertainty is quantified through Monte Carlo dropout across 100 stochastic forward passes with a dropout rate of 0.2. The model is evaluated on 105,312 meteorological records using five-fold time-series cross-validation. Across all observations, the model achieves a mean RMSE of 6.62 ± 2.22 W/m 2 and an R 2 of 0.98 ± 0.01. Daytime-filtered evaluation (GHI > 50 W/m 2 ) yields a mean RMSE of 8.28 ± 1.45 W/m 2 and an R 2 of 0.961 ± 0.015, providing a more conservative estimate of forecasting accuracy during operationally relevant periods. The full pipeline reduces RMSE by 28–40% compared to ablated model variants and achieves a 47% reduction compared to the persistence baseline. The proposed pipeline delivers interpretable solar irradiance forecasting with calibrated uncertainty estimates suited for grid management applications at arid-climate sites.