Solar irradiance forecasting with CNN-LSTM based on wavelet transformation and shap-supported feature selection

This thesis proposes a novel framework that combines a wavelet-based hybrid CNN-LSTM model with SHAP-supported feature selection and analysis to achieve high accuracy and interpretability in solar irradiance forecasting. The study used three years of meteorological data from Riyadh, Saudi Arabia, collected at 15-minute intervals. Manual feature engineering was applied to enrich the dataset. Then, Discrete Wavelet Transform (DWT, db4, level 2) was used for noise reduction and multi-scale component extraction. CNN layers modeled spatial patterns, while LSTM layers captured temporal dependencies. Hyperparameters were optimized using Optuna-based Bayesian optimization. SHAP analysis was employed to select the top 20% of variables with the highest average contribution, and the model was retrained. Ablation studies were conducted to validate component contributions. The evaluation, performed using 5-fold TimeSeriesSplit cross-validation, showed low RMSE and high R² scores. Variables such as zenith angle and cloud opacity played a dominant role in predictions. Monte Carlo Dropout was applied for uncertainty estimation, and PCA-based dimensionality reduction was used to improve generalization capability. The results show that the proposed method is a reliable tool for assessing solar energy potential, grid integration, and energy management. This approach offers a flexible prediction system applicable to regions with similar climate conditions and provides a valuable contribution to explainable deep learning-based solar irradiance forecasting studies in the literature.