Hybrid cloudiness-irradiance modeling for enhanced 24-hour solar forecasting with multi-source data

Accurate solar irradiance forecasting is critical for optimizing photovoltaic (PV) energy integration, enhancing market participation, and reducing im-balance costs in intraday trading. However, existing methods often depend on single data sources, costly real-time information, or infrequent updates in public databases, limiting their effectiveness. To address these challenges, we propose a system-level hybrid framework that integrates multi-source mete-orological datasets with advanced artificial intelligence techniques. The ap-proach combines Machine Learning and Deep Learning methods within a 24-ANN stacking architecture, merging cloudiness-based predictions from Ran-dom Forest and eXtreme Gradient Boosting models with irradiance-based predictions from a CNN-Transformer architecture. Leveraging open-access ERA5 and AEMET datasets, our framework delivers accurate deterministic 24-hour ahead forecasts updated hourly, achieving a Root Mean Square Error (RMSE) reduction of 43.86% compared to the persistence baseline and up to 31.88% relative to standalone component models. Systematic ablation exper-iments validate the contribution of each component, demonstrating robust performance across diverse weather conditions. The system offers a cost-effective, regionally scalable solution for solar energy management across the Iberian Peninsula.