Comparative analysis of conformal prediction techniques and machine learning models for very short-term solar power forecasting

Accurate and reliable short-term forecasting of solar power generation, along with robust uncertainty quantification, is essential for the effective integration of renewable energy into modern power grids. In this study, we present one of the first comprehensive evaluations that systematically integrates and compares multiple state-of-the-art conformal prediction algorithms—including Inductive Conformal Prediction (ICP), Jackknife+ after Bootstrap (J+aB), and Ensemble Bootstrap Prediction Intervals (EnbPI)—with a diverse set of advanced machine learning and deep learning models (Random Forest, XGBoost, LSTM, and Transformer) for very-short-term (minutes-ahead) solar power forecasting. Uniquely, we combine the Transformer architecture with conformal prediction techniques to construct prediction intervals, and systematically investigate the impact of temporal segmentation on forecasting performance. Our framework is validated on two real-world photovoltaic datasets from Ninh Thuan, Vietnam and SolarTechLAB, Politecnico di Milano, Italy, representing diverse climatic and operational conditions, and is assessed across 5, 10, and 15 min nowcasting horizons with a focus on both prediction interval coverage and efficiency. The results show that while traditional models like Random Forest and XGBoost, when paired with conformal prediction methods, consistently achieve well-calibrated and efficient intervals, deep learning models – particularly LSTM – tend to produce overly narrow intervals with significant undercoverage. Notably, the Transformer model demonstrates robust performance, balancing interval sharpness and reliability across both sites and all horizons. This work provides a detailed, cross-site comparison of conformal prediction interval techniques for solar nowcasting, highlights the importance of model selection, interval calibration, and temporal segmentation, and demonstrates the scalability, adaptability, and practical relevance of the proposed framework for operational renewable energy forecasting.