In the rapidly evolving landscape of renewable energy, accurate solar power forecasting is becoming increasingly crucial for grid stability and efficiency. A recent study published in the journal *Energy and Artificial Intelligence* offers a comprehensive evaluation of cutting-edge techniques for very short-term solar power forecasting, providing valuable insights for the energy sector.
The research, led by Nguyen Binh Nam from the Department of Energy – Electrical Engineering at Politecnico di Milano and the Faculty of Electrical Engineering at The University of Danang – University of Science and Technology, systematically compares multiple conformal prediction algorithms with advanced machine learning and deep learning models. The study focuses on forecasting solar power generation just minutes ahead, a critical timeframe for grid operators.
“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,” Nam explains. The study integrates and compares state-of-the-art conformal prediction algorithms—including Inductive Conformal Prediction (ICP), Jackknife+ after Bootstrap (J+aB), and Ensemble Bootstrap Prediction Intervals (EnbPI)—with machine learning models like Random Forest, XGBoost, LSTM, and Transformer.
One of the unique aspects of this research is the combination of the Transformer architecture with conformal prediction techniques to construct prediction intervals. The study also investigates the impact of temporal segmentation on forecasting performance, providing a detailed cross-site comparison of conformal prediction interval techniques for solar nowcasting.
The research was validated on two real-world photovoltaic datasets from Ninh Thuan, Vietnam, and SolarTechLAB, Politecnico di Milano, Italy, representing diverse climatic and operational conditions. The results show that traditional models like Random Forest and XGBoost, when paired with conformal prediction methods, consistently achieve well-calibrated and efficient intervals. However, deep learning models, particularly LSTM, tend to produce overly narrow intervals with significant undercoverage. Notably, the Transformer model demonstrated robust performance, balancing interval sharpness and reliability across both sites and all horizons.
This work 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. The findings could significantly impact the energy sector by improving the integration of renewable energy sources into the grid, enhancing grid stability, and optimizing energy management systems.
As the world continues to transition towards renewable energy, the ability to accurately forecast solar power generation will be crucial. This research provides a valuable framework for future developments in the field, offering insights that could shape the next generation of energy forecasting technologies.