• Dynamic cloud tracking enhances detection of rapid PV power ramps for minute-ahead forecasts. • A four-stage deep learning model fuses sky images, weather data, and historical PV power for accuracy. • Quantile regression provides calibrated prediction intervals with improved sharpness. • Proposed method reduces deterministic errors by 21–59% versus benchmark models. • Prediction intervals are 22–31% narrower than LUBE while maintaining high coverage. Fluctuations in photovoltaic (PV) power generation caused by rapidly moving clouds often lead to abrupt power ramps, posing challenges to stable grid operation. Developing an accurate ultrashort-term forecasting model is therefore essential for improving dispatch efficiency and maintaining system reliability. To address this problem, this paper proposes an ultrashort-term PV forecasting framework that jointly performs deterministic and probabilistic prediction, incorporating dynamic cloud tracking and sky-image-based correction. A cloud-coverage ratio forecasting module is developed using dense optical flow to estimate cloud movement, enabling early detection of potential power ramps. In addition, meteorological variables obtained from the weather research and forecasting model, together with historical PV power measurements, are integrated into a four-stage deep learning architecture enhanced with quantile regression to generate both point forecasts and calibrated prediction intervals. Comprehensive experimental evaluations, including comparisons with conventional and recent forecasting models as well as ablation analyses, demonstrate that the proposed framework consistently improves forecasting performance. Deterministic results show that the proposed model reduces NRMSE by 21–59% compared with conventional benchmark models and by 4.5–13% relative to the strongest baseline. For probabilistic forecasting, the method achieves coverage rates of 82–95% across confidence levels of 80–95%, while maintaining 22–31% narrower prediction intervals than the lower–upper bound estimation method, indicating sharper and more reliable uncertainty representation. These results verify the robustness of the proposed model for rapid PV fluctuations and demonstrate strong applicability for solar power scheduling.
Jittratorn et al. (Sun,) studied this question.