Ultrashort-Term Multistep Photovoltaic Power Forecasting Based on Whole-Sky Images and Conditional Temporal Fusion

Photovoltaic (PV) power output is strongly affected by the rapid evolution of clouds. Traditional methods that rely only on historical power sequences or low-spatiotemporal-resolution meteorological data struggle to accurately characterize rapid fluctuations and ramp events within a 10–20 min horizon, where PV power rises or drops sharply over a short period. To address this issue, this paper develops an ultra-short-term multistep forecasting framework centered on whole-sky images (WSIs) and recent PV observations. We first design a static image model and a static hybrid model based on the large-small network (LSNet) to mine spatial structural information from single-frame WSIs and use power-conditioned cross-attention to alleviate the representation ambiguity that similar sky appearances may correspond to different output levels. We further propose a dynamic image learning model based on LSNet and a temporal convolutional network, which models short-term cloud-motion evolution on top of frame-wise spatial encoding and introduces a multiscale radiative transmission index (RTI) and a ramp-aware multistep training objective to enhance sensitivity to rapid cloud-occlusion progression and high-risk samples. Experiments on the public Sky Images and Photovoltaic Power Generation Dataset (SKIPP’D) show that the proposed method outperforms the persistence model, single-layer neural network, long short-term memory, and recent baselines such as RTI-Net, SolNet, and SolarNexus in both general forecasting and ramp-event scenarios. Under general forecasting conditions, our method achieves a root mean square error (RMSE) of 27.48 and a mean absolute percentage error (MAPE) of 9.07, which are 29.1% and 28.1% lower than those of the smart persistence model (SPM), respectively. Under ramp-event scenarios, balanced performance reaches 76.84%, a 52.4% improvement over SPM, while ramp-event RMSE and ramp-event MAPE are reduced to 58.93 and 17.11, corresponding to decreases of 47.7% and 62.2%, respectively. These results indicate that jointly exploiting the spatial structure of sky images, recent power context, and ramp-oriented modeling strategies can substantially improve the accuracy and dispatch usefulness of ultra-short-term multistep PV power forecasting.