Skynet-Driven Prediction of Synthetic Sky Images for Ultra-Short-Term Solar Forecasting

Ultra-short-term solar forecasting plays a critical role in optimizing the efficiency and reliability of photovoltaic (PV) systems by predicting rapid fluctuations in solar irradiance caused by cloud movements. This study presents a novel framework that leverages Skynet, a deep learning-based approach, to predict synthetic sky images for ultra-short-term solar forecasting. Our model utilizes a CNN-BiLSTM architecture, which combines convolutional neural networks for spatial feature extraction and bidirectional LSTM for capturing temporal dependencies in sequential sky image data. The model is trained using the SKIPP’D dataset, consisting of high-resolution sky images collected from Stanford University’s campus, and is designed to predict solar power output at intervals of 1 to 15 minutes. The results indicate that the CNN-BiLSTM model significantly outperforms other models, such as ConvLSTM and PhyDNet, especially in forecasts up to 3 minutes, where rapid cloud movements have a critical impact. While the model shows high accuracy in short-term predictions, its performance gradually declines with longer forecast horizons, suggesting the need for further enhancements. This research contributes to the integration of solar energy into grid systems by providing accurate, real-time forecasting that can inform energy management strategies and reduce reliance on dispatchable power sources.