Ultra-Short-Term Solar Power Prediction Using Sky Image Sequences by a Residual Vision Reformer

The unpredictable nature of solar power generation, largely influenced by fluctuating cloud cover, poses a challenge to the stability of renewable energy systems. Considering this, accurate forecasting of solar power can lead to better grid management and operation. With the advent of deep learning models, various models have been suggested to enhance the ultra-short-term solar power forecasting performance. Given that cloud images offer more direct and comprehensive information about cloud patterns compared to the numerical weather prediction data, analyzing cloud images allows for more precise and efficient cloud change predictions, leading to a more accurate ultra-short-term solar power forecasting. In this way, aiming to enhance the forecasting performance, in this paper, we introduce a deep learning-based model, including three main blocks. In the first block, a Multi-Stream Video Vision Transformer (MS-ViViT) model is proposed for extracting different types of spatio-temporal features from the input image sequences. The output features from the first block are input to the second block, Fused Improved Reformer (Fused I-Reformer), including three Improved Reformer (I-Reformer) models equipped with a Fused Encoder as well as a new loss function for sequence learning. Finally, an Attentive Residual Fully Connected (ARFC) model is proposed for solar power value prediction. The comparison results with 36 comparative models on six real-world datasets using seven evaluation metrics confirm the effectiveness of the proposed ultra-short-term solar power forecasting model.