A ground-based cloud image classification method for photovoltaic power prediction based on Convolutional Neural Networks and Vision Transformer

Abstract

Cloud image classification plays a crucial role in accurately predicting solar radiation attenuation, which significantly impacts ultra-short-term photovoltaic power predictions. Despite recent advancements in cloud image classification using Transformer and convolutional neural networks, challenges remain, particularly in handling rapidly evolving cumuliform clouds. To address this, we propose CloudMViT, an improved model derived from a model that combines Convolutional Neural Networks and Vision Transformer (MobileViT). CloudMViT introduces the CloudMobileNetV2 (CMV2) Block, building upon the Mobile Network architecture (MobileNetV2). This block incorporates a triple-branch inverted residual structure consisting of depthwise separable convolution, standard convolution, and shortcut branches, along with a New Multi-scale Channel Attention Module (NMS-CAM). Simultaneously, CloudMViT proposes the CloudMViT Block, which integrates adaptive convolution to enhance local feature representation. CloudMViT improves local and global feature extraction, boosting cloud image classification accuracy. The proposed model was evaluated through ablation studies and comparative experiments on the Tianjin-Normal-University-Ground-based-Cloud-Dataset (GCD) and Tianjin-Normal-University-Ground-based-Remote-Sensing-Cloud-Database (GRSCD). The ablation experiments demonstrate that the modules introduced in CloudMViT significantly improve cloud image classification accuracy. Comparative results show that CloudMViT achieves higher accuracy than other state-of-the-art methods, reaching 91.40% on GCD and 98.25% on GRSCD datasets. Finally, CloudMViT was validated through experiments in conjunction with a photovoltaic power prediction model, achieving a prediction accuracy of 96.70%, which surpasses that of the original model. This further demonstrates that utilizing CloudMViT cloud image classification results enhances the accuracy of photovoltaic power prediction. Moreover, it verifies that improving cloud image classification accuracy can effectively enhance the precision of photovoltaic power forecasting.