CloudPVNet: A fine-grained ground-based cloud image segmentation method for photovoltaic power forecasting

Photovoltaic (PV) power generation depends on solar irradiance fluctuations, primarily caused by cloud movement. Specifically, different cloud types and distributions can substantially affect solar irradiance. Therefore, precise analysis of cloud conditions is crucial for reliable PV power forecasting. However, the lack of significant advancement in fine-grained segmentation of cloud images over PV power stations impedes the acquisition of the high-quality data necessary for accurate PV power forecasting. To address this limitation, this study quantifies surface irradiance attenuation under different cloud types by comparing theoretical and actual values using synchronized ground-based cloud images. Building upon this research, a fine-grained cloud image segmentation dataset tailored for PV power forecasting is constructed to establish the correlation between cloud optical properties and irradiance attenuation characteristics. Furthermore, we propose CloudPVNet, a fine-grained ground-based cloud image segmentation model for PV power forecasting, including the parallel convolution attention (PCA) module and the wavelet-enhanced feature fusion (WE-FF) module. The PCA module combines multi-branch convolution and directional attention to capture multi-scale cloud features. The WE-FF module uses wavelet convolution to extract low-frequency features for fusion, enhancing upsampling and boundary delineation. Experimental results demonstrate that the proposed model outperforms existing cloud segmentation methods, offering superior adaptability and reliability in PV applications.