MSSEAG-UNet: A Novel Deep Learning Architecture for Cloud Segmentation in Fisheye Sky Images and Solar Energy Forecast

This study analyses sky images captured using a ground-based fisheye camera, aiming to address the challenge of accurately segmenting clouds, which is difficult due to their fuzzy and indistinct boundaries and uneven lighting conditions. Accurate segmentation of clouds in ground based sky images are crucial for accurate solar energy forecasting. Motivated by these challenges, this paper has proposed a novel deep learning architecture called MSSEAG-UNet (Multi-Spatial Squeeze-and-Excite Attention Gated U-Net) for cloud segmentation in ground-based fisheye sky images. The proposed architecture integrates a Multi-Spatial Convolutional (MS-CNN) block and Squeeze-and-Excite (SE) blocks in the encoder path to improve multi-scale feature extraction and recalibrate feature maps, while an attention block is incorporated in the decoder path to emphasize key cloud features. The segmentation performance of the MSSEAG-UNet is compared with five benchmark models, and results show that the proposed model outperforms than all benchmarks models. Further, the segmented cloud images produced by the MSSEAG-UNet are used to calculate the cloud percentage, which is then integrated with the original sky images using a Multi-Column Convolutional model for Global Horizontal Irradiance (GHI) forecast. GHI forecast is conducted for 15-minute, 30-minute, and 60-minute ahead timesteps, with the best results achieved for the 60-minute forecast, yielding MAE, MSE, and RMSE values of 6.245%, 0.683%, and 8.265%, respectively. These results highlight the effectiveness of the proposed approach in improving both cloud segmentation accuracy and short-term solar irradiance forecasting.