Convolutional Deep Learning hierarchical classifier to identify synoptic sky conditions based on sky images

Accurate solar irradiation forecasting is essential for efficiently integrating and managing solar energy into grid systems. Solar irradiance nowcasting models based on sky cameras are based on different steps (cloud segmentation, altitude estimation, speed estimation, tracking, and optical depth estimation, among others) with great complexity and prone to errors. The amount of clouds in desert regions is lower than in other areas. Therefore, models require some intelligence to identify the cloud synoptic conditions smartly, as a human meteorologist would. This study uses a convolutional deep learning approach to classify sky synoptic conditions regimes in the images obtained by the all-sky imager (ASI) into four classes: clear-sky, high-level, mid-level, and low-level clouds. ASI provides comprehensive, hemispherical sky images, capturing various cloud conditions formations and their dynamics in an area dependent on the cloud bottom height (CBH). By analyzing these images and employing weighted and unweighted deep convolutional neural networks (CNN), we achieve about 93 percent accuracy in cloudlevel classification using sky-camera images. This hierarchical classification facilitates to use of different models depending on the sky conditions, reduces the complexity of the overall model $(O n)$, and limits the type of clouds in the segmentation. Quality controls the cloud height estimation to produce more precise solar irradiation nowcasts, ultimately supporting better operational strategies for solar power generation and grid management. Thus, integrating ASI-based sky synoptic conditions classification with solar nowcasting models represents a significant advancement since the model can acquire the capabilities of a meteorologist expert system incorporating knowledge base information emulating the decision-making ability of a human expert.