Semantic Segmentation of Cloud Images captured with Horizon-Oriented Cameras for Nowcasting Applications
Year: 2023
Type: preprint
Abstract: The goal of this research is to create a cloud identification model for short-term solar power generation forecasting using deep learning methods. Deep learning techniques such as U-net, HRnet, and Detectron2 were used to perform semantic and instance segmentation of clouds. Transfer learning was employed, with weights trained on the "2012 ILSVRC ImageNet" dataset and architecture configurations such as Resnet, Efficient, and Mask RCNN. As a validation metric, the models were trained using the Jaccard Coefficient. The best model, a U-net with Resnet18 at 486x648 resolution, achieved an average Intersection over Union (IoU) of 0.6, indicating good cloud segmentation performance. All models were affected by the data imbalance, with the tree class causing a favorable bias towards the overall results. The HRNet model demonstrated refined pixel-level segmentation but was ineffective at detecting dominant clouds. This research provides insights into effective cloud identification techniques and lays the groundwork for future cloud classification model development. The model may be able to predict cloud motion. For future works the team will investigate dataset balancing and data augmentation transformations for model improvement. The research demonstrates the effectiveness of deep learning in cloud identification for solar power forecasting.
Source: Research Square
Authors: Bruno Juncklaus Martins, Juliana Marian Arrais, Allan Cerentini, Sylvio Mantelli, Gilberto Perello Ricci Neto, Aldo von Wangenheim
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4381615565