"Development of advanced cloud classification and segmentation models for solar energy applications using deep learning and synthetic data augmentation"

Titel:	Development of advanced cloud classification and segmentation models for solar energy applications using deep learning and synthetic data augmentation
Untertitel:
Hochschule:	Hochschule Kempten
Fachbereich:	Computer Science
Studiengang:	Computer Science
Geschrieben von:	Raphael Gut

Vollständige Arbeit als PDF zum Download

Abstract

The transition to clean and renewable energy is one of the most present and defining challenges of our time. Among all available energy sources, solar power is the most abundant, yet its availability is highly variable due to fluctuations in solar irradi- ance. The primary cause of intra-hour fluctuations is cloud cover, which significantly impacts local irradiance levels. To maintain grid stability, mitigate ramp events in large-scale photovoltaic sites, and optimize the efficiency of Concentrated Solar Power (CSP) plants, reliable forecasting systems are essential. Nowcasting systems address this need by providing intra-hour forecasts, enabling the anticipation of short-term variations in solar irradiance.

A common approach to Nowcasting involves using ground-based All-Sky Imagers to capture sky images, detect clouds, and track their movement to predict future solar irradiance. The accuracy of these predictions largely depends on the quality of cloud detection, which is typically performed at the pixel level. Modern deep learning- based methods have emerged as the dominant approach, outperforming traditional techniques. However, one of the key challenges of these methods is their reliance on large, high-quality ground truth datasets for training and validation. Since manually annotating such datasets is labor-intensive and time-consuming, obtaining sufficient labeled data remains a significant challenge.

This thesis aims to enhance existing semantic cloud segmentation models by in- corporating temporal dependencies between consecutive image sequences captured by All-Sky Imagers. First, a semi-supervised video segmentation approach was em- ployed to expand an existing human-annotated ground truth dataset by a factor of 20, resulting in a total of 16.170 images. Four different methods were proposed and benchmarked against each other, demonstrating the effectiveness of this approach. Additionally, a new model architecture was developed that integrates motion cues from consecutive All-Sky Imager pairs alongside the primary image, leveraging cloud dynamics and transitions to provide richer prior information for the semantic seg- mentation process to learn better feature representation.

Evaluation on a validation dataset confirms the efficacy of both approaches. In par- ticular, the semantic cloud segmentation model benefits from the enlarged training dataset, achieving improvements in accuracy and Intersection over Union (IoU) by 2.6% points and 3.5% points, respectively, compared to the current state-of-the-art model. Moreover, the motion cue-enriched model significantly enhances the differ- entiation of cloud classes, improving detection accuracy by up to 3.3% points for previously challenging cloud types.