呂藝光Leu, Yih-Guang張奕舜Zhang, Yi-Shun2025-12-092025-02-062025https://etds.lib.ntnu.edu.tw/thesis/detail/bb3116e4ea737e71238fc7fb61b54346/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/125043本論文使用全天空影像進行分析，來發展雲層為特徵的日射量估計與預測系統。在全天空影像中，使用紅藍比例法擷取雲層特徵影像，使用自適應閾值在不同亮度的情況下更精確判定雲層資訊計算雲層整張圖的占比，以及太陽周圍雲特徵分析，透過RAFT (Recurrent All-Pairs Field Transforms)光流法推估雲層移動情形，製作未來數分鐘至小時之雲層情況，並提取雲層特徵，作為生成式對抗網路(GAN)模型之輸入，日射計測出之日射量為輸出。並使用三個評估指標來查看模型學習的狀況，有相對均方根誤差(rRMSE)、相對絕對平均誤差(rMAE)、預測技巧(FS)比較估計及預測成果。This study analyzes all-sky images to develop a cloud feature irradiance estimation and prediction system. In the all-sky images, the red-blue ratio method is used to extract cloud feature images, while adaptive thresholding enables more accurate cloud information detection under varying brightness conditions. The system calculates the cloud coverage ratio for the entire image and analyzes cloud features around the sun. The RAFT (Recurrent All-Pairs Field Transforms) method of estimating cloud movement produces forecasts of cloud conditions for the next few minutes to hours. Cloud features are extracted and used as input for a Generative Adversarial Network (GAN) model with solar irradiance measured by a pyranometer as output. Three evaluation metrics—relative root mean square error (rRMSE), relative mean absolute error (rMAE), and forecasting skill (FS)—are used to assess the model's performance in estimation and prediction tasks.全天空影像RAFT光流法生成式對抗網路All-sky imagesRAFT optical flowGenerative Adversarial Network學術論文