呂藝光Leu, Yih-Guang陳家豪Chen, Chia-Hao2023-12-082023-08-092023-12-082023https://etds.lib.ntnu.edu.tw/thesis/detail/34380e23578cc6be4b7f73b8d7db6e98/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/120341因應太陽能發電日益重要，又因太陽能發電量受制於日射量的影響，同時考量太陽能發電是間歇性的能源，故提出一種基於深度學習光流應用於全天空影像之日射量估計與預測的方法。由於日射量變化容易受到天氣狀況影響，本文藉由全天空影像，建立一個以影像特徵為基礎之日射量估計與預測系統，其影像特徵包括，利用紅藍比例法計算影像中雲層比例、雲層厚度，及藉太陽位置演算法得到影像中太陽位置，分析太陽附近雲層特徵，萃取全域與局部的雲層權重。同時用基於深度學習的光流法推算雲層移動的狀態，並製作未來數分鐘之天空預測圖像，將這些預測圖像作長短期記憶(LSTM)之輸入特徵，日射量作為訓練輸出，其深度學習光流法是透過卷積神經網路來實現。本文將資料集區分為月、季、半年與一年，分別進行10分鐘至60分鐘的日射量預測。同時，本文使用多個效能指標評估效果，包括平均絕對誤差(Mean Absolute Error)、均方根誤差(Root Mean Square Error)與判定係數(R^2)。最後，與文獻的方法進行比較，本文所提的方法具有較好的結果。Due to the increasing importance of solar power generation, the fact that solar power generation is affected by the amount of solar irradiance, and the fact that solar power is an intermittent energy source, the study of solar irradiance estimation and prediction based on deep learning optical flow method is presented. Since the change of solar irradiance is easily affected by the weather conditions, this study utilizes the all-sky images to establish a system for estimating and predicting solar irradiance by analyzing and processing image features. Meanwhile, the deep learning optical flow method implemented using convolutional neural network is used to predict cloud movement states and generate future sky images which are used as input features for LSTM.In this study, we categorize the datasets into monthly, quarterly, semiannual, and annual datasets and predict the irradiance for the next 10 minutes to 60 minutes, respectively. In order to illustrate the validity of the predictions and estimates, several performance metrics, including MAE, RMSE and R^2, are used. Finally, the proposed method is compared with several methods in the literature and the results show better performance.深度學習卷積神經網路全天空影像光流法長短期記憶Deep learningconvolution neural networkall-sky imagesoptical flowLSTMetd