在天氣觀測中,本論文以組合的方式結合觀測與統計的方法估算天空輻射值(GHI),並藉由雲層動態來預測未來短時間內天空輻射值的變化情形,藉以解決太陽能發電的相關問題。 在此論文中提出一種使用紅色與藍色色度比值(Cr/Cb)的方法來分析全天空影像(TSI)的雲層分佈特徵,接著依據長期的觀測資料,使用最小平方法(LSM)分析無雲層干擾下天空輻射值的時間特徵,並透過類神經網路(NN)訓練兩種特徵資料以估算天空輻射值,最後以光流法(OF)搭配最近行為優先估測法(RFE)來追蹤雲層的移動軌跡,並預測未來短時間內的天空輻射值。 由實驗結果顯示類神經網路的平均絕對百分比誤差(MAPE)可達5.8161%,顯示此方法具高度準確的估算能力,而動態預測在3分鐘內雲層特徵影像的均方根誤差(RMSE)少於0.02,顯示此方法可在短時間內有效地追蹤雲層動向。
In the meteorological observation, this dissertation proposed a hybrid method by combining the observation and the statistics to estimate the Global Horizontal Irradiance (GHI). Then the cloud motion estimation is adapted to forecast solar power. In this dissertation, we analyze the distribution of the Total Sky Image (TSI) by the Red and Blue Chroma ratio (Cr/Cb), quantize the Global Horizontal Irradiance without cloud interference by the Least Squares Method (LSM) according to the long-term observation data, train the above two types of characteristic data by the Neural Network (NN) to estimate the Global Horizontal Irradiance. Finally, the Optical Flow method (OF) and the Recent First Estimator (RFE) are used to track the movement trajectory of the cloud, and forecast the Global Horizontal Irradiance in a short time future. The experimental results show the accuracy of estimation by the Mean Absolute Percentage Error (MAPE) of the Neural Network is 5.8161%, which shows that this method has a high accuracy estimation ability. And the Root Mean Squared Error (RMSE) of the cloud feature image in the prediction less than 3 minutes is 0.02, it shows that this method can effectively track the cloud movement in a short period of time.