A Minutely Solar Irradiance Forecasting Method Based on Multidimensional Feature Extraction Using All-Sky Images

Abstract

Accurate solar irradiance forecasting is the basis for accurate photovoltaic (PV) power forecasting. However, previous minutely irradiance forecasting methods based on allsky images have difficulty in adequately extracting the cloud features that are essential for future irradiance fluctuations. Therefore, in this paper, a minutely irradiance prediction method based on multidimensional feature extraction of all-sky images is proposed. The raw images are first pre-processed and classified into four weather types by cloud-sky identification. Then the cloud displacement vector is calculated using the optical flow (OF) method, and capture the subimages of the cloud domain that will cover the sun in the future dynamically according to the calculation results. Subsequently, we use the convolutional neural network (CNN) to extract multidimensional features including the local features and overall features. The multidimensional features are combined with relevant meteorological factors and historical irradiance to construct irradiance mapping models for each of the four weather types to achieve irradiance prediction on a ten-minute scale. Simulation results show that the proposed method can better introduce the key cloud information and improve the prediction accuracy.