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

With the increasing penetration of photovoltaic (PV) power generation in the grid, minutely irradiance prediction, which is the basis of minutely PV power prediction, has become very important. Aiming at the problems of extracting feature redundancy, weakening key area features, and high cloud-sky misidentification rate in the current minutely irradiance prediction, this paper proposes a minutely solar irradiance forecasting method based on multidimensional feature extraction using all-sky image, which is beneficial to achieve higher accuracy in solar power forecasting. First, the improved clustering-boundary correction algorithm is used to identify cloud and sky pixels and classify the all-sky images into four cloud-sky types. Then capture the subimages of the cloud domain that will cover the sun in the future dynamically according to the results of the cloud displacement vector calculation and extract local features and overall features as multidimensional features based on convolutional neural network (CNN) and image RGB matrix, respectively. Finally combined the multidimensional features with meteorological factors and historical irradiance to respectively construct irradiance mapping models for four cloud-sky types to achieve irradiance prediction on a ten-minute scale. Compared with the benchmarks, the mean absolute percentage error (MAPE) of the proposed method is reduced by 0.21%, 20.21%, 2.53%, and 5.30% for four cloud-sky types: clear sky, block clouds, thin clouds, and thick clouds, respectively. The proposed method can be widely used in PV plants equipped with all-sky imagers to provide data support for the optimal operation and maintenance of the plant.