Research On Feature Extraction And Identification Model Of Cloud In Ground-Based Sky Images

Cloud space identification of ground sky image and cloud prediction are the premise of studying on the distribution and change of photovoltaic(PV)power in regional power grid.It is of great significance for supporting scheduling and increasing the proportion of consumptive PV power generation.This paper is supported by National Science Foundation of China for Research on Ultrashort-Term Prediction of Photovoltaic Power Generation on Basis of Mathematical Description and Quantitative Representations of Cloud Motions,Research on Major Weather Factors and Power Prediction Model of Photovoltaic Power Generation as well as Scientific Program of Yunnan State Grid,and Research on Cloud Movement Vector Field and Cloud Identification Method.The paper adopts Otsu and K-means clustering algorithm to carry out research about cloud identification model.The proposed model is applied for cloud identification of sky image in Yunnan Province.This paper takes TSI-VIS-J1006 as an example to introduce the operation principles of ground-based sky image capturing device and some technical parameters including sampling frequency,resolution and file format.Then,RGB color model and HSV model are introduced according to visual characteristics.By analyzing characteristics of cloud clusters and clear sky shown in sky images,the paper establishes two cloud identification models by using Otsu and K-means algorithm.To verify the effectiveness of these two models,a simulation study is conducted on sky images captured by TSI-VIS-J1006 total-sky imager of a PV power plant in Yunnan Province.Finally,this paper compares the processing outcomes of these two models with the outcome of fixed threshold method,the results show that the cloud identification method based on ratio of red and blue components which adopts fixed threshold value presents poor effect and high fall-out rate.Otsu algorithm can select optimum threshold value to divide saturation matrix of the sky image into two parts,so as to seek the maximum value of between-class variance of these two parts.The maximum value represents the most significant selectivity.In addition,this paper demonstrates it is feasible to reduce identification errors caused by imbalanced sky image pixels by adding additional images to clear or cloud sky images.Through this way,researchers can efficiently and accurately select cloud clusters from sky images.K-means algorithm uses the high ratio of red and blue components and the low ratio of red and blue components as fixed threshold values to divided ground based sky images.Sequentially extract the location information of the sky pixels and the cloud pixels in the identificationresults and get the corresponding position RGB value of cloud pixel and the sky pixels in the original image.And then calculate average of the sky pixels and cloud pixels obtained and make mean value of the red and blue components be divided to get the initial cluster center.Then use the K-means algorithm,use Weighted Euclidean distance to calculate the distance between each cluster sample and cluster center.Get the clustering results by several iterations.Then the clustering results are reduced to a matrix,so the cloud space identification results of ground sky images are obtained.After processing in above method,K-means algorithm overcomes instability and inaccuracy of traditional K-means caused by random selection of clustering center.In addition,this paper comes up with a method to identify cloud of the sky image by adopting several attributes and setting weighing coefficients of the attributes of the sky image.Research outcome of this paper reveals ground-based sky image identification based on the above two cloud identification models which adopt Otsu and K-means algorithms show better identification performance than the model based on fixed threshold value.Accurate cloud identification is the prerequisite for study of distribution and change of PV power generation of power grid,which lays a solid foundation for the implementation of accurate and reliable PV power generation prediction.