Research On Cloud Motion Prediction Method Based On Ground-Based Cloud Image Sequence

Tower solar thermal power generation system has attracted much attention in the field of solar power generation due to its advantages such as large concentration ratio,high photoelectric conversion efficiency,and the possibility of large-scale integration.However,the sudden change of solar irradiance caused by the movement of clouds over the heliostat field will affect the stability of power output at least,and damage the key equipment in the power generation system at worst.Therefore,it is of great significance to accurately predict the cloud movement trend in real time and take protective measures before the sudden change of solar irradiance to ensure the safe and stable operation of the tower solar thermal power generation system.Based on the project"Cloud Prediction System Development" in cooperation with Dongfang Boiler Co.,Ltd,this thesis uses image processing technology to predict cloud motion in the ground-based cloud image sequence carried out research and exploration,the main work and innovations are as follows:(1)Aiming at the problems of cloud deformation,complex motion and difficult segmentation of ground-based cloud images,a BKM-SIFT-KF cloud motion prediction algorithm is proposed.In view of the different chromaticity information between the sky and the cloud layer,this algorithm enhances the blue-red difference(BRD)between the cloud and the sky by adjusting the chromaticity of the cloud image,and uses the K-Means method to cluster the BRD features of the cloud image,to realize the accurate segmentation of clouds and sky;considering the deformation of clouds and the different motion states of different clouds in the same cloud image,the connected domain detection is used to obtain each cloud layer in the segmented cloud image,and the SIFT feature point matching algorithm is used to obtain The corresponding position of the feature point of each cloud layer in the cloud image sequence;finally,the Kalman Filter(KF)is used to model the position of the feature point of the cloud layer,and the parameters of the prediction model are dynamically updated to adapt to the change of cloud movement,thus accurate prediction of cloud position and movement speed is realized.(2)When the cloud is thin or has a high degree of deformation,it is difficult for the traditional prediction method to obtain the matching characteristics of the cloud in the cloud image sequence,so that it is impossible to establish a cloud motion prediction model.To solve this problem,this thesis introduces deep learning video frame prediction into cloud motion prediction,and proposes a PredRNN-SIFT-BKM prediction model.Different from traditional forecasting methods,this model first constructs the PredRNN cloud image prediction network,which captures the deformation characteristics and complex nonlinear motion process of the cloud through its adaptive feature learning and strong nonlinear mapping capabilities,so as to obtain accurate forecast cloud images.Based on the predicted cloud image,using the matching and detection method proposed in research content(1),the position and motion velocity of each cloud layer are calculated,and the cloud motion prediction based on the predictionmatching-detection framework is realized.The experimental results show that this PredRNNSIFT-BKM prediction model can accurately predict the future frame cloud image,which not only solves the problem that the cloud in the cloud image sequence cannot be modeled and predicted due to lack of matching features,but also improves the prediction accuracy of the cloud position and motion speed.