Research On Ultra Short Term Power Forecast Of Photovoltaic Generation Based On Ground Based Cloud Images

In the context of the rapid development of global economy and society and the increasing demand for energy,the development and sustainable utilization of solar energy has been rapidly improved.Accurate photovoltaic power prediction is not only helpful to regulate the safe and stable operation of the power grid,reduce photovoltaic power limiting,improve the absorption capacity of the power grid,but also can provide accurate information for the power market trading and improve its own economic benefits.However,photovoltaic power is easily affected by environmental factors such as weather,and has high randomness and volatility.It is difficult to ensure the safe and stable operation of the power system when large-scale photovoltaic is connected to the power grid.The movement of cloud blocks the sun is the main factor that causes the fluctuation of photovoltaic power.To solve the above problems,in order to accurately quantify the solar shading of cloud clusters and further improve the accuracy of photovoltaic power prediction,this thesis is based on historical photovoltaic data and groundbased cloud image data set,combined with deep learning related technologies,respectively carried out cloud image restoration and data processing,ground-based cloud image feature extraction and ultra-short-term photovoltaic power prediction combined with attention mechanism.Firstly,aiming at the occlusion problem of the shadow part in the acquired ground-based cloud image,the area to be inpainted is marked by the method of image mask,and a cloud image inpainting method based on improved sample block is designed.This method solves the problem of large error caused by determining the area to be repaired according to the astronomy method,and improves the priority function of the algorithm,which improves the speed and effect of cloud image repair.In order to better screen out the main variables that affect photovoltaic power prediction,Pearson correlation coefficient method is used to analyze the correlation between different variables in photovoltaic historical data characteristics and photovoltaic power,and the variables with weak correlation are removed,which effectively reduces the dimension of the model input.Then,in order to extract the feature information of cloud image,a feature extraction method of cloud transmittance factor based on image brightness is designed.Aiming at the complex changes of cloud movement,the speeded up robust feature algorithm is used to match cloud image features,and the cloud velocity features are extracted to predict the changes of cloud displacement vector at the next time.Considering that cloud thickness has a significant effect on solar radiation transmittance,a maximum entropy multi-threshold segmentation method based on Grey Wolf optimization is established to identify thin cloud,thick cloud and clear sky pixels in cloud images.Photovoltaic feature data is constructed by synthesizing the extracted cloud image feature information and photovoltaic historical data,which provides data support for the input of prediction model.Finally,a combination prediction model based on cloud image features and fusion attention mechanism is designed.In this study,photovoltaic feature data are taken as input,and after processing by convolutional layer and maximum pooling layer,the extracted features are input into the short-duration memory neural network model with attention mechanism,which fully integrated the advantages of each method and improved the overall performance of the photovoltaic power prediction model.In this thesis,the photovoltaic power generation data used by the Renewable Energy Laboratory of the United States is taken as an example,and the combined prediction model is tested in the non-rainfall scenario.Compared with various benchmark methods,the advantages of the combined prediction model designed in this thesis are verified in terms of prediction performance,and the prediction accuracy of photovoltaic power of the combined prediction model with and without cloud transmission factor is compared.The validity of introducing cloud transmittance factor as the input of prediction model is verified.At the same time,considering the requirements of the ultra-short time scale,the operation time of the combined model and the single model algorithm is compared,and the operation time and prediction accuracy of the model are comprehensively analyzed.The results show that the designed combined prediction model can meet the requirements of the ultra-short time scale with high accuracy.