Research On Ultra-Short-Term Photovoltaic Power Forecasting Based On Cloud Image Data And Deep Learning Technology

Compared with traditional thermal power,photovoltaic power generation has obvious random volatility and uncontrollability.Large-scale installation and grid connection of photovoltaic power generation bring severe challenges to the secure,stable and economic operation of power systems.Photovoltaic generation power forecasting can provide an important support for suppressing the uncertainty of photovoltaic generation power,and it is an important means to improve the level of photovoltaic generation power consumption.This paper focuses on photovoltaic generation power forecasting need under ultra-short-term time scale,and is based on the data sets including ground-based cloud images and satellite cloud images,and deep learning technology,and studies minute level cloud distribution prediction,station level ultra-short-term photovoltaic power forecasting and cluster level ultra-short-term photovoltaic power forecasting,respectively.For minute level cloud distribution prediction,digital image processing technology(DIPT)commonly used has two shortcomings: the input spatiotemporal information is relatively limited and the hypothesis that the adjacent cloud distributions are the same.To address these two shortcomings,cloud distribution prediction models based on two-dimension(2D)and three-dimension(3D)convolutional auto-encoder(CAE)are proposed.Two typical DIPT methods,including particle image velocimetry(PIV)and Fourier phase correlation theory(FPCT),are used to build the contrast models.The performance was compared in five different prediction time resolution scenarios,and the results show compared with DIPT models,CAE models can better predict the structure information of images,and they also show some advantages in predicting the gray information of images.Besides,compared with 3D CAE models,2D CAE models using a simpler algorithm could achieve an overall slightly higher performance than the former.For ultra-short-term photovoltaic power forecasting,its purpose is to provide output data in the following four hours.For station level ultra-short-term photovoltaic power forecasting,the observation method based on ground-based cloud images is difficult to achieve hourly level prediction due to the limitation of observation range.To introduce cloud image information under a longer time scale,a CAE-based automatic extraction method of cloud distribution features(CDF)is proposed.In addition,the usual feature fusion method is only one-way splicing,which doesn’t consider time step correspondence among information.Therefore,an ultra-short-term forecasting method based on long short-term memory(LSTM)and time step feature fusion is proposed.Based on the dataset including global horizontal irradiance(GHI)and CDF,the station level ultra-short-term forecasting model LSTM-GHI-CDF is used to predict the GHI in the next 1 to 6 step whose interval is 10 min.Results show the proposed model is superior compared with the two contrast models under multiple scenarios.For cluster level ultra-short-term photovoltaic power forecasting,considering the similarities and differences between station level and cluster level ultra-short-term forecasting research in this paper,this research draws lessons from the above station level ultra-short-term forecasting research.Firstly,CDF is automatically extracted from satellite cloud images with relatively large observation range by CAE.Secondly,construct the dataset containing photovoltaic power(PVP),CDF,and numerical weather prediction(NWP).Thirdly,based on the above dataset,cluster level ultra-short-term forecasting model LSTM-PVP-CDF-NWP is used to predict cluster power in the next 1 to 16 step whose interval is 15 min.Results show the proposed model is superior compared with the multiple contrast models under multiple scenarios.