Intra-hour Forecasting Of Cloud Images And Direct Normal Irradiance Trend Based On Total Sky Images

With the development of science and technology,mankind’s demand for energy is also increasing.As a kind of clean energy,solar energy takes up an increasing proportion of energy sources due to its renewable,environmental protection and other advantages.The tower-type solar thermal power generation system has the advantages of large concentration ratio,large capacity,high efficiency,etc.,so it has important commercial prospects and has received more and more research.The tower-type solar thermal power generation system uses solar receiver to focus the direct normal irradiance(DNI)on the system solar receiver to absorb and convert to complete power generation.The occlusion and departure of clouds are the main factors that cause the direct normal irradiance change of the sun.When the cloud suddenly enters the field,the direct normal irradiance received by the heliostat field changes suddenly,and the temperature decreases,which causes the heat storage medium to solidify and may cause pipe blockage or even shutdown,which greatly reduces the energy utilization rate of the system.In addition,the temperature of the wall surface of the tube where the storage medium is solidified increases sharply,which may cause damage the heat absorbing device.Therefore,it is of great significance to the tower solar thermal power generation system to predict the cloud cluster and the direct irradiance change trend affected by it and take measures in advance to deal with the changeable weather.This paper analyzes the demand for cloud image prediction and DNI change trend prediction in CSP systems and designs a set of plans to realize cloud image prediction and DNI trend prediction.First,preprocess the total sky image to obtain the input sequence images required by the total sky image prediction model proposed by this paper,which is sent to the images prediction model to predict sequence cloud images,and the performance of the model under different hyper-parameters and different weather conditions is discussed,then verifies the effectiveness of the long-term prediction of the model.Then locate the center of the sun in the predicted image according to the sun-earth model.Finally,taking the sun position as the center,intercept the region of interest(ROI)of the target size and use the deep neural network to extract DNI to complete the DNI trend prediction.The main research contents of this paper are as follows:(1)In this paper,3D convolution is incorporated into the long short-term memory(LSTM)model,and combined with the Attention mechanism,a spatiotemporal network model is designed to realize the prediction of the total sky sequence images.Different from the previous digital image processing technology based on feature engineering to predict cloud images,this model can realize end-to-end prediction cloud images.(2)Research on the performance of the sequence image prediction model.Specifically,the research is conducted from the following three aspects:the structural parameter analysis of the model,while verifying the prediction limit of the model under a certain input;discussing the performance of the model under different weather conditions;verifying the reliability of the model’s long-term prediction.(3)Introduce the application of the predicted sequence full sky map-DNI trend prediction.In this paper,we design an algorithm to accurately calculate the position of the sun in the whole sky image.In the predicted whole sky image,ROI of the target size is cut with the sun position as the center as the feature map,and the ResNet model is used to predict the DNI trend end-to-end.Experiments show that the whole sky sequence cloud image prediction model proposed in this paper can predict the whole sky image for long time,and the model performs well.At the same time,the forecast of the whole sky cloud map can be used to predict the corresponding DNI value.Experiments show that the predicted DNI change trend is basically true.Even the error caused by the machine measurement in a few locations,the forecast results based on the cloud map in this paper are more realistic.