Research And Application Of Cloud Pattern Recognition And Ultra Short Term Direct Solar Radiation Prediction Based On Neural Network

Solar energy is green pollution-free clean energy,solar photovoltaic power generation is the main way to use solar energy.However,due to solar radiation intensity,cloud cover,temperature and humidity and other meteorological factors,ultra-short-term photovoltaic power generation becomes unpredictable,and brings a strong volatility and intermittent to photovoltaic power generation system and network power,which makes the power grid dispatching more difficulty,meanwhile increases the security risks of the power grid.Especially in integrating into the grid for the large-scale photovoltaic power station,it is particularly important to accurately predict the photovoltaic power generation for the stable operation of the grid.For ultra-short-term or real-time projections,variation in solar irradiance are largely dependent on cloud changes.The analysis of the distribution and type of cloud by the whole sky image imager on the ground is helpful to the prediction of the short-term power generation,especially for the prediction of the abrupt change of the radiation caused by the cloud shelter,which has high application value.In order to solve the problem of ground cloud image recognition,this paper introduces the deep neural network algorithm for the first time,establishes the deep convolution neural network model which can extract the feature of the cloud class,and forms the end-to-end training mode to classify the cloud image.As a comparison,we use the characteristics of the cloud image which selected by hands and support vector machine classifier to build the classification model.Found that the deep of convolution neural network model has been significantly higher than the traditional method of identification accuracy.In this experiment,the effects of different depths of convolution neural network structure,training methods and parameters on the experimental results are studied,and the influence of various network operations on network performance is discussed.In the short-term direct solar radiation prediction problem,the single neural network model is built to predict the solar radiation.The prediction results of the continuous model,the linear regression model and the multi-layer perceptron model are compared respectively.And then we proposed a Hybrid-model by combining the cloud-image classification model and the continuous mode.The experimental results show that the hybrid model has higher stability and accuracy than single model.In order to push the model to the practical application,a network application service platform is created.After obtaining the data of the photovoltaic power plant,the platform uses the trained cloud classification model and the solar radiation prediction model to predict the short-term radiation,and gives the early warning information after the analysis.The user can access the platform remotely and use the data to verify and get the results after analysis.This paper mainly introduces the design and implementation process of the application platform,including server deployment,model porting,front-end design,platform security and stability consideration,and gives the result of test.