Research On Inter-hour Forecast Of Direct Normal Irradiance Based On Ground-based Cloud Images

With the continuous consumption of traditional fossil fuels and the environmental pollution and global warming,the abundant and clean solar energy has attracted more and more attention.In recent years,solar energy has been developed rapidly all over the world,and its proportion in the power grid has become larger and larger.However,the solar power is intermittent and fluctuant.The larger the proportion is,the greater the risk is to the power grid,when predicting the solar power inaccurately,which limits the application and proportion of the solar power in the power grid.Solar radiation is the most important factor causing the intermittence of solar power.Therefore,the vital thing is to predict solar radiation precisely in order to predict the solar power generation.The solar radiation forecast is the necessary demand of the design and operation of solar energy system,the premise of power grid dispatching decision,and the foundation to guarantee the security of the power grid.Therefore,it is an urgent task for solar power grid-connected to realize the regional,high-precision and full-time solar radiation prediction,and establishing the prediction system of solar radiation that is of great significance for promoting renewable energy generation,which will improve the market competitiveness of solar power.The research contents of this paper can be summarized into the following five aspects:(1)From the perspective of solar radiation transmission,the statistical analysis on the global monthly means of direct normal irradiance(DNI)in the CERES database were carried out to research on its spatial and temporal distribution characteristics,which would provide the theoretical basis of data selection for constructing the DNI forecasting model.On the other hand,a novel method based on neural network(NN)has been proposed to quantitatively analyze the influence of meteorological variables on radiation from the perspective of climate feedback in atmospheric science.The NN-based method not only overcomes the two shortcomings of the published the Kernel method in the atmospheric field in calculating cloud feedback directly and estimating the non-linear relationship in climate feedbacks,but also realizes the more accurate calculation and decomposition of the radiative feedback.Taking the experiments to analyze the effectiveness of meteorological variables on DNI by the NN-based radiative feedback method,the results show that meteorological variables have different effects on DNI for different observation stations,and that aerosols play a major role in DNI under the clear sky,while clouds play a major role in the all sky,and different components of cloud have different effects on DNI.(2)An adaptive semi-empirical hybrid clear-sky model has been proposed in order to bal-ance the accuracy and flexibility of the clear-sky methods for estimating the clear-sky DNI.The inputs of the proposed clear-sky model are adjusted according to the meteorological variables available at a specific observation station,and the model parameters are adjusted based on the historical measured data,which reduces the application restriction and improve the forecasting accuracy.The measured data from five observation stations in the United States were used to validate the performance of the proposed method,and the experimental results show that the adaptive semi-empirical clear-sky model is able to estimate the clear-sky DNI in a variety of climates,and that its forecasting performance is better than the state-of-the-art published clear-sky models.These results provide the great theoretical values for subsequent cloud recognition,cloud classification and DNI prediction in all sky.(3)Since cloud is the most influential factor on the DNI in the all sky,it is added into the inter-hour forecasting model of DNI with the ground-based cloud image(GBC image)taken by the total sky imager(TSI-880).According to the characteristic of GBC images and the demand of the photovoltaic power generation forecast,a pre-processing method of the GBC image has been proposed to overcome the defects of occlusion and distortion and a strong light source of sun in the GBC image.Then,a local threshold segmentation method for cloud detection has been proposed by analyzing the features of sky condition in every region of GBC images.Experimental results show that the proposed method is able to detect cloud under various sky conditions and it does better than the traditional method in cloud detection.The GBC image preprocessing effectively restores the distribution of clouds in the sky,and the local threshold segmentation method detects clouds accurately,which provides valid research resources for DNI forecast.(4)Since the attenuation effectiveness of different clouds on DNI,a DNI forecasting method based on cloud classification(CC-SVM)has been proposed in all sky.Firstly,a neural network classer was constructed to classify the current sky based on cloud physical characteristics and attenuation of DNI into five classifications:clear sky,cirrus,stratus,cumulus and nimbus(nimbostratus and cumulonimbus).Next,five forecasting sub-models were constructed with historical measured data for each cloud type,and the cross-validation method was used to learn the sub-model parameters.The experiments were carried out with data from the open database(NREL).The results show that cloud classification contributes to improve the performance of the CC-SVM model in predicting the inter-hour DNI and that its forecasting accuracy is comparable to the state-of-the-art published methods.(5)To overcome two defects of the CC-SVM model:the rationality of artificially designed image features and the ignoration of the transformation between different types of clouds,a Deep Learning method(CNN-RBF)has been proposed to predict the inter-hour DNI in all sky.Taking the advantage of the learning skill of Deep Learning,the operations,such as convolution,regularization,pooling and so on,are used to extract image features automatically,and the full-joined operation is used to fuse image information and measured data in order to solve the problem on the image-numerical multi-type data resource.The experimental results show that the CNN-RBF model achieves the greater prediction performance compared with the CC-SVM model and realize the precise inter-hour DNI forecast under various sky conditions.