Solar Irradiance Prediction Based On Numerical Data And Images Fusion

Solar energy,characterized by its abundance,pollution-free nature,and wide distribution,has been widely applied in various energy-related fields as one of the primary clean energy sources.However,the fluctuating and intermittent nature of solar radiation,which is the source of solar energy,is detrimental to the grid integration of photovoltaic power stations and the scheduling,operation,and maintenance of power systems.As a result,accurately forecasting solar radiation is an extremely important task,holding significant implications for the power system.Solar radiation is susceptible to numerous meteorological factors,such as cloud movement,aerosols,and optical air quality,making it challenging for general-purpose time series models to achieve precise predictions without targeted design.In this paper,A neural network prediction model that fuses meteorological numerical data and ground-based cloud images was designed.To enhance the effectiveness of numerical data,the model was build based on an automatic modal attention mechanism to directly predict solar radiation while filtering and analyzing various meteorological data.To improve the effectiveness of image data,a two-stage fast partial convolutional network,which takes physical and chemical properties of the atmosphere into consideration was proposed to reconstruct ground-based cloud images.Finally,by constructing a modal decomposition attention module with strong data feature extraction capabilities,time series images and data are fused to predict solar radiation accurately.The specific research content can be summarized in the following aspects:(1)A statistical analysis was conducted on the dataset provided by the National Renewable Energy Laboratory’s Solar Radiation Research Laboratory to limit the data volume that was input into the prediction model.It also increased the proportion of effective information in the input data.Meteorological numerical data were classified and grouped into categories,including clouds,wind,temperature and humidity,atmosphere,and sun.Since ultra-short-term solar radiation prediction is mainly influenced by cloud movements,different meteorological data combinations were formed within each group under various cloud cover levels.An Automatic Modal Attention Mechanism(AMAM)based network was proposed for selecting meteorological and solar radiation variables,with fixed training parameters.Each data combination was used to directly predict solar radiation through the network.By using prediction accuracy as a criterion,effective combinations were filtered out,thus eliminating irrelevant meteorological data with less useful information for the study’s application.Further analysis and fusion of these combinations,along with cloud cover,served as optional inputs for the subsequent prediction model.(2)Addressing the limitations of existing all-sky imaging devices,where the clouds near the sun’s position are obscured by a sun shade to protect the image sensors,which could limits the accuracy of ultra-short-term solar radiation prediction models based on ground-based cloud images.In this study,a fast supervised attention module and momentum embedding were proposed to reflect the time series characteristics of images,combined with a cyclic residual partial convolution algorithm to construct a ground-based cloud image restoration and reconstruction network.The input consists of ground-based cloud image time series and effective meteorological data time series,while the output is the reconstructed image after the removal of sun shade obstructions.The reconstructed images,to some extent,would contain genuine and effective meteorological data,serving as inputs for subsequent prediction models.(3)Considering the time series characteristics of solar radiation,as well as its periodicity and randomness,a Modal Decomposition Attention was proposed,which fuses the automatic modal attention mechanism for numerical time series structures and ground-based cloud image sequences.Adopting a decoder structure similar to Autoformer and based on the encoder network constituted by the Modal Decomposition Attention,an image attention side path containing ground-based cloud image features was added,constructing a numerical-image fusion solar radiation prediction model.This model is designed to enhance the ability of network to learn relationships between different time instances,achieving accurate predictions of total solar radiation,and ensuring that the predicted total radiation is entirely based on historical data.The ground-based cloud image sequences are input into the side path added outside the decoder,integrating the intermediate features of each Modal Decomposition Attention module.The final output is combined with the output of decoder to participate in the feedforward network,forming ultra-short-term solar radiation predictions.Experiments showed that the prediction accuracy of the proposed model was better than the persistent model,the model without reconstructed ground-based cloud images as inputs,as well as Siamese CNN-LSTM that fused the data in a single-layer.