Short-term Solar Irradiance Prediction Based On Spatiotemporal Graph Convolutional Recurrent Neural Network

Solar energy is a kind of clean and renewable energy.More and more countries and regions take it as an important direction to develop renewable energy.Because solar irradiance is an important parameter for solar power generation,it is important to predict solar irradiance accurately for the reliable operation and sustainable development of solar energy.However,most of the current studies are based on time series prediction models to predict short-term solar irradiance,ignoring the spatial correlation between sites.This method cannot make full use of the correlation information between the sites in the region,thus affecting the accuracy of short-term solar irradiance prediction.To solve the above problems,a spatiotemporal series prediction model combining Graph Convolution Network(GCN)and Long Short-term Memory Network(LSTM)is proposed to integrate spatiotemporal information for short-term solar irradiance prediction.The model uses a convolution network to extract spatial correlation between sites in an area,and a long-term and short-term memory network to extract temporal correlation of solar irradiance series.Compared with traditional time series prediction models,this method makes full use of spatial correlation information and improves the accuracy of short-term solar irradiance prediction.At the same time,in order to meet the practical application requirements,a hardware acceleration scheme based on programmable gate array(GA)is proposed,which can guarantee the model performance and improve the computational efficiency.Through this study,we can better understand the spatiotemporal characteristics of solar irradiance,improve the accuracy and efficiency of short-term solar irradiance prediction,and further promote the development of solar power technology.The spatiotemporal series prediction model and hardware acceleration scheme proposed in this paper can provide important reference value for the research and practical application of solar power generation.The main contents of this paper are as follows:(1)In this study,six sites in Tacheng,Yining,Aksu,Urumqi,Yanqi and Kashgar in Xinjiang were selected as the study objects,and they were constructed into an undirected graph structure,where each site acts as a node and the edges between adjacent sites represent the spatial correlation between them.By collecting solar irradiance data and meteorological data from the six sites,we performed a correlation analysis of the meteorological factors affecting solar radiation to determine the input to the model.(2)In this study,we first propose a spatiotemporal graph convolution recurrent neural network(GCN-LSTM)model,which uses GCN to extract spatial correlation between sites and LSTM to extract temporal correlation of solar radiation data.By combining GCN with LSTM,the spatial and temporal information can be effectively integrated to improve the prediction accuracy of solar irradiance.Next,the model is trained using historical meteorological data and solar radiation data,and meteorological factors such as temperature,humidity,pressure,wind speed,solar zenith angle are used as input to the model.Finally,the prediction performance of the model is evaluated under different experimental conditions and compared with other commonly used prediction models,including seasonal differential autocorrelation moving average model(SARIMA),convolution neural network(CNN),recurrent neural network(RNN)and LSTM.The results show that the mean square root error(RMSE)of GCN-LSTM model at six locations is 62.06 W/m~2,which is 9.8%,14.3%,6.9%and 3.3%lower than other models,respectively.The mean absolute error(MAE)is 25.38 W/m~2,which is 27.7%,26.5%,20.1%and 11%lower than other models,respectively.The average coefficient of determination(R~2)is 0.94,which is 1.4%,2.2%,0.8%and 0.4%higher compared with other models,respectively.This indicates that the spatiotemporal graph convolution recurrent neural network model can significantly improve the prediction accuracy of short-term solar irradiance,which provides strong support for the practical application of solar power generation.(3)The GCN-LSTM model proposed in this study has good performance and application prospects in short-term solar irradiance prediction.However,due to the high computational complexity of the model,the large amount of data and the large scale of parameters,it will face the problems of high-power consumption and low efficiency when deployed to actual photovoltaic power stations.Therefore,this paper presents a hardware acceleration scheme based on the field programmable gate array,which uses fixed-point arithmetic and pipelining technology to optimize computational performance.The results show that the calculation speed and efficiency of GCN-LSTM model can be significantly improved by using hardware accelerator,and the prediction accuracy can be guaranteed,which provides the feasibility for the application of the model in practical photovoltaic power stations.