Comparative Study Of Solar Radiation Simulation Based On Different Machine Learning Methods On The Loess Plateau

Due to the scarcity of solar radiation observation stations in China,it is difficult to obtain higher-precision solar radiation data by interpolation or extrapolation of radiation stations.There are very few solar radiation observation stations on the Loess Plateau,and only 14 radiation stations are in line with this study.However,high-precision solar radiation data about climate,hydrology,and ecology studies is required and particularly important on the Loess Plateau.Existing research shows that machine learning can simulate solar radiation well,but different machine learning have different simulation accuracy in different regions.In order to obtain higher-precision solar radiation on the Loess Plateau,This paper first analyzes the correlation between meteorological factors and measured solar radiation values at stations,and selects the average pressure,average temperature,average water vapor pressure,sunshine hours,cloud cover,cloud optical thickness,ozone,precipitation water vapor,DEM,slope,and aspect as the input parameters of the model,and then the simulation results and simulation errors of three different machine learning methods: random forest(RF),BP neural network,and support vector machine(SVM)are compared.In order to make the model have more training samples,solar radiation data of 14 radiation sites on the Loess Plateau in 2003~2016 are selected.In order to have more training samples of the model,the measured data of 14 radiation sites on the Loess Plateau in 2003~2009 and 10 radiation sites in 2010~2016 are selected,Remaining measured solar radiation data from 4 radiation sites in 2010~2016 as verification data of the model.The verification results show that RF model has higher simulation accuracy on the Loess Plateau and surrounding areas.In this paper,some existing radiation products with better time continuity including ERA5,NOAA_AVHRR products and CERES radiation products are selected,this four radiation products and the solar radiation results obtained by RF model simulation are further compared and analyzed.Finally,RF model with less error was used to simulate the solar radiation data of 97 meteorological stations on the Loess Plateau.The kriging method was used to obtain high-precision solar radiation data on the Loess Plateau.The temporal and spatial characteristics of solar radiation on the Loess Plateau were analyzed.The main research conclusions of this article are as follows:(1)This paper compares the error indexes of the three models of RF,BP neural network,and SVM during training and verification,and finds that the RMSE of RF model is the smallest during model training,when the model is verified,the average deviation and the RMSE of RF model are the smallest,and the correlation coefficient is the largest.(2)This paper validates the simulation results of the three models including RF,BP Neural Network,and SVM by using measured data of solar radiation.The verification results show that the simulation results of the RF model are closer to the measured data,each error index is small,and the degree of fit of the RF model obviously better than BP Neural Network and SVM.In general,the RF model has the best simulation effect of solar radiation on the Loess Plateau,followed by SVM,and BP Neural Network has the lowest simulation accuracy.(3)By comparing ERA5,NOAA_AVHRR,and CERES-BSAF satellite products and RF model simulation results in 2003~2016,the error indicators show that the accuracy of solar radiation obtained by RF simulation is higher.The ERA5 reanalysis data is next,and the NOAA_AVHRR satellite product and CERES-BSAF reanalysis data have a large deviation.It can be known that the combination of RF model with meteorological data and remote sensing data to simulate monthly mean value of solar radiation is a highly accurate and reliable simulation method,which can effectively solve the problem of missing solar radiation data on the Loess Plateau.(4)Solar radiation showed a spatial distribution on the Loess Plateau in 2003~2016,which is more in the northwest of the Loess Plateau,and solar radiation is less in the southeast the Loess Plateau.The 14-year change trend showed an average slow rise,and the rising trend of solar radiation was more obvious in the eastern of the Loess Plateau.the solar radiation trend of the Loess Plateau showed a trend of rising first and then falling from January to December in 2003~2016.The seasonal change trend of solar radiation on the Loess Plateau showed a slight upward in 2003~2016.And,the solar radiation trend was more obvious in winter,followed by spring and summer,and the upward trend was not significant in autumn.