Satelite Retrieval Model Of Solar Irradiance Based On Machine Learning Algorithm

As the low-carbon economy develops,renewable new energy represented by wind power and photovoltaics plays an increasingly important role in the energy landscape.However,the production and measurement of new energy such as wind power and photovoltaics are easily affected by environmental factors such as wind speed,wind direction,sunshine,temperature,and air pressure,which pose problems for the operation efficiency and equipment safety of power generation equipment and grid connection.Therefore,accurate prediction and simulation of power generation are particularly important.The most important factor affecting photovoltaic power generation is surface solar radiation.Highprecision radiation data can help the grid dispatch system to adjust and optimize power generation plans reasonably,greatly improving the peak regulation capacity of the grid.At present,photovoltaic power stations in China are rare and unevenly distributed,and some observation data have low data accuracy and many missing values.In order to effectively improve the data accuracy of solar irradiance,this study introduces a variety of machine learning models,and through comparative analysis,obtains models suitable for single-point verification and cross-validation scenarios of solar radiation.Current research has shown that machine learning models can well simulate solar radiation and explore the mathematical relationship between solar radiation and various influencing factors.Based on six machine learning models including Support Vector Regression(SVR),Random Forest(RF),Bayesian Ridge Regression(BR),Gradient Boosting Decision Tree(GBDT),BP neural network(BP),and Deep Belief Network(DBN),this study uses real satellite data and ground observation data from four photovoltaic power stations located in Henan from January 2020 to April 2021 as the dataset for the following research:(1)Through comparative analysis of four feature selection methods including correlation analysis,LASSO regression,importance analysis of random forest,and RFE feature selection,the features used as inputs for machine learning models are finally determined to be shortwave radiation data(SWR),solar zenith angle(SoA),aerosol index(AE),aerosol optical thickness(AOT),temperature(TEM),pressure(PRS),humidity(RHU),and cloud optical thickness(CLOT),and sensitivity analysis of variables is completed.(2)By constructing an hourly dataset,the simulation effects of five machine learning models including SVR,RF,BR,GBDT,and BP under a single-point verification scenario are compared and analyzed.The study found that the RF simulation effect is the best,with a minimum root mean square error of only 84.49 W·m-2.(3)By constructing an hourly dataset and a machine learning simulation result dataset,the simulation effects of DBN coupling model and single machine learning model under cross-validation scenario are compared and analyzed.The study found that under the cross-validation scenario,BP simulation effect is the best among single machine learning models,and the minimum root mean square error is only 94.85 W·m-2.Compared with the single model,the DBN coupling model has better effect,with a minimum root mean square error of only 86.01 W·m-2.(4)By constructing a daily average dataset,the simulation effects of empirical model A-P and five machine learning models are compared and analyzed.The study found that the simulation effect of the machine learning model is significantly better than that of the empirical model.Among them,the RF simulation effect is the best,with a root mean square error of only 33.40 W·m-2.This study concludes that compared with traditional physical models,using machine learning models to simulate solar radiation is feasible and effective.Among them,RF is more suitable for single-point verification with less data among single machine learning models.The research results have important implications for improving the accuracy of solar.