Estimation Of The Surface-layer Atmospheric Optical Turbulence Intensity Based On Machine Learning And WRF Model

When propagating in the atmosphere,laser is affected by atmospheric turbulence,resulting in turbulence effects such as beam wander,beam spread and scintillation,which seriously restrict the performance of various laser systems in atmospheric environment.The refractive index structure constant C_n~2 is an important description parameter for the intensity of atmospheric optical turbulence.The acquisition of C_n~2 values by means of instrumentation measurement not only costs a lot of resources,but also meets many difficulties in measuring the underlying surface of complex situations.Therefore,the research on the parametric model of optical turbulence for the estimation of C_n~2 from conventional meteorological parameters has received much attention.However,for the estimation of surface-layer C_n~2,the Monin-Obukhov similarity theory fails in the case of stable atmospheric junction.In view of this limitation,it is of great significance and value to establish a parametric model of surface-layer optical turbulence based on machine learning.In this paper,by using a large number of measured data,the study on estimation of surface-layer C_n~2 in the inland gobi region is carried out based on three machine learning regression algorithms:Back Propagation Neural Network(BP),Support Vector Regression(SVR)and K-Nearest Neighbor(KNN).In the above-mentioned algorithms,three conventional meteo-rological parameters of temperature,relative humidity and wind speed,as well as time,are used as input characteristics,while the output is C_n~2.The training results show that the three estimation models can well show the diurnal variation characteristics of optical turbulence in the inland gobi region in terms of trend and magnitude.The estimation results of the time of strong turbulence at noon,the transition time of sunrise and sunset and the strong turbulence values at noon are in good agreement with the measured results.According to the statistical analysis of lg (C_n~2),the mean absolute errors of the estimated and measured results of BP,SVR and KNN models are 0.27,0.30,and 0.31,the mean square errors are0.13,0.16,and 0.17,and the overall Pearson correlation coefficients are 85%,81%and 80%,respectively.As a conclusion,BP model has the best estimation performance,since the mean absolute error and the mean square error are the smallest and the overall Pearson correlation coefficient is the largest.In order to forecast the value of C_n~2,a preliminary study on C_n~2 estimation using BP model combined with mesoscale atmospherical model WRF is carried out in this paper.The conventional meteorological parameters obtained by numerical simulation of WRF model are taken as input and substituted into the established BP model.The results show that the C_n~2 estimated by this method is consistent with the diurnal variation of optical turbulence in the inland gobi region.According to the statistical analysis of lg(C_n~2),the mean absolute error of the estimated and measured results is 0.33,the mean square error is 0.21,and the overall Pearson correlation coefficient is 76%.It can be concluded that the method for estimating the C_n~2 in strong turbulence and medium-intensity turbulence based on machine learning algorithms and mesoscale atmos-pherical model WRF is feasible.