Application Research On Temperature Forecast And Precipitation Spatial Downscaling In Guangxi Based On Machine Learning

Numerical model forecasting performs numerical calculation and prediction of atmospheric physics,dynamics and thermal processes through supercomputer numerical simulation,aiming at obtaining the results of weather forecasting and climate forecasting.However,due to the mechanical error of the hardware equipment,the initial value error of the atmospheric motion equations and the imperfection of the model,there are inevitably some errors in the forecast results of the numerical model.In order to improve the refinement of numerical model forecasts,this thesis starts with the two most closely related meteorological elements in daily life,temperature and precipitation,hoping to improve the accuracy of numerical model temperature forecasts and reproduce a more refined spatial distribution of precipitation.The main research work is as follows:(1)In view of the fact that a single feature selection method may filter out some potential information during the feature selection process,resulting in poor stability of the results,this thesis proposes a SpearmanXgb hybrid feature selection method,which complements the advantages and disadvantages of Spearman feature selection and XGBoost feature selection.The results show that compared with the above two single feature selection methods,the SpearmanXgb hybrid feature selection reduces the training time of the machine learning model by 19.7%and 10.3%,respectively,and the root mean square error decreases by 0.94%and 0.64%,respectively.(2)In order to compare the correction effects of different machine learning algorithms on the temperature forecast of numerical models,three temperature correction models of Random Forest,XGBoost and LightGBM were constructed in this paper.The experimental results show that the three models have positive correction skills compared with the numerical model,and the average root mean square error has decreased by 7.04%,7.47%and 7.37%compared with the TIGGE data set;The XGBoost model has the best revision effect for the first and middle forecast period(24 h-144 h),and the LightGBM model has the best revision effect for the late model period(168 h-240 h).The study concludes that there are significant spatial differences in the forecast effects of both the numerical model and the three machine learning models,that is,the forecast errors are larger in both southeastern and northeastern Guangxi.(3)This thesis proposes a precipitation spatial downscaling model based on multi-scale and attention-intensive residuals.Aiming at the limitations of singlescale feature extraction in the SRGAN algorithm and the poor effect on image texture details,firstly,multi-scale convolution with convolution kernels of 3 × 3 and 5 × 5 is used instead of a single scale for shallow feature extraction,making full use of the feature information of the low-resolution precipitation image;Secondly,the channel attention mechanism is combined with the densely connected residual block to replace the original residual block,which improves the attention of the network to the important features of the image and enhances the texture details of the image.The results show that the model proposed in this thesis outperforms traditional spatial downscaling methods and similar image super-resolution algorithms in terms of objective evaluation indicators,realizes the spatial downscaling of numerical model precipitation products,and further improves the refinement of precipitation forecasting.