Research On Model Weather Wind Field Forecast Correction Model Based On Deep Learning

The research of urban regional wind field is of great significance to people’s daily life and the improvement of ecological environment.However,due to its complex geographical characteristics and seasonal differences in thermal characteristics,the numerical model has always been ineffective in forecasting wind field.Therefore,it is particularly important to study the error correction model of model prediction results.In order to meet the needs of regional correction,this paper applies deep learning technology to regional wind field correction based on model prediction and observation data.The research contents and main work of this paper are as follows:(1)Based on the summary and analysis of wind field correction technology at home and abroad,it is found that the current correction method is relatively simple.Most of the meteorological elements with high correlation of wind field are used as model inputs to correct the wind field.These schemes lack the consideration of the spatial correlation of wind field.In this paper,convolution neural network is innovatively applied to regional error correction.(2)Firstly,the original data(the output of numerical model and the data of national observation station)are preprocessed.After data preprocessing,the forecast data of rmaps-st system are analyzed from two aspects: the variation law of forecast error with forecast time and the daily variation law of forecast error.It lays a foundation for the design of the model scheme and the test of the revised results.(3)According to the results of model error analysis,the output of the model is corrected respectively according to the prediction time.In the meantime,due to the fact that the number of observation stations is less than the number of output grid points of the model,and the amount of data is less after grouping the prediction time,a network training scheme based on transfer learning is designed.The correction results are compared with the original error of the numerical prediction system,and some results are obtained.However,due to the defects of the model,the fuzzy phenomenon will appear at the change of wind field.(4)The noise reduction self coding network is improved,and a correction model based on u-net is proposed.In the process of up sampling,the model will fuse the high-dimensional features extracted from the corresponding down sampling,which makes up for the loss of input matrix information in the process of up sampling.The results show that the correction effect is better than the denoising self coding model.