Adjustment Of Atmospheric Profiles Based On Deep Reinforcement Learning

Atmospheric profiles include temperature,humidity,wind speed,etc.,are the main support for atmospheric scientific research such as numerical weather forecasting and typhoon trajectory prediction.Restricted by factors such as observation facilities,there are various degrees of errors in the measured atmospheric profiles,and these errors will reduce the accuracy of atmospheric scientific research.In order to reduce the influence of atmospheric profile errors,in numerical weather forecasting,data assimilation methods such as variational and filtering are used to fit multi-source data such as atmospheric profiles and brightness temperature in a certain space-time region,thereby improving the forecast accuracy.The correction of the atmospheric profile error by the assimilation method is indirect,and the corrected atmospheric profile cannot be extracted for other purposes,so there are certain limitations.The rise of deep reinforcement learning provides new ideas for reducing atmospheric profile errors,and its powerful decision-making ability can guide the adjustment of atmospheric profiles.Based on this,this thesis proposes an atmospheric profile adjustment method based on deep reinforcement learning.temperature data,and adaptively adjust the atmospheric profile.Adjustable atmospheric profiles include wind speed in U,V directions,60-level data for temperature and 60-level data for humidity.The deviation of the CRTM brightness temperature converted from the adjusted atmospheric profile will gradually decrease.The ideal goal of adjustment is to reduce the brightness temperature deviation as much as possible with the smallest adjustment amplitude possible.Since this method uses the ATMS brightness temperature as a reference,that is,it is assumed that there is no error,the CRTM brightness temperature converted from the adjusted atmospheric profile is closer to the ATMS brightness temperature,indicating that the adjustment of the atmospheric profile is more successful.In the above steps,an action network is designed in this thesis,and an atmospheric profile parameter is selected as the adjustment object in each cycle.At the same time,an evaluation network was designed to score the adjusted atmospheric profiles,thereby enhancing the credibility of the assimilated atmospheric profiles.Experiments show that after using this adjustment method on the test set,the brightness temperature deviation of22 channels is reduced by 2.37 K on average.