Research On Meteorological Downscaling And Short-term Forecasting Based On Deep Learnin

With various environmental pollution,global warming,ozone layer destruction and other reasons,the meteorological environment is becoming more and more unstable.And the frequency of disasters is also increasing in recent years.Frequent occurrence of disasters(such as floods,typhoons,droughts,blizzards,etc.)usually leads to a large number of casualties,property losses and damage to public facilities.In order to reduce the impact of these man-made uncontrollable disasters,the concept of refined forecasting is put forward.The refined forecast in this paper mainly involves the detail of data accuracy and the accuracy of forecast data.The research of this kind of task is to use computer algorithms to obtain data beyond the accuracy of hardware collection equipment,so that the internal data of the rasterized system has the characteristics of more detailed spatial dimension,more dense time scale and more accurate prediction results.Therefore,the corresponding solutions are downscaling algorithm and algorithm of short-term forecast.The main research achievements of this paper are as follows:(1)Aiming at the problems of few time-scaling schemes of existing deep learning and undifferentiated data fusion methods,FC-ZSM(Feature Constrained Zooming Slow-Mo)is proposed,which provides a frame filling scheme based on visual theory for the first time for meteorological spatio-temporal scaling problems.In this scheme,bidirectional deformable Conv LSTM was used to make up for the lack of time sequence information processing,dynamic and static data types were classified,and a dynamic and static data differentiation fusion mechanism was constructed by deformable convolution.(2)In view of the problem that the loss function of the existing downscaling and spatio-temporal super resolution methods is concentrated in the spatial domain,ignoring the frequency domain,the image frequency domain analysis method is introduced,and the frequency domain FL1 L loss function which can obtain better performance than the common spatial threshold loss function(such as MSE,MAE,perceived loss)is summarized.The characteristics of PL1 L phase loss function destroying phase spectrum are expounded and the performance of different loss functions is analyzed in detail.(3)Aiming at the problems of noise introduction,lack of timing processing and redundancy of network structure in the design of UNet which is used in the task of short-term forecast,Sma At-ST-UNet is proposed.In this model,pixelshuffle reconstruction is introduced instead of linear interpolation to eliminate noise.Two-layer small channel Conv LSTM is used to process timing information,and network channel parameters are adjusted appropriately.Finally,we got the network T-UNet with optimal quantitative index,network ST-UNet with optimal heat map effect,and the most lightweight network Sma At-ST-UNet.