Research On Meteorological Forecast System Based On Deep Learning Super Resolution

Weather forecasting is the estimation and recording of climatic and meteorological changes over a certain period of time,and weather data is the information output in the form of matrix values.The denser the distribution of coordinate points in an area within the matrix,the more refined the weather forecast for that area,and the weather refinement forecast data can provide richer weather forecast information within a certain area.The refined forecast designed in this thesis is a study of super-resolution forecasting in the spatial dimension,which is the downscaling of low-resolution(low-precision)meteorological observation data in the spatial dimension to generate high-resolution(high-precision)meteorological forecast information,expecting to improve the meteorological forecast accuracy from the provincial and municipal(low-precision)level to the township(high-precision)level.Also,this thesis tries to propose a meteorological refinement method in the time dimension.The method used to achieve refined forecasting in the field of meteorology is the downscaling technique.Traditional downscaling techniques are implemented using interpolation algorithms or statistical models,where the value of each pixel point on the generated image is constructed as a distance relationship between several pixel points around it,or is calculated by learning from a specific distribution of low-and high-resolution precipitation data.However,the high-resolution images mapped into based on such methods do not retain the details of the original images well.Meanwhile,it is found that meteorological observation data is a kind of spatially structured information,which is vulnerable to the influence of different meteorological elements.To address some problems in the field of meteorological spatial refinement forecasting,this paper conducts an in-depth study.1)When using downscaling techniques to build weather refinement forecasting models,the raw weather data must first be processed into weather images that can be trained by downscaling forecasting models.Meteorological observation data is a kind of spatially structured information,and refined meteorological forecasts are easily influenced by a variety of related meteorological elements.In this thesis,we only study precipitation refinement forecasting,and after several sets of comparison experiments,we find that considering the effects of three meteorological elements,namely temperature,humidity,and topography,can significantly improve the final precipitation refinement forecasts.We also use convolutional neural networks to preprocess the raw meteorological data to generate meteorological pictures that can be used for training of downscaled forecasting models,and finally obtain a standard dataset applicable to the precipitation domain.2)Using different upsampling layers may introduce tessellated streak noise to the generated images.In order to ensure the quality of the generated images,this paper also does a comparative analysis of several common upsampling layers and finds that the pixel reorganization method has better results.3)In order to recover the details of the original image as much as possible,this paper proposes two spatial downscaling forecasting methods based on deep learning super-resolution.The first RDBLap model introduces Laplacian structure and residual learning mechanism,and although it still uses root mean square loss error to optimize the model,it has made a big progress compared with the traditional algorithm.The core idea of the second Pre GAN model is to train a generator and discriminator network separately,with the aim of making it difficult for the discriminator network to distinguish the high-resolution predicted images generated by the generator from the labeled images.The advantage is that more local images at high textures can be repaired,thus further improving the quality of precipitation refinement forecasts.In this paper,a variety of meteorological downscaling models are compared and analyzed in two types of meteorological standard datasets,and the experimental results find that both RDBLap and Pre GAN models achieve better results in all evaluation indexes.Among them,the Pre GAN model can achieve more accurate meteorological high-precision forecasting compared with the RDBLap network.