Study Of Precipitation Nowcasting Based On Deep Learning And Radar Observations

Precipitation is an important weather phenomenon,an important part of the water cycle process,has a far-reaching impact on all aspects of people's lives,and is one of the important factors triggering natural disasters,accurate and timely grasp of the upcoming precipitation can avoid economic losses and loss of life,and help safeguard people's lives and property.Meteorological radar can effectively observe precipitation and use a series of radar echoes to forecast precipitation over the next two hours to provide information on the development of precipitation and to make the right decisions in response to the possible effects of precipitation.However,the temporal and spatial characteristics of the precipitation development process are highly uncertain and its trends of change and movement are difficult to predict.The current main prediction method is the optical flow method;however,the optical flow method has some limitations in that it separates the extrapolation of precipitation echoes from the prediction,is not conducive to the determination of parameters,and lacks the full exploitation of the time-space characteristics of the observed historical data.In recent years,as the level of computer hardware continues to rise,deep learning has been developed and has been successfully applied to a large number of scenarios and domains.Therefore,on the basis of summarizing and analyzing the existing relevant studies,we systematically study and analyze the ability of the depth learning method to predict different precipitation intensities and forms under the echo of precipitation generation,development and extinction,using the precipitation data from radar observations and the depth learning method,and compare it with the optical flow method to test the applicability of the depth learning method to forecast with radar data.Therefore,the following studies have been carried out in this paper.(1)Based on the consideration of spatio-temporal information data processing,summarize the temporal definition of short-cycle precipitation forecasting and the various forecasting methods available,select appropriate radar data types and methods,analyze and transform the radar data to construct a grid data set of radar-observed precipitation sequences in preparation for the further completion of short-cycle precipitation forecasting studies.(2)To improve the prediction accuracy and model generalizability,based on the theory of the neural network approach and its operational steps,to update the optimized neural network parameters by applying the grid data set of the established radar precipitation observation series so that they can learn and capture the precipitation characteristics of the region.(3)Using quantitative analysis methods commonly used in the field of hydrometeorology,the model is compared with the optical flow method to test the ability of neural networks to capture and predict precipitation events and to verify the feasibility of short precipitation prediction based on depth learning and radar observations.Validation of the neural network model and light flow by two-for-two precipitation processes leads to the following conclusions.(1)Convolutional long and short-term memory neural networks predict the dissipation of precipitation echoes,but the predicted echoes dissipate more rapidly and thus differ somewhat from actual observations,while the optical flow method retains most of the precipitation echoes that would have passed over time.(2)Convolutional long and short-term memory neural networks are more effective in predicting the situation in the main precipitation areas,with the predicted echoes maintaining the shape of the main precipitation areas and their boundary contours as the forecast time increases,while the optical flow method has difficulty maintaining the shape of the boundary.(3)Convolutional long-and short-term memory neural networks have a high detection rate for precipitation and are much higher than the optical flow method with increasing forecast time.(4)In the forecasts of two randomly selected precipitation processes,the correlation between the forecasts and actual observations for both methods over one hour remained above 60%.