Research On Prediction Model Of Precipitation Nowcasting In South China Based On Radar Echo Extrapolation

The short-term precipitation forecast generally refers to the precipitation forecast from 0 to 6 hours,and the forecast accuracy can reach the kilometer level and the minute level.With the development of meteorological radar technology,short-term precipitation forecasting based on radar echo extrapolation technology has received extensive attention from the business and scientific communities.However,real-time and refined forecasting still has great challenges.The research object of short-term precipitation forecast is mainly small and medium-scale precipitation systems.For disaster warning,short-term forecast has more practical needs than medium and long-term and short-term forecast.This study uses multi-scale feature extraction technology to establish a data-driven deep learning forecast model for short-term precipitation by studying the extrapolation of radar echoes.The research results are expected to further improve the level of short-term precipitation forecasting.The research work of this paper is as follows:(1)First,considering that the optical flow method is the most basic radar echo extrapolation method,it is now introduced into the forecast of short-term precipitation in South China.Based on the radar echo data in South China,the optical flow method is used to extrapolate the precipitation forecast of the 1-h radar echo in South China;the critical success index(CSI),the correct rate(POD)and the false alarm rate(FAR)are further used to evaluate the extrapolation forecast.The effect of each index under multiple reflectivity is given,and the forecast results are visualized.(2)Because the precipitation itself has the characteristics of randomness,complexity and high nonlinearity,the actual "generation and disappearance" of radar echo has a more complex evolution law.Therefore,the prediction effect of traditional radar echo extrapolation algorithms(such as optical flow method)is not significant.In this study,based on the South China radar echo puzzle data,a radar echo extrapolation model based on Convolutional Long Short-Term Memory(Conv LSTM)was constructed,and the Guangxi region was selected for precipitation forecast verification.The research further compares and analyzes the Conv LSTM and the traditional optical flow method.The results show that the CSI and POD of the Conv LSTM algorithm are respectively 0.12 and 0.11 higher than the traditional optical flow method,while the FAR is decreased by 0.11,indicating that the prediction effect of the Conv LSTM model is better than that of the traditional optical flow method.(3)In view of the serious loss of echo evolution detail information in radar echo extrapolation as the extrapolation time increases,on the basis of Conv LSTM,this study proposes a deep learning based on multi-scale feature extraction and fusion.Short-term precipitation forecast model(Multi-scale feature extraction and fusion convolution network,MSF2).First,the model uses convolution kernels of different sizes to extract features from the shallow information of the network to make up for the shortcomings of single feature detection.Secondly,the feature information of different channels is spliced ??and channel shuffled to enhance the information flow and information expression ability between channels.Finally,the extracted multi-scale feature information is fused to effectively retain the channel information after feature map fusion.Based on the data of the South China radar echo puzzle data,the ablation experiments were carried out under three different radar echo reflectivities,and a comparative analysis was carried out with the two mainstream algorithms Conv LSTM and the traditional optical flow method.Under the condition of multiple reflectivity,the CSI and POD of the MSF2 algorithm are improved by0.056 and 0.06 respectively compared with the Conv LSTM method,and the FAR is decreased by 0.08,indicating that MSF2 performs the best among all evaluation indicators.The experimental results show that the introduction of multi-scale mechanism can improve the feature extraction ability of the model.Compared with the current mainstream Conv LSTM and optical flow methods,the MSF2 model proposed in this study has better applicability and higher prediction accuracy for short-term rainfall forecasting.