Research On Precipitation Nowcasting Based On Radar Echo Extrapolation Algorithm In China

Precipitation nowcasting can predict the future high resolution rainfall in a local region in 0?3 hour,which is of great significance for early warning of disasters.Precipitation nowcasting based on radar echo extrapolation algorithm is widely used,but it is still a challenging task to achieve real-time and highly accurate prediction results.Firstly the real-time problem of the forecast system requires that one prediction task should be completed within 6 minutes.When the forecast coverage is large,the total time of radar echo extrapolation based on optical flow method and quantitative precipitation estimates is more than 20 minutes,making the forecasts fall behind observations.Secondly,in the radar echo extrapolation,the echo extrapolated by Horn Schunck algorithm is prone to generate noise making poor extrapolation accuracy.Then in the quantitative precipitation estimation,because of the rapid change of the meso-micro scale weather system and the large scale of the observation data,the estimation algorithm based on the off-line model can not satisfy the requirements of the real-time and efficiency.As a kind of streaming data,radar echo data and station observation data are suitable for online learning to update the model.In this paper,three aspects are studied: radar echo extrapolation,quantitative precipitation estimation and product parallel output.The main research work includes:Firstly,In this paper Horn Schunck and Lucas Kanade algorithms are introduced into radar echo extrapolation prediction.Specifically,the Gaussian convolution template of Lucas Kanade algorithm is added to Horn Schunck algorithm to solve the problem of dense optical flow and reduce the noise in the extrapolated echo.In order to reduce the computation time of optical flow field,the multigrid method is used to solve the optical flow equations and the multithread parallel technology is used for optimization.The results show that the average time of solving the 4200x6200 km equations by multigrid method is about 43 s,and after parallel optimization,it is about 25 s,which meets the requirements of echo extrapolation for computing time efficiency.Secondly,In the quantitative precipitation estimation,the online method is introduced to replace the offline model,and the precipitation estimation model based on the online random forest algorithm is established.Experimental results show that on the one hand,the model based on the online random forest algorithm is better than the random forest model in the low and medium rainfall intensity prediction,on the other hand,it greatly reduces the training and prediction time of the model.Finally,A precipitation nowcasting system over large area is realized and the output time bottleneck of the product is optimized by parallel computation.Using MPI parallel technology,the output tasks of multi time extrapolation products are assigned to different nodes.Because there is no dependency between the extrapolation tasks,a linear relative speedup can be achieved.The results show that: 1)the Probability of Detection of radar echo extrapolation prediction are basically kept above 0.5 within 1 hour and kept at 0.7 within 30 minutes;2)the Probability of Detection of one hour precipitation forecast are kept above 0.45 and the Probability of Detection are kept above 0.51 when the precipitation intensity greater than 10mm/h.