Analyse Of The Blending Algorithm Based On Radar-Nowcasting And Meso-scale Numerical Weather Prediction Model

As for short-range severe convection nowcasting,the most difficult part is topredict the developing trend of the weather system.Without considering the intensityand shape changing in storm motion,but only depends on radar echo extrapolation,theforecast skill decreased rapidly with lead time.Over longer time scales,the meso-scalenumerical weather prediction model assimilating a variety of high-resolutionobservations would perform better than radar nowcasting methods as theyincorporated dynamic and physical processes of weather system.However,the intrinsic“spin-up” problem in NWP model would reduce the reliability of model,especially inthe first two hours.It becomes feasible to apply the NWP output to optimally mergewith radar nowcast for improved rainstorm predictions in the1to6hours forecastrange. To this end,we conduct our research on three aspects.First, a multi-scaletracking radar echoes by correlation algorithm is proposed to make radar-basedprecipitation nowcasting.We then comparing the short-term precipitation forecastbetween the radar-based extrapolation and the Meso-scale Numerical WeatherPrediction named BJ-RUC to identify their characteristics in precipitationforecasting.The final step would consist of a statistical blending of the two forecaststo produce an optimum rainstorm forecast. Preliminary results are listed as follows:(1)Based on the traditional radar tracking method named TREC(Tracking Radarechoes by correlation),a new radar echo tracking algorithm calledMTREC(multi-scale tracking radar echoes by cross-correaltion) is proposed in thispaper to analyze movements of radar echoes at different spatial scales.Cases studiesshow that：the final MTREC motion vectors are synthesizing the large-scale motion ofthe whole precipitation and small-scale movement of the embedded storm cells so thatits have a more explicit physical meaning.At the same time,generation of zero vectorsand erroneous vectors in the TREC caused by null echo region or rapid changes ofreflectivity within the radar patterns can be avoided in MTREC method,thusingobtaining a smoother and more continuous motion field.Since backward extrapolationscheme could result in obvious biases in echo motion prediction as the decimalfractions of echo pixel positions accumulate with time,in this study,an explicitremapped particle-mesh semi-lagrangian advection scheme is used to extrapolate echoor precipitation fields from time t0to t1.Ten cases evaluations suggested the improvedalgorithm has a better performance than TREC. (2) A meso-scale Numerical Weather Prediction (NWP) called BJ-RUC(AWRF-based Rapid Updating Cycling forecast system of Beijing MeteorologicalBureau) is choosed as model input in blending algorithm.Four rainstorms events inBeijing during the summer of2011were selected to analysis the precipitation abilityof MTREC and BJ-RUC. The results show that: Firstly,the bias of rain-area are bothexisting in MTREC and BJ-RUC, while BJ-RUC is more sensitive to the rainfalltypes and MTREC is more stable than BJ-RUC. Secondly，the forecasting score ofMTREC varies with the type of precipitation event: the larger of the rainfall scale is,the better of the performance shows, while the forecasting ability of local convectiveprecipitation by BJ-RUC is limited. Thirdly,for0-6h preciptation, a cross point existsduring the forecast ability between MTREC and BJ-RUC. The time of cross pointdepends on the rainfall types, asit s more later for large-scale and well-organizedprecipitation. The results above would provide some references for developingblending technic based on MTREC and BJ-RUC.(3)Hourly rain derived from MTREC and BJ-RUC are used to conduct blendingtest with radar nowcasting and NWP model.The result shows that the final QPF afterblending with these two methods are improving compared to both radar extrapolationand numerical weather prediction.It made the precipitation area forecast moreaccurate and also enhanced the precision in the forecasts of rainfall intensity.Thealgorithm is effective and feasible,having a certain reference value in the short-timeforecasting and nowcasting research. With appropriate parameter at both ends ofweighting curve,the merging result can reduce missing prediction rate caused by radarnowcasting in the first2hours.2hours later,the blending result largely depend on themodel output,the preciser the model performs，the better the result will be.A maincontributer to model QPF error is the spatial location error.As for this,a phasecorrection technic based on FFT(Fast Fourier Transform) and MOVA(Multi-scaleoptical flow by variational analysis) is used to deduce the phase correction vectorfield, then applied to Model QPF in the subsequent forecast hours.In our cases, thededuced phase correction vector field cause somewhat over correction.At the sametime, an “object_based” phase correction scheme is tried to comparing with originalone.It seems to be more accurate in phase correction，however, its stability andreliability also requires a large and long-term system validation.