Research On Rainstorm Probability Forecast Based On Ensemble Forecast

This research was funded by the National Key R&D Program of China(Project number: 2017YFC1502000).This paper through to evaluated the prediction capability of the ECMWF ensemble forecast model and ECMWF high resolution numerical forecast model in areas of middle part of China,analyzed the advantages and Error characteristics of these models.On this basis,the clustering algorithm commonly used in machine learning is used to divide the middle part of China into 8 regions by the differences in forecasting characteristic,and analyzed the characteristics of regional precipitation forecast and the climate characteristics of graded precipitation.Finally,the traditional Poor Man equal-weight probability forecasting method and artificial neural network algorithm are used to calculate the probability of regional rainstorm,and compared the two methods to the heavy precipitation’s forecasting in different regions.Summarizes the primary coverage which this article studies as follows to show:(1)Firstly,the traditional deterministic prediction tested method and improved standard deviation R value are used to analyze the precipitation forecast capacity of ECMWF high-resolution model,ECMWF ensemble forecast’s control prediction and the ensemble average forecast.The results show that ECMWF high resolution model’s description of observed precipitation’s mean value is rather random.Overall,The concentration ratio is larger in the east of China and smaller in the west of China.The average value of ECMWF ensemble forecast’s control prediction is basically larger than observation.The concentration is higher in the southeast of China,lower in the northwest of China.The average value of ECMWF ensemble averaged forecast basically is higher in the southeast of China,lower in the northwest of China.The phenomenon of high concentration in the eastern region is magnified.Overall,the ensemble control forecast is closer to the observed precipitation under the combined measurement of concentration and mean value.But the ECMWF high-resolution model is closer to observed precipitation in the term of concentration.(2)And through the traditional probabilistic forecast methods of calibration and,conducted the " BS score with precipitation" and the "BS score without precipitation" to avoid uneven sample causes the BS score lose efficacy,evaluated the probabilistic prediction ability of ECMWF ensemble forecast,and The Talagrand distribution was used to analyze the main causes of errors in the forecasting of precipitation.The analysis results show that the southeast of China,north of China and Sichuan Province are the high value areas of the dispersion,the reliability of Precipitation forecast is low.The northwest region of China is a low value area of dispersion,the reliability of Precipitation forecast is high.At the same time,ECMWF ensemble forecast in the middle of China,especially in Sichuan Province and southeast of China is much higher than the standard deviation of Ensemble average forecast,so it is necessary to reduce the ensemble dispersion appropriately to make the ECMWF ensemble forecast system more intensive.And then the single model equal weight probability forecast of ECMWF ensemble forecast is not good for the light rain,although it can increase the hit rate,but also increase the false alarm rate.It performs well in moderate or heavy rainfall,which can make a valuable distinction for the occurrence of precipitation events.On the rainstorm scale precipitation,the false alarm rate is low,but the missing alarm rate is high and the identification of Rainstorm forecast is insufficient.Therefore,it is necessary to explore a new method to integrate the ECMWF ensemble prediction system to obtain more uncertain information in the Rainstorm forecast.(3)By using the scoring results in evaluated the probabilistic prediction and deterministic prediction as clustering factors and using the k-means clustering method divided the middle of china into eight different regions,and the analyzed the characteristic of forecast and graded precipitation characteristics of each region.The analysis results show that region 2 and region 4 is the key area of the northern area of China in rainstorm precipitation forecast,region 7 and region 8 is the key area of southern area in china in rainstorm precipitation forecast,The precipitation frequency above the rainstorm(≥50mm)in these regions is high,and due to the complexity of terrain and climate,the accuracy of rainstorm forecast is low,so it is necessary to further explore the appropriate forecasting methods in these areas.(4)Using the Artificial neural network algorithm to make the regional probability prediction and compared with the forecast results of Poor Man equal-weight probability forecasting method.The analysis results show that the probability forecast result of rainstorm precipitation made by artificial neural network algorithm is more accurate than that made by Poor Man equal-weight probability forecasting method,and the predicted area precipitation is also closer to the observation precipitation.In region 2,4,7 and 8,the prediction reliability is relatively high,so it is suitable to use the artificial neural network algorithm to make probability prediction for the prediction of rainstorm precipitation in these regions.