Observation Based Deep Learning Model for Short-duration Heavy Rain Nowcasting

Abstract

Short-duration heavy rain（HR） causes serious disasters.However， because of its local abrupt occurrence and evenly distribution， it is difficult to be nowcasted and warned with lead times of 0~2 hours by traditional extrapolations of radar echos.Radar echo images contain little information about the atmosphere environment which breeds the convective weather.To overcome the deficiencies of existing methods， two variables are selected and input from ground meteorological observations which is in enough high spatial and temporal resolution due to the large growing number of regional automatic weather stations build in recent years.Based on dew point temperature（DPT） and 3 hours of pressure change range（PCR） 2 meters above the ground and Doppler radar echo images from major flood period of 2016 to 2019， a new deep learning model is build for predicting HR in 0~2 hours by combining Convolution Neural Networks（CNN） and Long Short-Term Memory（LSTM） networks.Data quality control methods include that， PCR is standardized to be non-diurnal variation， convolution filter is used to filter off speckle noise and clutter in radar echo pictures.Feature engineering include that， DPT and PCR which are selected as most important factors rank top two through the permutation importance diagnosis.Evaluation results from April to September in 2020 show that， Double Channels Deep Learning（DCDL） model which merges radar echos and ground meteorological observations performs better than Single Deep Learning（SDL） model which only uses radar echos and CMA-SH3（Shanghai Meteorological Service-WRF ADAS Rapid Refresh System， SMS-WARR） used frequently in HR nowcasting business.The key reason is that the genesis and destruction of convective weather systems can be predicted relatively accurately by The DCDL model， while they are not forecasted accurately by the traditional extrapolations of present radar echos.The 0~2 hours probabilities of detection（POD） of DCDL model are 46.7% and 39.6%， while the false alarm rates（FAR） are 67.9% and 76.8%， TS scores are 23.4% and 17.1%.As to SDL model， POD are 45.9% and 34.4%， FAR are 69.1% and 77.3%， TS scores are 22.7% and 16.2%.Analysis of two typical examples of HR events in 2020 flood season indicate that， DCDL model performances better than SDL and CMA-SH3 for nowcasting HR events whatever caused by moving convection storms or newborn storms.Therefore， adding physical meaningful observation factors to the HR nowcasting model can develop the accuracy of the deep learning method.DCDL model proposed by this paper could be applied to the future auto-identify system of the meteorology operation.

Keywords

• short-duration heavy rain
• deep learning
• double ensemble network
• cnn
• lstm