Short-term Ensemble Flood Forecasting Based On Numerical Weather Forecasting

Human activities and climate change have led to frequent occurrence of extreme hydrometeorological events.For example,we are affected by flood seriously,especially in eastern areas,which usually causes large economic loss.With rapid development of computer,meteorology and other disciplines,the ensemble forecast of coupled numerical weather forecast and hydrological forecast model has become an important means of flood control and disaster reduction.In this thesis,by evaluating the performance of various deep learning models,a variety of numerical weather forecast data were processed after precipitation.Coupled with the numerical weather forecast products and the three-source Xin ’anjiang model,an ensemble flood forecast system was built.In addition,Yonganxi Basin and Shifengxi basin of Jiaojiang River Basin were taken as research objects to analyze and evaluate the ensemble forecast system’s ability to forecast and simulate typical typhoons.The main work and findings of this thesis are as follows:(1)The precipitation data of ECMWF(European Center for Medium Range Weather Prediction),CMA(China Meteorological Administration)and GEFS(National Center for Environmental Prediction)during flood season were evaluated by various indices and EOF analysis.The results show that the EOF analysis results are similar to the CMA-CMOPRH analysis results,but there are some deviations in some areas.The prediction ability of ECMWF is stronger than that of CMA and GEFS in short forecast period.However,the performance of GEFS data is more steady in each forecast period.The original forecast data of ECMWF,CMA and GEFS all have large errors,and the RMSE of the original forecast is generally more than 5mm,and the maximum value is more than 14 mm.The bias span is large,and the index is distributed in the range of-0.5 ～ 1.5.The maximum MAE index is 4mm;the maximum value of CC index does not exceed 0.6 and is concentrated in the range of 0 ～ 0.5.Therefore,the original numerical weather forecast needs post-processing.(2)Using CNN,LSTM and CNN-LSTM deep learning models,the ECMWF,CMA and GEFS numerical weather forecast precipitation data during flood season were processed.The results show that the CNN model can improve the accuracy of the forecast data when the forecast period is 0 ～ 24 h,but the accuracy of the precipitation data is very limited when the forecast period is greater than 36 h,and the index has no obvious improvement.The LSTM model improved the accuracy of precipitation data in each forecast period,but quite limited.The CNN-LSTM model greatly improved the accuracy of precipitation data in each prediction period and in the whole basin.After post-processing,the maximum value of RMSE index decreased from 14 mm to 5mm,and the proportion of descending points under the RMSE index was 100%.CC index rose from the maximum value of 0.6 to 0.9,and the proportion of descending points under CC index was 100%.MAE decreased from 1 ～ 4.2mm to 0 ～ 2mm,and the proportion of descending points under CC index was 100%.Compared with the CNN and LSTM post-processing models,CNN-LSTM has greater improvement capability on indicators,and the performance is optimal in all aspects.Based on the postprocessing precipitation data of this model,the typical typhoon events of "Morakot" in2009,"Fitow" in 2013 and "Lekima" in 2019 were evaluated.The simulation of rainstorm center and maximum rainfall intensity has been greatly improved.After the post-processing of "Morakh" and "Fitow",the accumulated precipitation error of rainstorm center is within 12%.(3)The Yonganxi and Shifengxi basins were divided into sub-basins,and the threesource semi-distributed Xinan River hydrological model was constructed.The NSE,NSE of peak flood discharge and the overall bias of flood volume were obtained,and the parameters of the model were automatically determined by using NSGA II genetic algorithm.Meanwhile,the parameters were manually fine-tuned according to the physical significance of the parameters.Using the CMA-CMOPHR fusion data as the input data of the hydrological model,the parameters of the Xin ’anjiang model in the Yong ’anxi Basin with the overall NSE of 0.93 and the qualified rate of flood prediction of 89.5% with grade A were obtained.And the parameters of the Shin ’an model with the overall NSE of 0.91 and the qualified rate of flood forecast of 89.5% were derived.(4)Based on the three-source semi-distributed hydrological model of Xin ’an model,12 flood events were selected,and the ECMWF,CMA and GEFS numerical weather forecast precipitation data before and after post-processing were used as input.The flood forecasting ability before and after CNN-LSTM post-processing in the Yongan and Sipungxi basins was evaluated by using the index of peak discharge,peak time and flood pass rate.The results show that the model produced much better results using post-processing precipitation data than that without post-processing.The qualified rate of simulated flood of ECMWF forecast precipitation data after post-processing is 83.5%with Grade B,close to Grade A,the qualified rate of simulated flood of CMA forecast data is 71% with grade B,and the qualified rate of simulated flood of GEFS forecast data is 71% with grade B.The simulation performance of ECMWF data after postprocessing is the best.To sum up,the ensemble flood forecast studied in this thesis can provide more accurate flood forecast with longer forecast time and provide good references for ensemble flood forecast applications in other river basins.Ensemble flood forecast in this study can well improve the level of flood forecast and early warning and ensure the safety of people’s lives and property.