Study On Flood Discharge Of Three Gorges Reservoir And Hydrological Response Law Of Middle And Lower Reaches

After the completion of the Three Gorges Reservoir,the flood control capacity of the middle and lower reaches of the Yangtze River has been greatly improved.There are many studies on the change of downstream water level/flow caused by the discharge of the Three Gorges Reservoir,but there are few studies on the catastrophic flood on the water level and flow in the middle and lower reaches of the Yangtze River.This paper focuses on the impact of floods caused by the operation of the Three Gorges Reservoir on the downstream.The M-K trend analysis method,moving average method and wavelet analysis method are used to analyze the hydrological situation changes and periodic characteristics of Yichang,Chenglingji and Hankou from 1952 to 2020.The HEC-RAS one-dimensional hydrodynamic model is used to discuss the simulation effect of distributed hydrological model in conventional flood and catastrophic flood.The random forest(RF)model,convolutional neural network model(CNN),long short-term memory model(LSTM)and convolutional-long short-term memory coupling model(CNN-LSTM)were used to predict the flow/water level.The optimal neural network model was selected to explore a method for quickly estimating the water level amplitude of the downstream Zhicheng,Shashi and Jianli caused by the Three Gorges discharge.It provides technical support for flood control optimization dispatching,emergency rescue and dispatching impact analysis of Three Gorges Power Station.The main research contents and results of this paper are as follows:1.The catastrophic flood is defined.The outflow of the Three Gorges from 2008to 2020 was analyzed,and 2010,2012 and 2017 were determined to be the year of catastrophic flood.2.The hydrological situation changes of the main stations in the middle and lower reaches of the Three Gorges are analyzed.Using the daily flow and daily water level data of Yichang station,Chenglingji station and Hankou station from 1952 to 2020,combined with M-K trend analysis method and moving average method,the variation trend and characteristics of annual average flow,annual average water level,flood peak flow and 3-day flood volume of the above three stations are analyzed.The results show that the average annual flow of Yichang station and Chenglingji station shows a decreasing trend,and the average annual water level of Yichang station has a significant downward trend.The annual average flow change trend of Hankou station is stable,and the annual average water level has shown an upward trend since 1990,but the trend is not significant.The peak flow of Yichang station and Chenglingji station showed a decreasing trend,and the decreasing trend of Chenglingji station was significant for a long time.The peak flow of Hankou station showed an increasing trend,but not significant.The 3-day flood volume has a similar trend to the flood peak.The wavelet analysis shows that the annual average flow and flood peak flow of each station have a main period of more than 20 years.The main periods of the annual average water level of Yichang station are quasi-19-year change and quasi-29-year change.The main periods of annual average water level in Hankou station are quasi-24 years and quasi-34years.3.The HEC-RAS one-dimensional hydrodynamic model is used to simulate the flow process of Hankou station.The efficiency of HEC-RAS one-dimensional hydrodynamic model was evaluated by four evaluation indexes:Nash efficiency coefficient(NSE),Klein efficiency coefficient(KGE),root mean square error(RMSE)and symmetric mean absolute percentage error(SMAPE).In conventional flood years,the simulation effect of HEC-RAS is better,NSE is between 0.70 and 0.98,KGE is between 0.74 and 0.97,RMSE is mostly within 0.2 to 0.3,and SMAPE is mostly within5 to 10.In the catastrophic flood,the simulation effect is poor and the flow process cannot be simulated well.4.Based on the neural network model,the water level/flow of Hankou station is simulated.In this paper,four evaluation indexes of NSE,KGE,RMSE and SMAPE are used to compare the prediction efficiency and stability of four models:random forest model(RF),convolutional neural network model(CNN),long short-term memory network model(LSTM)and convolutional-long short-term memory coupling network model(CNN-LSTM).The results show that the coupled model CNN-LSTM is significantly better than the single model in the simulation of water level and flow.It is suitable for the prediction of water level flow.5.The method of rapid estimation of water level variation is discussed.There is a strong linear relationship between the flow variation of the Three Gorges Reservoir and the flow variation of Yichang station.The flow variation of Yichang station is divided into three levels:3000～5000m~3/s,5000～7500m~3/s,7500～10000m~3/s.It is found that the water level variation range of Zhicheng station is 0.5～1.3m,0.9～1.6m and 1.5～2.0m respectively,and the water level variation of Shashi station is 0～1.5m,0.3～1.5m and0.7～1.5m respectively.However,the correlation between the amplitude of water level change and the magnitude of flow change at Jianli and Hankou stations is gradually weakened,and the law of water level in the year of catastrophic flood is not obvious.