Application Research Of Ensemble Probabilistic Inflow Forecasting Method For Three Gorges Reservoir

The operational hydrologic forecasting in China mainly takes the form of deterministic point estimations since long,which has ignored various uncertainty soureces throughout hydrologic forecasting process.The flaws of deterministic forecasting method are clear that it fails to provide stakeholders with useful uncertainty information and decisions based on deterministic forecasting method are imcomplete from the respect of risk.In recent years,hydrologic uncertain forecasting methods draw much attention and develop fast,of which the most representative one is the Hydrologic Ensemble Probablistic Prediction System(HEPPS).HEPPS can generate ensemble forecasts which fully consider various uncertainty sources and then apply univariate post-processing and multivariate post-processing to the ensemble forecasts to obtain more reliable and accurate probabilistic forecasts,forecasting intervals and ensemble flood hydrographs compared with deterministic forecasts.It has been widely acknowledged nowadays that HEPPS can quantitatively descript the inherent uncertainty of forecasting process and help improve the social and economic benefits of hydrologic forecasts.The main objective and motivation of this thesis is to carry out research on both method and application of HEPPS with Three Gorges Reservoir as case study as well as enrich the current uncertain forecasting techniques.The research findings of this thesis are also expected to provide technical support for TGR's operational HEPPS.The main works and findings are summarized below:(1)Reviews and discussions about the hydrologic uncertain forecasting theory at home and abroad are illustrated detailly.It is pointed out that applying hydrologic uncertain forecasting is the inevitable trend of future hydrologic forecasting.Application progress of hydrologic ensemble probabilistic forecasting methods are introduced and discussed,and the drawbacks of hydrologic uncertain forecasting methods and their possible solutions are also investigated in this section.(2)Hydrologic regime of TGR basin is changed due to TGR's operation.In order to quantatively distinguish the differences between inflow floods and dam site floods,Multi-Inputs-and-Single-Output(MISO)model is used to simulate the inflow series of TGR and a Copula-based extension model is developed to exchange historic dam site flood features into inflow ones.The MISO simulated inflow series of 1960-2002 are validated by comparing the Muskingum routed inflow series with dam site series at Yichang station.Flood frequency analysis(FFA)results reveal that means of TGR's annunal maximum daily flood peak and 3d flood volumes increase 5.58%,3.85%,respectively,and decreased 1.82% and 1.72% for 7d and 15 d flood volumes than dam site ones,respectively.The flood control risk analysis(FCRA)results of TGR show that flood risks of small-and-middle inflow floods increase about 2% while little change for large inflow floods,which is in accordance with the FFA results.The corresponding results can serve as reference for TGR's scientific operation.(3)Ensemble inflow forecasting scheme with 18 members for TGR is proposed,which comprehensively considers the major uncertain sources of inflow forecasting,for instance input uncertainty,model structure uncertainty and model parameter uncertainty.By coupling upstream flood forecasts and precipitation predictions over TGR's interval basin,the 18-member real-time ensemble inflow forecasts of TGR are generated with lead times from 1d to 3d.The NSE modes of the ensemble inflow forecasts are 0.98,0.95 and 0.91 at 1-3d lead times,respectively.The spread of evaluation metrics among ensemble members increase as the lead time increase,which indicates forecasts with longer lead time have larger forecasting uncertainty.It is observed that more inflow observations fall outside the ensemble ranges with the increase of lead time,especially at high-magnitude parts,indicating a decreasing reliability of ensemble inflow forecasts.In this premise,the raw ensemble inflow forecasts cannot correctly guide decision making process and thus call for ensemble post-processing.(4)Two most popular ensemble post-processing methods,namely Bayesian Model Averaging(BMA)with association of data transformation technique and Copula-based BMA(C-BMA)are employed to calibrate the raw ensemble inflow forecasts of TGR.A coherent set of evaluation metrics is used to make result diagnosis.The raw ensemble inflow forecasts of TGR are proven to be unreliable and underestimate forecasting uncertainty with the badly U-shaped Probability Integral Transform(PIT)histograms and the smallest Coverage Rates(CR)at all lead times.While both BMA and C-BMA probabilistic forecasts show better reiliability and can provide more accurate forecasting intervals than the raw ensemble forecasts,C-BMA method performs better with smaller Calibration Deviations(CD)of the PIT histograms and Contiouns Ranked Proability Scores(CRPS)than BMA method.The CR values of C-BMA results are approximate to the given value 90% at 1-3d lead times,which are 85% for BMA method.As a conclusion,post-processing by C-BMA can efficiently remove the probabilistic bias of ensemble spread and provide reliable risk information for TGR's operational managements.(5)A common problem for inflow post-processing is that none of the parametric distribution candidates can have satisfactory fitting effect on the marginal distributions,which hinders reliability of the probabilistic forecasts.To solve this problem,a C-BMA coupling Kernel Density Estimation(KDE)method(KC-BMA)method is proposed and tested.Evaluation results indicate that KC-BMA probabilistic inflow forecasts of TGR achieve better PIT histograms with smaller CD values at 1-3d lead times and meanwhile ensure the other evaluation metrics no worse than C-BMA results,revealing that KC-BMA probabilistic forecasts have better reliability than C-BMA results.It is suggested that KC-BMA method has strong theoratical basis and abundant scope of application in ensemble post-processing field.KC-BMA is flexible for implementation and remains a valid option of C-BMA method.Non-parametric KDE method strongly relys on the data samples,which the conduction of KC-BMA post-processing demands sufficient data for marginal distribution construction.(6)By introducing Ensemble Copula Coupling(ECC)stemming from meteorology to hydrology,multivariate post-processing of TGR's ensemble inflow forecasting hydrographs is carried out to restore the inherent temporal dependency structure of inflow hydrographs.Three different sampling schemes,i.e.Equidistant Quantile Sampling(EQS),Uniform Random Sampling(URS)and Quantile Transform Sampling(QTS)are used to draw samples from the calibrated KC-BMA probabilistic forecasts independently at 1-3d lead times.The raw ensemble hydrographs by different sampling schemes show biased temporal dependency structure compared with TGR's historic inflow observations.After ECC post-processing,the dependency structure of the three sample sets are correctly restored in terms of Kendall rank correlation coefficient(?).Better Energy Score(ES)and Variogram Score(VS)performance are also observed on the ECC post-processed ensemble forecasting hydrographs.Results indicate good ability of ECC method on hydrologic multivariate post-processing and ECC-EQS results show an overall best performance for the case study of TGR.What's more,it is observed that the PIT histograms of ECC-QTS results at 1-3d lead times appear anomalous periodic oscillation and fail to maintain the reliability of marginal distributions due to the inherent differences between atmospheric circulation and hydrological cycle.Thus the QTS sampling method is not recommended for hydrologic multivariate post-processing.