Storm Flood Ensemble Probability Forecasting Methods And Application

The reliablity and accuracy of precipitation prediction directly determines the quality of hydrological simulation and forecasting results.The traditional hydrological forecast focuses on observed rainfall in land surface,which has short forecasting time and cannot meet the requirements of reservoir operation and management.With the rapid development of numerical weather forecasting technology in the world,the introduction of numerical precipitation forecasts into the traditional deterministic hydrological forecast has made it possible to extend the prediction period.However,due to the complexity and variability of weather systems,the numerical precipitation forecasts often have errors,which need to be corrected.The traditional hydrological forecast only provides users with a deterministic forecast values without considering the uncertainty of hydrological forecasts.It is urgent to expand from deterministic forecasting to probabilistic forecasting which can provide more risk information for reservoir scheduling decisions.In this paper,the ensemble probability flood forecasting is carried out based on the correction of the numerical precipitation forecasts.The main research works and conclusions are summarized below:(1)A time-varying sliding window correction model based on the Copula function is proposed.Classification and continuity evaluation indicators are used for evaluating the forecast accuracy of ECMWF and WRF models for 15 zones in the upper reaches of the Yangtze River.It is shown that the ECMWF model performs better than the WRF model,and the advantage become more obvious as the forecast period is extended.The MK method is used to test the non-stationary of correlation series between the predicted and measured precipitation.The Copula function and the distribution mapping correction method are used to correct the numerical precipitation forecasting bias.All index results are significantly improved after correction,and the conditional distribution based on the Copula function performs better than the distribution mapping correction method.(2)The selection of the marginal distribution in the conditional probability prediction model based on the Copula function is a key issue.A Gaussian mixture distribution is used to fit flow series and compared with the GEV,Weibull,and Gamma distributions.Results show that the Gaussian mixture distribution can best fit inflow series of Lushui reservoir,the CRPS value in the lead-time of 1～ 4 days are decreased by 25.6% ～ 29.5%.(3)The flood probability forecasting based on error correction was carried out to increase flood forecasting accuracy and reduce the hydrological uncertainty.Results show that the ANFIS model performs much better than ARXM and AR models because it not only can effectively identify the non-linear relationship in hydrological process but also the relationship between errors and flood magnitude;the error correction method is an effective tool to improve the accuracy and reliability of probabilistic forecasts.(4)To investigate the impact of precipitation prediction uncertainty on flood forecasting,the TIGGE data extracted from the CMA,NCEP,and ECMWF were input to the GR4 J hydrological model.The effectiveness of four statistical post-processing methods,including BMA,Copula-BMA,EMOS and M-BMA methods,were compared and analyzed.The results show that each of the four methods could provide a reasonable and reliable confidence interval on prediction.Compared with the raw deterministic forecasts,the forecast accuracy of expected values associated with the four methods was improved,where the forecast error in water volume was significantly reduced.The M-BMA method performs the best because it considers the heteroscedasticity of the predictive distribution,without conducting a normal transformation,which could be much simpler and more flexible in practice.(5)The current common used statistical post-processing methods assume that the distribution parameters are related to the ensemble forecast values.Considering the antecedent runoff factor to estimate the distribution parameters,a modified M-BMA probability forecast method is proposed.From the comprehensive view of the index results of deterministic and probabilistic prediction,the modified method performs much better than the M-BMA method.The probabilistic forecasting can be significantly improved through considering the antecedent runoff factor in parameter estimation.