Combining Ensemble Weather Forecasts And Statistical Post-Processing Methods For Streamflow Predictions

Under the influence of global change and human activities,the problem of the short forecast period and insufficient accuracy or reliability are increasingly prominent in streamflow prediction,which poses a negative impact on people’s abilities to cope with the future extreme hydrological events,i.e.floods.The promising way to deal with the above challenges is to conduct ensemble streamflow prediction(ESP)which combines the ensemble weather forecasts(EWFs)and the hydrological models.Ensemble forecasting helps to extend the forecast periods of the streamflow forecasts and reduce the forecast uncertainty induced by the deterministic forecasting,which improves the availability and accuracy of hydrological forecasting.However,raw EWFs are usually biased and underdispersive.The uses of EWFs are also restricted by the low spatial resolution and the inconsistent spatiotemporal dependence with the observations.The post-processing method is required for correcting the EWFs before being used as the meteorological inputs of the hydrological models.Two problems have emerged when choosing the proper post-processing method,including the evaluation of the existing univariate post-processing method applicable for the single weather variable,and the development of the multivariable method which considers the dependencies among the weather variables.Based on the forecasts from the Global Ensemble Forecasting System Reforecasts(GEFS reforecasts),and the observations from the weather stations in China and the hydrological stations in Xiangjiang basin,this study aims to evaluate the applicability of the EWFs,compare the performances of different univariate post-processing methods,and propose a novel multivariable postprocessing method which considers the multivariable dependence structure.The main works and corresponding conclusions are as follows:(1)This study proposes a novel method for evaluating the applicability of EWFs.The proposed method is based on two indexes,the significant correlation period and the relative valid forecast period.This method was used to evaluate the applicability of GEFS forecasts in China and some application suggestions were given.It found: The precipitation forecasts have good applicability in eastern monsoon regions in China,including Northeast China,Northern China,Central China,and Southeastern China.The valid forecast period for these regions is above 7 days.The precipitation forecasts have poor applicability in Northwestern China.The performance of precipitation forecasts in Southwestern China depends on the post-processing methods and is influenced by the under-dispersion of the EWFs.Temperature forecasts have good applicability in China,and the valid forecast period for air temperature is generally above 12 days.Specifically,over 80% of stations are found to own a valid forecast period of 15 days.(2)The study compared different univariate post-processing methods and evaluated the possible influencing factors in weather forecasting and streamflow prediction.The chosen post-processing methods include Generator-based Post-Processing(GPP),Extended Logistic Regression(Ex LR),Bayesian Model Averaging(BMA),and Affine Kernel Dressing(AKD).The first two methods belong to the post-processing methods whose parameters are coherent with the observations,while the latter two methods belong to the post-processing methods whose parameters are calibrated from the EWFs.The four methods were evaluated in correcting the EWFs in China and generating the ESP in Xiangjiang Basin.It found the postprocessing methods whose parameters are coherent with the observations(GPP and Ex LR)perform better than the post-processing methods whose parameters are calibrated from the EWFs(BMA and AKD)in both weather forecasting and streamflow prediction.Among the four methods,GPP is recommended considering its overall performances.The influencing factors for choosing the post-processing methods include weather variable type,the regions,and climate type,and the forecast time.Specifically,the key to improve the EWFs lies in improving the ensemble spread for precipitation and correcting bias for air temperature,which results in an obvious performance difference in post-processing the precipitation forecasts compared to the temperature forecasts.The performance of the post-processing methods is generally better in northern regions compared to other regions for precipitation forecasts;poorer in the Qinghai-Tibetan regions for temperature forecasts;and generally poorer in summer compared to other seasons for temperature forecasts.(3)This study proposes the multivariable post-processing method where the dependence structure of the forecast variables being considered.Under the formulated framework proposed in this study,the dependence reconstruction method is combined into the univariate post-processing method in two ways,Pre-coupling and Post-coupling.This study chose GPP as the univariate post-processing method and provided three dependence reconstruction methods,Rank shuffle(RS),Gaussian Copula(GC),and Empirical Copula(EC).Based on the proposed framework,a total of six multivariable methods are developed,including RSPost,RS-Pre,EC-Post,EC-Pre,GC-Post,and GC-Pre.The six multivariable methods were evaluated and compared in weather forecasting,streamflow prediction,and real-time flood forecasting in Xiangjiang Basin.It found,1)when using the post-processing method for correcting the EWFs,the inter-variable and inter-site correlations considered in multivariable post-processing help to improve the probabilistic performance of the EWFs,improve the runoff simulation across space,increase the accuracy performance,probabilistic performance and forecasted interval performance.Among the six methods,the six methods except GCPre are shown effective to reconstruct the inter-variable and inter-site dependencies and improve the performances in weather forecasting and streamflow prediction.RS-Post is recommended considering its accuracy,probabilistic,and forecasted interval performance.