Research On Distributed Ensemble Flood Forecasting Based On TIGGE Dataset

Due to climate change and urban expansion,extreme events have occurred with an increasing frequency.Flooding became a serious issue in China and worldwide,with large losses of life and damage to property.Enhanced flood forecasting early warning system is beneficial to take appropriate prevention measures.With the development in meteorological technology and computational power,utilization of numerical weather predictions products to generate ensemble flood predictions is accelerated and has been proved to be effective.This study attempted to couple TIGGE dataset with VIC distributed hydrological model for ensemble flood forecasting system.We performed our research in Lanjiang and the Yarlung Zangbo River(YZR)basin to investigate the predictability of peaks,flood volumes,components and so on.The main effort and conclusions are as follows:(1)Two post-processing methods,CSGD-EMOS and QM were used to reduce bias and to improve the QPFs from ECMWF,NCEP and CMA.Results were evaluated by using three incremental precipitation thresholds and in two different seasons(plum rain season and typhoon season).It is indicated that QPFs from NCEP and ECMWF presented similarly skillful forecasts,although the ECMWF QPFs performed more satisfactorily in the typhoon season and the NCEP QPFs were better in the plum rain season.Lighter precipitation tended to be overestimated,but heavier precipitation was always missed.In general,raw forecasts can provide acceptable QPFs 8 days in advance.After post-processing,the useful forecasts can be significantly extended beyond 10 days.QM had a greater effect on removing bias but was unable to warrant reliabilities.CSGD-EMOS was more effective for probabilistic metrics.(2)VIC model was automatically and parallelly calibrated by ?-NSGAII based on MPICH.Then,the daily(VIC/D)and POT peak model(VIC/P)frameworks were established.The modular approach was adopted to combine flows from VIC/D and VIC/P.It is found that the improved calibration strategy and modular approach could impressively increase the accuracy of hydrological model,with NSE for daily flow greater than 0.8 and near 0.9 for peak flows selected by the POT method.It is also demonstrated that the forecasts from ECMWF were the main contributor to multimodel ensembles,with the relative contribution rate up to 20%.Various ensemble methods differed from each other,and those methods assigning higher weights to ECMWF had higher skill score.The flow can be predicted in 10 days ahead,whilst the POT peaks were generally underestimated by ensembles.(3)A set of preferred solutions was selected by POR from the Pareto-optimal fronts provided by ?-NSGA.Together with the extreme value for different objective functions,those solutions were used to represent model parameter uncertainty for investigating the influence on flood and its components forecasting.ECWMF and VIC were coupled for ensemble flood forecasting and hydrological separation was achieved by snowmelt tracking algorithm.We found that the VIC with considering model uncertainty(NVIC)performed better than conventional one in which no model parameter uncertainty was involved(SVIC).However,NVIC generally lost the advantages in forecasting applications Our forecast system was able to capture the annual maximum flood events with 10 days in advance,and the meltwater-related components can be predicted about 7 days ahead.The accuracy of forecasts for the first floods is inferior,with a lead time of only 5 days.The meltwater-induced surface runoff was the most poorly captured component by the forecast system,and the well-predicted rainfall-related components were the major contributor to good performance.(4)Well-marked seasonal and sub-seasonal variations in error statistics from VIC model in the Yarlung Zangbo River(YZR)were detected by the analysis of temporal variation in error statistics.Error statistics in summer and winter were homogeneous,consistently underestimated in summer and overestimated in winter.Error distributions in spring and autumn were obviously transitional.It was caused by the varied degree of success of VIC in simulating the cooccurrence of snow,rainfall and glacier in these periods Accordingly,three temporally varied ERRIS models were constructed at semiannual(ERRIS-H),seasonal(ERRIS-S)and monthly(ERRIS-M)temporal granularity.Results showed that the temporally varied ERRIS models were considerably more effective than the temporally unvaried one(ERRIS-A),with 34%reduction in continuous ranked probability score(CRPS)and 23%increase in Nash-Sutcliffe Efficiency(NSE)for prediction applications.ERRIS-A even had a negative impact on raw model simulations With respect to forecasting applications,improvements of about 7%in CRPS were achieved by the temporally varied models.The performance of different temporally varied error models roughly followed the same order as the level of temporal granularity Generally,ERRIS-S and ERRIS-M are similarly effective when no sub-seasonal error statistics were found.When complicated hydrological processes existed,temporally finer error models were expected to lead to higher accuracy and reliability.