Scribd
Upload
7 views11 pages

Can Ensemble Forecasts

artigo

Uploaded by

Lais Faria
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
7 views11 pages

Can Ensemble Forecasts

artigo

Uploaded by

Lais Faria
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 11
Can ensemble forecasts improve the reliability of flood alerts? J. Dietrich’, M. Denhard? and A.H. Schumann? | Institute of Water Resourees Managment, Hydrology and Agicutural Mycraule Engng, Letina University, Mannover, Gennany 2 Deutscher Wettecienst (German Notional Meteoriogtal Sense), Cifenbach, Gesmary ‘3 intute of Hydrology, Water Resources Management and Eméeonmental Engnesing, Ruli-Uniersy, Bochum, Germany Correspondence Jorg Deve, ctu of Water Resources Management, Hytology and Agriculture Hyarauic Engingeong, Leia University, D167 Hanna, Germany Tek 4.495518 762 2309 mall detich@ mi unénannver. de ‘Thispaper was presented at the th International Srpelum on Flood Defence, Torta, Canada(6-B May 2008) and wil be pubisnedin a special sue of Journai of Ficod Ask Management edited ty Profesce Sobodan P Senenove, University of Western conta, Abstract. A probabilistic evaluation of ensemble forecasts can be used to communicate uncertainty to decision makers. We present a flood forecast scheme, which combines forecasts from the European COSMO-LEPS, SRNWP-PEPS and (COSMO-DE (lagged average) ensemble prediction systems with a rainfall-runoff model, The methadology wes demonstrated with a case study for the Central European Mulde River basin. In this paper, we summarize results from hindeast simulations for seven events from 2002 to 2008. The ensemble spread resulting from uncertainty in rainfall forecast was very high at 2-5 days lead time, The median of the medium- and short-range forceasts and a lagged average ensemble ofthe very short-range forecasts praved to be reliable regarding the probability of exceeding flood alert levels. However, the limited number of observed events does not allow for the postulation of prescriptive binary decision rules. Flood managers have to adapt their decisions when new information becomes available. (pot 10.11184,1753-318« 2008 01098 ey words [eiion sspport system; ensemble technique, etree fled, food Forecast: lod rik, ‘ait al-runat mae: laity Introduction Because of the highly nonlinear behaviour of the atmo- spheric system and the land-atmosphete interaction, the Iknovledge of the future meteorological and hydrological development is inherently incomplete and uncertain. Parti- ccularly the quantitative forecast af rainfall events is subject to uncertainty, which often (but not necessarily) increases, with a longer lead time of the forccast. Within a flood forecast chain the metcarological forecasts can farce hydro- logical and hydraulic models. These models add additional sources of uncertainty, eg. the availability and quality of input data, the initial and boundary conditions for the ‘models, model parameters and model structure. Inaccurate hbuman interaction and technical problems may also affect the output of a flood forecast chain. Resulting from these luncertainties it is not possible to issue a perfect flood forecast During the last decades the traditional determi ‘modelling paradigm has more and more been superseded by the prababilistic paradigm, which admits the imperfect- (92000 he ctor [nur Compl #2008 eel ress, respectively, the uncertainty of the forecast. Ensemble forecasts aim at framing the uncertainty of the potential fature evolution of the hydro-metcorological situation (Anderson, 1996; Kalnay, 2002; Toth et al, 2003). An censemble-hosed forecast produces a set of valucs instead of fa single value of the variable under consideration. ‘The ascessment and the aggregation ofinformation are necessary for decision support, eg. the comput probabilities for threshold values of crt ‘causing inundation, ‘Among the numerous ensemble generation methods are physics ensembles (perturbation of model parameters or use ‘of different schemes within one model), multimodel ensem- bles (combination of different models} and lagged average ensembles (combination of different model runs of a single model), Different meteorological ensemble prediction systems {EPS) became operational at global seale(Buizza etal, 2005). ‘During the last years the spatial resolution of the EPS has been continuously improved. Nowadays many weather services offer EPS at regional scale with different horizontal resol aand different lead times jon of exceedance sal discharge levels, oo anager 22008) 222-282 Camensetle forecasts improve te eablity flood ales? EPS can force hydrological models, which simulate the rainfall-runoff process, river routing and inundation, A chain of meteorological and hydrological models can com- pute flood forecasts. The development of hydrological applications of ensemble forecasts has been demonstrated by several studies (eg. de Roo et al, 2003; Gouweleeuw ial, 2005; Verbunt eta, 2006; Red et a, 2007; Dietich ‘eta, 2008; Diomede era, 2008). Different duties and responsibilities in water resources ‘management, particulary inthe field of flood management, ‘can potentially benefit from. ensemble forecast: reservoir contra, issuing flood alets, initiating food defence mea- sures. An ensemble-hased operational flood management system (OFMS) should fulfil diferent requirements regard- ing the reevance of uncerinty accuracy needed) and the lead times of the forecast. This study proposes an adaptive flood management strategy, which sims at an efficient, use case-drven processing and evaluation ofthe available fore- ‘ests and observations The authors developed a scheme for the combination of ensembles from diferent sources as a basis for an OFMS (Dictich et al, 2008). The OFMS intogrates cosemble forecasts from thes systems, which are ‘operationally provided by meteorological services: 'COSMO-LEPS (Molten eral, 2001) isa regional scale physics EPS with 16 (10) members, approximately 10km horizontal reslution and 5 days lead time. 1 SRNWP-PEES (Denhard and Trepte, 2006) combines up 023 deterministic forecasts fom 21 national metcorolo- gical services. This multimodel ensemble has approxi- ‘ately 7km horizontal resolution and 2 days lead time + COSMO-DE (Stepper et al, 2003) is @ deterministic local model with 2.8km horizontal reslution and 21 h lead time. The most recent model run can be combined with earlier model rans to build a lagged average ensem= ble ‘Water manages from local authorities were involved in the case study reported within this paper. Practitioners ‘wondered about the reliability of probabilistic forecasts, which can be derived from the forecast ensersbes, Thit paper addresses the following questions related to flood suanagement * Can the additional knowledge about uncertainty provide audded value for decision making? s Are the probabilistic ood ser level predictions rlshe? [Av this time, the authors can only rely on a limited amour of data in particular for heavy and extreme events (ovhich are of special interest fr flood managers). Thus i ‘will not be possible to drawe general concusions from this ‘ese study alone. Nevertheless we provide examples for the tue of raw ensemble forecans in one important wse case of flood managers issuing of flood alerts, Further work will seal with refinements of the OEMS by integration of more advanced ensemble postprocessing methods nd data astimi- ced ik Managment 22608) 292-282 28 ilation, which is of special importance for application near real time, when flood defence measures have to be initiated and when unexpected events (eg. local extremes, dike breeches, loss of sensors) have tobe tackled by the system. ‘The next sections of the paper give a short introduction into the ensemble-based flood forecast frameveork. We evaluate and discuss (a) implications from knowledge about ‘uncertainty for issuing flood alerts and (b) the reliability of flood alerts derived from the different EPS. ‘Combination of different types of ensembles for operational flood forecasting ‘When designing an OFMS, « compromise among computa- tional efficiency, availabilty of data, predictive capability of the models and the cognitive burden for the lood manager thas to be found. An OFMS is typically buile from compo- nents, which are among the generic components of decision support systems (DSS), namely a knowledge system, a problem processing system and a user interface (language system and presentation system; Dos Santos and Holsapple, 1989). Thus the OFMS can be seen asa specific type of DS ‘As mentioned in the introductory section, there is a variety of ensemble approaches. For the specific forecast situation, only asubset of the information may be accessible and useful at the same time. Furthermore, only a subset of the problem processing tools may be needed to obtain the desired ‘outcome ofthe computation, An adaptive systems approach, ‘can support the efficient combination of the sources of information, which are available and useful for the curcent situation and the current user of the system, The OEMS presented here follows an adaptive approach in the wider sense, Adaptive DSS in the narrower sense as defined by Holsapple etal. (1993) include additional components for unsupervised machine learning. This means thatthe system is able to learn and adapt itself in order to improve the quality of the outcome by using the same input data. Extending the adaptation capabilites of the OFMS is subject cof further research ‘The OFMS prototype described in this study combines ‘medium-range forecasts (3-5 days lead time), short-range forecasts (1-2 days lead time) and very short-range forecasts (<1 day lead time) from different operational meteorolo- sical prediction systems with hydrological models (Figure 1), Mediur-range flood forecasts forced by COSMO-LEPS provide the basis for decisions about reservoir management and early warnings previous to a potential large or extreme flood event. Additional short-range forecasts fram SRNWP- EPS can be used for issuing flood alerts and first planning ‘of flood defence measures. For the incorporation of forecast refinements with 2.8km horizontal resolution and 3-hourly update we use the COSMO-DE model, 2008 the ator hurd Caplan #2109 tat beg ° S(t d.updatey ° ° . 20-0080 ~~ 2 (6hi24 h update) 290000 24 Wh update) physics ensemble meso-sosle, medium range multi-modal encemble meso-ecale, short range ‘Lagged average ensernble local scala, vary short range prcbabaisraticcmaio_ Fal COSWO-LEPS SANWP-PEPS ‘COSMO-DE ensemble post precessing rina art od -ACEGMO probabilistic ‘acEGMO realm eneorble Bayesian updaia of lonsenble weighs ot ensemble postprocessing accion support Tot eaigimunaion Fe Figure’ Scheme ofan operational lod frecast system with an adaptive combination of meteorological ensemble forecasts fem diferent sources ‘ hyerlagical made k enanged fam Bleich et a, 2008), ‘the meteorological ensembles are fed into hydrological ‘models to simulate stream flow ensembles. Note that we do not process meteorological probability forecasts within the (OEMS. We use the complete ensemble to force the hydro- logical models. Thus a stream flow ensemble has at least as ‘many members asthe forcing meteorological ensemble, The skil of the hydrological forecast strongly depends on the skill of the precipitation and temperature forceast, because these two climate variables dominate the generation of fast runoff processes and snow melt, The QFMS prototype presented in this paper includes the disteluted conceptual ‘ainfall-runoff model ArcEGMO (Becker ef al, 2002). Most ofits parameters have a physical meaning and can be derived from catchment characteristics. ArcEGMO is a modular ‘modelling system, whose modelling keme! can be controlled isan external flood management application in a computa- tionally efficient way. Ths allows for the simulation ofa large umber of forecasts near realtime, Thus the rainfill-ranof ‘model can simulate ensemble forecasts of stream flow at several points of interest like gauges and vulnerable sites The OFMS can be operated in two modes: standard and extended. The standard mode is designed for the fast computation of raw (uncalibrated) stream flow ensembles (outside the grey area in Figure 1). The stream flow simula tions are updated whenever one of the three operational 2009 the tues, Jounal Compan 08 ac sig ensemble systems provides anew meteorological forecast via file transfer or database connection. The update interval is between 3h for COSMO-DE and I day for COSMO-LEPS. Figure 1 shows the update sequence in advance of a time step of interest (black dots). The SRNWP-PEPS is updated every Gh, but not all participating systems deliver data within that interval. ‘The ArcEGMO model is run in a continuous mode forced by observed climatic input and in 4 forecast made forced by the meteoralagical ensembles. ‘The OFMS can save system states ofthe hydrological model at every time step, In forecast mode, the model is restarted for every single member of the meteorological forecast using, the system states of the continuous model as consistent initial states. ‘Currently, the standard mode does not include ensemble postprocessing. Thus the probabilistic evaluation is merely based on a relative frequency approach here. There are more advanced techniques available, which may improve the probabilistic evaluation of the ensemble forecasts and there- fore provide a better picture of uncertainty. The most important methods are the generation of closed probability distributions (eg. by using probability density functions or kernel density estimates) and the calibration of ensembles (eg. by computing weights ofthe ensemble members andior by correcting the bias). The shortcoming of these methods is ag anager 208) 282282 ‘Can ensemble forecasts improve the rehab af fd aarts? the introduction of more parameters and the neod of training against observations, In the case of heavy oF ‘treme events, there are not a suflcient number of stua- tons with forecast and observation availabe. ‘The extended modules of the OFMS are subject of ‘ongoing research and will be published in subsequent papers. These modules will support the generation of hydrological ensembles, the calibration and postprocessing of ensembles and a more advanced prababilstic valuation (yey arca in Figure 1). Hydrological ensembles were de- signed to regard the uncertainty ofthe initial state and the parameters of the model (Dietrich eral, 2008)."The mem= bers of the hydrological ensemble, respectively, alternative system states and parameter sts of the modal, are selected by inference of antecedent precipitation, the type of the ‘expected event and efficient parameter sets obtained by mode! calibration against past flood events. In the extended mode the forccast can be updated in 1-3 hourly intervals ding assimilation of observed rainfall and. discharge data. More details and a case stady for this option will be presented in a subsequent paper Case study ‘The upper Mulde river basin is situated in the Ore Moun tains (Germany and Caech Republic; Figure 2). Narow and steep valleys cause a fast reaction of the watershed and segeeegees 100. 200: 300: 400. 500. 600. 700 800 900 a8 critical superimposition of flood waves. Several cities are located in the flood plain of the lower Mulde river basin During west-cyclonic rainfall events, which caused several extreme flood events in the past, the uncertainty of pre-

You might also like