Flood forecastingFredrik Wetterhall
European Centre for Medium-Range Weather Forecasts
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Outline
•Introduction
•Operational forecasting systems (EFAS, GloFAS)
•S2S hydrological forecasting
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Flooding – a global challenge
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Flooding – a global challenge
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS October 29, 2014
Causes of flooding
•snowmelt runoff
•rainfall
•ice jams and other obstructions
•coastal storms (tsunamis, cyclones, hurricanes)
•urban stormwater runoff;
•dam failure (or the failure of some other hydraulic structure).
•Etc …
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Forecasting chain
EPSHydrology
Warning
Preprocessing/calibration
Postprocessing
Verification
Feedback to the model
Forecasts can fail because:• The initial conditions are not accurate enough, e.g. due to poor coverage and/or observation errors, or errors in the assimilation (initial uncertainties).• The model used to assimilate the data and to make the forecast describe only in an approximate way the true atmospheric phenomena (model uncertainties).• A combination of the two phenomena
As a further complication, the atmosphere is a chaotic system!
t=0
t=T1
t=T2
Why do forecasts fail?
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS October 29, 2014
Early probabilistic flood warnings across Europe
Transboundary
50 partners
Partners provide:• Observations• Feedback on warning performance• Development of decisions
EFAS has the largest collection of hydro-meteorological observations in Europe!
European Flood awareness system EFAS
What are the Benefits ?
National Hydrological Services European Commission
Novel information
Added value
- Catchment based information
- River basins larger than 4000 sq.km and regional cross-border dimension
- Longer lead-times up to 15 days through probabilistic information
- Network of operational services
- Promotion of novel tools, techniques and data sets (e.g. satellite data)
- Comparable information across Europe
- Tool for anticipation of crisis management:
- Civil Protection aid assistance during crisis
- COPERNICUS Mapping Service
Expert Knowledge of Member States
Real-time data (EU-FLOOD-GIS/ETN-R)
Europ. Data Layers
Meteo – Data / forecasts
Historical Data
Static Data
0
500
1000
1500
2000
2500
8/23/02 0:00 8/24/02 0:00 8/25/02 0:00 8/26/02 0:00 8/27/02 0:00 8/28/02 0:00 8/29/02 0:00 8/30/02 0:00 8/31/02 0:00
Dessau/Rosslau Wittenberg Torgau Riesa Dresden Labe Decin Labe/Usti N.L. Vltava/Prague
EFAS partner network
Alert email
EFAS user interface
DA
TA
Hydrological modeling
Linz, AT – 31/05/2013 00 UTC
How do we actually do flooding prediction? Schematic view
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
EFAS - Time series
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
EFAS - Time series simplified
Single deterministic forecasts
EPS forecasts
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
EFAS - Condensing information
Nr of EPS exceeding thresholds
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Pre
viou
s fo
reca
sts
Today’s forecast
Event forecast
Evaluation of persistence in time and consistence between forecasts are important
EFAS - Looking back in time
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 15
How reliable and accurate is our flooding prediction?
Flood warning for Passau, Germany 30 May 2013
Donau at Passau (DE), 30/5/2013
Although the flood was predicted it was not high enough. Top discharge should have been 10000
2013-05-31 to 2013-06-03
Going higher resolution: TL3999 (5 km) TP fc (+72h)
32 km ENS
16 km HRES
65 km
5 km
Observations
2013-05-31 to 2013-06-03
Increased resolution + modified cloud physics
HRES (16 km)
Observations
Higher resolution and physics: TL3999 (5 km) TP fc (+72h)
UK floods, December 2012 – Trent River
Trent at Dunham Bridge near Gainsborough, 27/12/2012
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 20
EFAS flash floods Medium
High
Severe
Alert Extremity
A global challenge? The Global Flood Awareness Systems
hydrological model
Decision support informationGlobal probabilistic weather forecast (ECMWF)
GloFAS – A global challenge
Streamflow simulations with ERA-Interim global atmospheric reanalysis as meteo input
Comparison with observed discharge data (~1400 world stations)
Ensemble streamflow predictions
Forecast peak flow detected ~10-15 days in advance
SE Asia floods Sept/Oct 2011 (Chao Praya and Mekong)
www.globalfloods.eu
Username: trainingPassword: tra1000
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Forecasting chain in future EFAS
EPSHydrology
Warning
Preprocessing/calibration
Postprocessing
Verification
Feedback to the model
S2SMultimodel
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 27
S2S hydrological forecasting
•EFAS (European Flood Awareness System): operational system for early flood and flash flood warnings over Europe (up to 15 days lead time)
•Growing incentive for hydrological forecasts at longer lead times:
– Applications: hydropower management, spring flood prediction, low flows prediction for navigation, agricultural water needs...
– Increase in NWP skill
•Aims:
– Produce seasonal streamflow predictions for Europe using ECMWF dynamical seasonal forecasts
– Provide probabilistic outlooks against model reforecasts for seasonal predictions beyond 15 days
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 28
Data
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 29
Evaluation strategy
•Scores computed:
– On weekly catchment discharge averages
– 1990 - 2013
– For each season (DJF, MAM, JJA, SON)
– Lead time: 1 - 8 weeks
– Against EFAS-WB
•Two main studiesEuropean catchments map used for the analysis (74 catchments)
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 30
Evaluation strategyMeteorological forcings (MF) versus initial
conditions (IC)
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 31
Evaluation strategyMeteorological forcings (MF) versus initial
conditions (IC)
•Reverse-ESP: 15 resampled years of initial conditions and ‘perfect’ meteorological forcing data (Wood and Lettenmaier, 2008)
•MF lead the uncertainty over the IC variance ESP > variance rESP
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 32
Results1) Seasonal predictability over Europe
•Decreasing accuracy with lead time
•On average still some accuracy until 8 weeks
•Increasing geographical disparities with lead time
•Seasonal more accurate than ESP on average until 4 weeks
•Increasing gap during 2nd week between seasonal and ESP
KGE for all seasons combined
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
KGE over all catchments and seasons
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 34
Results1) Seasonal predictability over Europe
•Higher predictability in summer
•Gain of using seasonal forecast increases in winter for lead times 1 to 4 weeks
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 35
Results1) Seasonal predictability over Europe
•Seasonal shows highest gain in predictability in winter:
– Iberian Peninsula
– Scandinavia (Baltic Sea)
•In summer predictability largest for:
– Scandinavia (Baltic Sea)
– Around Mediterranean Sea
– South of North Sea
Lead time at which CRPSS ≤ 0
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 36
Results CRPSS SYS4 vs ESP
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 37
Results1) Seasonal predictability over Europe
•Decreasing skill with lead time, but still skilful until about 6 weeks
•Seasonal and ESP show similar ROC score for week 1, then seasonal’s ROC scores higher
•Large decrease in skill for ESP between 1 and 2 weeks
•Both systems more skilful to resolve low flows than high flows
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 38
Results2. Meteorological forcings (MF) versus initial conditions (IC)
• Critical Lead Time (CLT) (Yossef et al., 2013): lead time at which var ESP > var rESP
1. High CLTs, leeward: groundwater fed rivers during winter
2. Low CLTs, windward: moist westerly winds
3. Low CLTs: precipitation driven flows in winter
4. High CLTs: drier antecedent moisture conditions
5. High CLTs: snowmelt driven discharges
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS 39
Take-home messages
Overall gain of using seasonal forecasts from 1 – 4 weeks lead time
Especially in winter: Iberian Peninsula and Scandinavia (Baltic Sea)
MF leads uncertainty over IC from 2 weeks of lead time on (average for Europe)
Seasonal transitions between hydrological states (wet, dry) crucial in this process
Seasonal more skilful to resolve low and high flows from the 2nd - 8th week lead time
Lower flows more skilfully resolved than upper flows
Multimodel system: test of three models
Hydrological models
LISFLOODHTESSEL/ CAMA
EHYPE
CharacteristicHydrological model with channel routing
HTESSEL coupled with CamaFlood
Semi-distributed conceptual model
Resolution 5 km gridded80 km land surface model 25 km routing
Catchment based (varying resolution)
Driving data5 km gridded observed data
ERA-Interim corrected with GPCP/5 km gridded observed data
ERA-Interim corrected with GPCC/5 km gridded observed data
Test period: 1990-2010Observational discharge: 212 stations from GRDC
LISFLOOD
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
HTESSEL and CaMa-Flood
E-HYPE – pan European HYPE application
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Results: Nash-Sutcliffe and mean relative error
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Multimodel – how useful is it for decision making?
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Create one supermodel to rule them all?
•Bayesian Model Averaging (BMA) accounts for the model uncertainty by averaging over the best models according to posterior model probability
•The BMA methodology applied for this study performs analysis assuming a uniform distribution on model priors and using a simple BIC (Bayesian Information Criterion) approximation to construct the prior probabilities on the regressions coefficients (Raftery, Hoeting, Volinsky, Painter & Yeung 2010).
•
•If M = {,…, } denotes the set of models and if is the quantity of interest (e.g. stream-flow), then the posterior distribution of given the data D is:
This is an average of the posterior distributions under each model weighted by the corresponding posterior model probabilities.
𝑝𝑟 (∆|𝐷 ¿=∑𝑘=1
𝐾
𝑝𝑟 (∆|𝑀𝑘 ,𝐷 )𝑝𝑟 (𝑀𝑘∨𝐷)
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Results: Bayesian model averaging
BMA improves NSE in 76% of the cases
EUROPEAN CENTRE FOR MEDIUM-RANGE WEATHER FORECASTS
Conclusions
•The different models perform differently depending on basin characteristics, such as size and elevation
•BMA improves the performance of the models in most cases, and gives reasonable results even if the individual models are not doing well in a particular catchment
•Without proper calibration it is difficult to see a great benefit from the 5km gridded dataset
•Combining models will be a challenge