Using ensemble forecasts for hydrological prediction in the Great Lakes

1
Using ensemble forecasts for hydrological
prediction in the Great Lakes St. Lawrence basin

• Vincent Fortin
• Recherche en prévision numérique environnementale
• Meteorological Research Division
• Environment Canada
• NAEFS workshop
• Camp Springs,6 8 October 2008

2
Outline

• HEPEX Great Lakes testbed project
• MESH hydrological modelling system
• Hydrological Ensemble Prediction System (H-EPS)
• Sub-projects based on the use of MESH and the
  H-EPS on the Great Lakes
• International Upper Great Lakes Study
• Decision support system for ensemble flood
  forecasts
• Ensemble urban prediction system
• Summary

3
Great Lakes and St.Lawrence basin

• Largest lake group in the world
• lake area 250 000 km²
• watershed area 1 000 000 km²
• population 40 millions
• 30 of Canada's population
• 10 of US population
• Regulated according to an international
  agreement between Canada and the US
• implemented by the International Joint Commission

4
Hydrological Ensemble Prediction Experiment
(HEPEX) Great Lakes testbed

• Demonstrate the added value of hydrological
  ensemble forecasts in the Great Lakes basin
• Plan A Develop an ensemble hydrological
  prediction system using in-house resources
• Plan B Support RD and demonstration projects
  with unified hydrological modelling system
  (CaPACaLDASMESH)
• Currently funded projects
• International Upper Great Lakes Study
  Understanding and predicting the water balance of
  the Great Lakes (April 2008 March 2011)
• Decision support system for ensemble flood
  forecasts in Québec (Jan 2008 Dec 2012)
• Ensemble urban prediction system improving
  stormwater management (Jan 2008 Dec 2010)

5
Canadian Land data assimilation system (CaLDAS)
CaPA
Observations
GEM

• rain gauges
• radar
• satellite

Atmospheric
Precipitation analysis
Model
mass, energy and
prognostic variables,
momentum fluxes
e.g. soil moisture
diagnostic 2m
temperature
and relative humidity
CaLDAS
ISBA or CLASS
Observations

• temperature
• humidity
• soil moisture

Land data
Land surface
soil moisture
Assimilation system
scheme
and
temperature
Surface runoff and groundwater recharge
Recycle bin
6
CaPA-CaLDAS-MESHmodelling system
CaPA
Observations
GEM

• rain gauges
• radar
• satellite

Atmospheric
Precipitation analysis
Model
mass, energy and
prognostic variables,
momentum fluxes
e.g. soil moisture
diagnostic 2m
temperature
and relative humidity
CaLDAS
ISBA or CLASS
Observations

• temperature
• humidity
• soil moisture

Land data
Land surface
soil moisture
Assimilation system
scheme
and
temperature
Surface runoff and groundwater recharge
prognostic variables,
soil moisture
e.g. water levels in the river network
WATROUTE
WATDRAIN
River and lake routing model
2D Hillslope model
lateral flow
streamflow
7
CaPA-CaLDAS-MESHmodelling system
CaPA
Observations
GEM

• rain gauges
• radar
• satellite

Atmospheric
Precipitation analysis
Model
MESH
mass, energy and
prognostic variables,
momentum fluxes
e.g. soil moisture
diagnostic 2m
temperature
and relative humidity
CaLDAS
ISBA or CLASS
Observations

• temperature
• humidity
• soil moisture

Land data
Land surface
soil moisture
Assimilation system
scheme
and
temperature
Surface runoff and groundwater recharge
prognostic variables,
soil moisture
e.g. water levels in the river network
WATROUTE
WATDRAIN
River and lake routing model
2D Hillslope model
lateral flow
streamflow
8
15-day hydrological ensemble prediction system
(H-EPS) is it what users want?
H-EPS member
EPS
CMC EPS
prognostic variables,
mass, energy and
e.g. soil moisture
momentum fluxes
Smoothed ensemble forecast of mean monthly
outflowfor Lake Superior
MESH
Land surface
scheme and
Hydrological spaghetti plots
hydrological model
Streamflow forecast Black River near Washago
streamflow
Lake Ontario level
and lake levels
9
International Upper Great Lakes Study (April 2008
March 2011)

• Understand and predict the water balance of the
  Great Lakes
• Understand why the Upper Great Lakes (Superior,
  Michigan-Huron) came close to their all time low
  in the fall of 2007
• Less precipitation, more evaporation and/or more
  outflow because of erosion?
• Assess monthly precipitation, evaporation and
  runoff into each lake for the past 50 years
• Predict same variables for the next 50 years
• Our contribution
• Assessment of monthly precipitation, evaporation
  and runoff into each lake for the past 5 years
  using MESH (FY 12)
• Ensemble prediction of the same variables for the
  next two weeks using NAEFSMESH (FY 23)

10
Overlake precipitation estimation

• CaPA analysis vs interpolation of observations at
  GLERLfor Lake Superior

June 2004 July 2007
2005
11
Overlake evaporation estimation

• MESH model vs GLERL model for Lake Superior

June 2004 July 2007
2005
12
Estimation of runoff into the lakes

• MESH model vs GLERL model obs. for Lake Superior

June 2004 Dec 2005
2005
13
Simulation of NBS with MESH compared to GLERL
estimates

• Net basin supply overlake precipitation
  overlake evaporation runoff into the lake

June 2004 Dec 2005
2005
MESH GLERL Residual method
14
Simulation of lake levels (no data assimilation
except for the use of CaPA)

• Results in simulation mode bias in precipitation
  analysis shows up

Lake Superior Lake Michigan-Huron Lake
Erie Lake Ontario
15
Simulation of lake levels (no data assimilation
except for the use of CaPA)

• Results in simulation mode with precipitation
  boosted by 25

Lake Superior Lake Michigan-Huron Lake
Erie Lake Ontario
16
Current priority for UGLS

• Reduce error and assess uncertainty in
  precipitation analysis
• Correct bias in precipitation observations
• We are asked to provide a dynamical assessment of
  uncertainty in the precipitation analysis
• Use streamflow and lake level observations to
  avoid drift
• Particle filter
• Main source of uncertainty remains precipitation
• What we plan to explore
• Use the future regional, mesoscale, ensemble
  prediction system to build an ensemble of
  precipitation analyses and as the corner stone
  for the particle filter

17
Decision support system for ensemble flood
forecasts in Québec

• Stochastic dynamic programming for regulated
  watersheds (dams and hydropower houses)
• Can take advantage of forecasts with low skill as
  long as uncertainty in correctly represented
  (good spread, little bias in PQPF)
• Flood warnings on unregulated basins
• Acting on flood warnings (evacuation orders,
  preventive downstream flooding) requires a lot of
  confidence in the precipitation forecast
• Difficult to take advantage of forecasts with low
  skill in precipitation forecast
• Unregulated basins in populated areas are
  relatively small (1000 km²) short response time
  (a few hours to one day)

18
Deterministic vs probabilistic hydrological
forecasts for days 1-3
Blue MAE meso-global Red CRPS ensemble 17
cases, October 2007
Lac Saint-Paul
Mitchinamecus Kiamika
Project led by François Anctil, U. Laval
19
Deterministic vs probabilistic hydrological
forecasts for days 1-3

• Benefit of using hydrological ensemble forecasts
  increases with lead time
• Spread of hydrological ensemble forecasts too
  small
• Variational data assimilation uncertainty in the
  stateof the watershed not taken into account
• We plan to move to a particle filter

Blue MAE meso-global Red CRPS ensemble 17
cases, October 2007
Lac Saint-Paul
Mitchinamecus Kiamika
Project led by François Anctil, U. Laval
20
Ensemble urban prediction system

• To improve water quality in urban rivers and
  reduce flood risks, surface runoff is accumulated
  in stormwater ponds and suspended particles fall
  to the bottom
• Water looks cleaner, but smallest, most toxic
  particles generally dont have time to settle

Spillway for extreme events (T10-100 yrs)
Runoff from lawns streets
Normal outflow to the river
Large Smaller particles particles (most toxic)
21
Ensemble urban prediction system

• If we know that large precipitation events are
  unlikely in the next few days, we can accumulate
  runoff from one storm and let it stand for 1-3
  days
• Urban river watershed size 1000 km²
• For case study (July 2006, Québec City), accurate
  3 day forecast for total precip needed to filter
  all runoff events

Spillway for extreme events (T10-100 yrs)
Runoff from lawns streets
Normal outflow to the river
Large Smallest particles particles (most
toxic)
22
Ensemble urban prediction system

• Volume of new stormwater ponds could potentially
  be divided by a factor of two with same
  reliability with respect to flood risk and same
  water quality in outflow
• Good spread and little bias in PQPF required
• Skill of PQPF must be high, and must be known
• Project led by Peter Vanrolleghem, Université
  Laval

Spillway for extreme events (T10-100 yrs)
Runoff from lawns streets
Normal outflow to the river
Large Smallest particles particles (most
toxic)
23
Summary Meeting user needs

• Calibrated PQPF
• at the basin scale, not for individual gauges
• skill must be known for the basin of interest
• Mesoscale ensemble forecasts
• for a dynamic assessment of the uncertainty in
  the precipitation analysis
• for hydrological data assimilation
• to improve the quality of PQPF
• to resolve smaller watersheds