Ensemble River Flow Forecasting Forced by GEFS Runoff Forecasts: Preliminary Experiments

1
Ensemble River Flow Forecasting Forced by GEFS
Runoff ForecastsPreliminary Experiments

• Dingchen Hou, Kenneth Mitchell, Zoltan Toth,
• Dag Lohmann and Helin Wei
• Environmental Modeling Center/NCEP/NOAA
• 5200 Auth Road Camp Springs, MD 20746
• SAIC at EMC/NCEP/NOAA
• Risk Management Solution Ltd.
• Acknowledgement
• Dongjun Seo, Pedro Restrepo and John Schaake,
  OHD/NOAA
• George Gayno, Yuejian Zhu, Jesse Meng and Youlong
  Xia,
• EMC/NCEP/NOAA

2
Current NCEP Routines (No Stream Flow
Forecast)Operational GEFS real time
NLDAS quasi-real timeForecast (Ensemble)
Analysis (Single)

• Background
• Land Surface component of NWP systems
  (Mitchell et al, 2004) facilitates stream flow
  forecast in NWP.
• River routing experiment in analysis mode of
  the North American Land Data Assimilation System
  (NLDAS) project (Lohmann et al, 2004) revealed
  potential benefit of river flow forecast in NWP.
• Existence of uncertainty in initial
  conditions, model structure and forcing needs to
  be considered with an ensemble approach.

3
Purpose and Strategy

• Purpose
• Demonstrate feasibility of gridded river flow
  forecast in operational GEFS.
• Establish optimal configuration for river
  routing.
• Develop suitable strategy to account for
  uncertainties.
• Develop suitable methods of post processing and
  generating final products.
• Validate model output of stream flow.
• General Strategy
• NLDAS stream flow analysis used as initial
  condition and verification
• Water management issues avoided by focusing on
  natural flow forecast
• Global domain in mind with domestic and
  international users
• River flow forecast capacity as a component of
  the ESMF system
• Hind cast data set to be generated for post
  processing.

4
Ensemble Stream Flow ForecastTwo Possible
ApproachesA)
B)
5
Configuration

• Design of Preliminary Experiment (Approach A)
• River Routing Model linear program, distributed
  approach, same as used in NLDAS (Lohmann et al.,
  1998, 2004).
• CONUS domain, 1/8 degree grid size (same as
  NLDAS).
• River Flow Direction Mask A D8 model, river
  stream in each grid point is discharged to 1 of
  the eight main directions (Lohmann, et al, 2004).
• Forcing Runoff from global ensemble forecasts
  (GEFS) and the high resolution control forecast
  (GFS), interpolated to NLDAS grid.
• Downscaling not considered yet
• Uncertainty considered in river routing only in
  the forcing, added partially.
• No initial uncertainty is included yet
• Hydrological model error is ignored but
  systematic model errors can be corrected via post
  processing
• Initial Conditions NLDAS Analysis.
• Evaluation For this preliminary test, natural
  flow is predicted and evaluated against NLDAS
  analysis, which is generated from observed
  precipitation.

6
Forecast Example (initiated April 1st, 2006)
Stream Flow Forced by GFS,GEFS Forecast and
NLDAS ProductForecast Starting at 00Z, April
1st, 2006. Lead time 15 days
GFS High Res. control
Ensemble (low res.) Control
Single control forecasts similar to each
other Ensemble mean is similar to the
analysis. This suggests the ensemble mean has
its value in stream flow forecast.
Ensemble Mean
Analysis (NLDAS)
7
Forecast Example (initiated April 1st, 2006)
Stream Flow Stream Flow, Analysis and Ensemble
Mean ForecastError of Ensemble Mean and Ensemble
SpreadForecast Starting at 00Z, April 1st, 2006.
lead time 12 days
Analysis (NLDAS)
Ensemble Mean
Ensemble mean is similar to the Analysis Geograp
hic distribution of positive and negative
errors. Spread and error are Comparable. Note
the scales for error and spread is 1/10 of that
for analysis and Ensemble mean.
ErrorEns. Mean - analysis
Ensemble Spread
8
May 4th
Mississippi, River Vicksburg, MS The Large
Basin May 4th case A major mid-range event well
predicted Significant spread in extended
range April 1st case With out a major event,
all simulations are similar and spread is
small. Trend and events picked up. Short lead
time dominated by initial condition, showing
little spread. Spread Increases with time.
----- GEFS members ----- GEFS ens. mean -----
GEFS control ----- GFS high resolution ----- NLDAS
0 2 4 6
8 10 12 14
16
Lead Time (days)
April 1st
9
Potomac River A Medium Sized Basin In both
cases Single forecasts are insufficient. Non-li
near evolution of ensemble members help to
improve forecast and catch major flood events.
May 4th
----- GEFS members ----- GEFS ens. mean -----
GEFS control ----- GFS high resolution ----- NLDAS
0 2 4 6
8 10 12 14
16
Lead Time (days)
April 1st
10
Merrimack-Concord River Lowell, MA A Medium Sized
Basin Major Problem Underdispersive ensemble in
grid and subgrid scale precipitation. Mid-May
Flood Event Compared with the Early-April event,
the Mid-May event is harder for the model to
simulate. Nevertheless, the ensemble shows some
skill indicating a major event with 10 day lead,
various amplitude and timing. Early April Major
event is correctly forecast despite short range
over- forecast and insufficient spread.
----- GEFS members ----- GEFS ens. mean -----
GEFS control ----- GFS high resolution ----- NLDAS
May 4th
0 2 4 6
8 10 12 14
16
Lead Time (days)
0 2 4 6
8 10 12 14
16
Lead Time (days)
April 1st
11
Nehalem River, FOSS OR A Small Basin A challenge
for the model. April 1st, large forecast
discrepancy from day 1 despite significant spread

• Possible causes of the problem in the short range
  forecast
• Lack of spread in precip. fcst. on grid and
  subgrid scale.
• Spatial and temporal resolution of the runoff.
• Bias of precipitation (and runoff) forecast

----- GEFS members ----- GEFS ens. mean -----
GEFS control ----- GFS high resolution ----- NLDAS
April 1st
0 2 4 6
8 10 12 14
16
Lead Time (days)
May 4th
12
Summary of Results

• Distributed river routing ensemble system
  (coupled GEFS, NOAH and the Lohmann River Flow
  model) works.
• Spread Error of ensemble mean forecast
• Large basin forecasts look reasonable.
• GEFS provides reasonable forcing?
• Downscaling and river initial condition error do
  not matter?
• Medium/small basin forecasts suffer from initial
  underdispersion.
• Downscaling and initial perturbation needed.
• Systematic errors in hydrological models to be
  addressed via post processing using hind casts.
• Anomalous river flow forecasting.
• Uncertainty in hydrological model to be addressed
  via
• Multi model
• Post Processing (Bayesian and Frequentist)

13
Background
14
Forecast Example (initiated April 1st, 2006)
Stream Flow Stream Flow, Analysis and Ensemble
Mean ForecastAbsolute Error of Ensemble Mean and
Ensemble SpreadForecast Starting at 00Z, April
1st, 2006. lead time 12 days
Analysis (NLDAS)
Ensemble Mean
Same as previous slide except for the error,
where absolute value is plotted to compared with
the spread. Spread is comparable to error, but
the value is smaller, especially in the West.
Absolute Error
Ensemble Spread
15
Further Development----Issues

• General Strategy
• Cooperation with HEPEX, NAEFS and THORPEX.
• Facilitate probabilistic forecasting.
• Configuration
• Expand to global domain (Initially at 0.5 degree
  resolution).
• Simplified lumped approach.
• Coupled and Uncoupled Approaches
• Coupled (Approach A) A complete forecast system
  of the Atmosphere-Land Surface-River Routing
  System.
• Uncoupled (Approach B) Facilitate bias
  correction of precipitation and separation the
  uncertainty in forcing and in the hydrological
  (land surface) model.
• Bias Correction
• Precipitation for better input runoff.
• Stream flow output for better final product.
• via a hind-cast data set for a better estimate of
  bias (Toth).
• Representation of Uncertainties in
• Precipitation (subgrid downscaling, stochastic
  perturbations).
• Land surface and river routing model (multi model
  or varying parameters).
• Initial state of surface water (how?).
• Evaluation
• Quantitative evaluation with rigorous statistics
  (EMC).