Combining statistical and dynamical methods for hydrologic prediction

1
Combining statistical and dynamical methods for
hydrologic prediction

• Andy Wood
• Seminar
• Land Surface Hydrology Research Group
• Civil and Environmental Engineering
• University of Washington
• Dec 19, 2006

2
Outline

• Ensemble Forecast Calibration
• Synoptic Scale Hydrologic Indices

3
The importance of Seasonal Hydrologic Forecasting
water management hydropower irrigation flood
control water supply fisheries recreation
navigation water quality
4
How does one make a forecast of river flow?

• Naïve forecast (climatology) simply use
  historical averages
• Persistence (of states or anomalies)
• (Multiple) Regression Forecast
• Traditional Predictors
• snowpack (SWE), accumulated precipitation,
  current or past river flow, measured over the
  drainage basin
• More advanced predictors
• ENSO state indicators (Nino3.4, SOI)
• Predictand daily, monthly or seasonal
  streamflow at some lead time in the future.
• Model-based approaches

5
Introduction Hydrologic prediction and the NRCS
PNW
Snow water content on April 1
SNOTEL Network
McLean, D.A., 1948 Western Snow Conf.
April to August runoff
6
Results for Winter 2003-04 volume runoff
forecasts
7
Introduction Hydrologic prediction and ESP

• NWS River Forecast Center (RFC) approach
• rainfall-runoff modeling
• (i.e., NWS River Forecast System,
• Anderson, 1973
• offspring of Stanford Watershed Model, Crawford
  Linsley, 1966)
• Ensemble Streamflow Prediction (ESP)
• used for shorter lead predictions
• used for longer lead predictions
• Currently, some western RFCs and NRCS coordinate
  their seasonal forecasts, using mostly
  statistical methods.

8
Forecast Calibration Hydrologic Simulation
Uncertainty
Simulation error results from -- parameter
uncertainty -- forcing uncertainty -- model
physics/structure Techniques for addressing each
exist -- multi-algorithm approaches --
calibration science -- forcing preparation
techniques Other approaches for improving
simulation -- data assimilation -- multi-model
approaches -- bias-correction
9
Forecast Calibration Effect of Uncertainty on
ESP
Model-based ensemble forecasts contain both
hydrologic uncertainty (associated with the
input data, model parameters physics) and
future climate uncertainty.
ESP accounts mostly for the latter, but not the
former, hence ESP forecasts have an inherent
tendency to be overconfident. One approach that
can be used to correct this is called forecast
calibration.
10
Forecast Calibration Overconfidence example
11
Forecast Calibration Approach
Following a technique suggested by John Schaake
for 15-day temperature forecast ensembles 1.
use only forecast ensemble means 2. correlate
forecast means with observations 3.
reconstruct forecast uncertainty A hindcast
dataset is needed for training of the
parameters. Also - correlation - mean and
variance of hindcasts and observations
12
Forecast Calibration hindcast dataset
13
Forecast Calibration Approach
Algorithm
hindcast long term mean
obs long term mean
one forecast mean
correlation, obs hindcast means
one calibrated forecast mean
obs long term std. dev
hindcast long term std. dev
correlation, obs hindcast means
obs long term variance
calibrated forecast variance
14
Forecast Calibration Results
15
Forecast Calibration State Dependent Approach
16
Forecast Calibration Raw ESP
17
Forecast Calibration Results
calib w/ entire hindcast
calib w/ sample size N35
bias-corr only
18
Forecast Calibration Reliability Improved
19
Outline

• Ensemble Forecast Calibration
• Synoptic Scale Hydrologic Indices

20
UW Real-time Daily Nowcast SM, SWE (RO)
½ degree VIC implementation Free running since
last June Uses data feed from NOAA ACIS
server Browsable Archive, 1915-present
We are currently migrating the daily update
methods to the west-wide forecast system (1/8
degree)
21
The challenge of changing observing systems
Meteorological stations that still report in real
time today
1920s
1990s
22
Surface Water Monitor Archive
March 1997 La Nina conditions bring the
highest recorded snowfall to the PNW
July 2002 the western U.S. drought centers on
Colorado
23
Surface Water Monitor Archive
August 1993 the highest recorded flow on the
Mississippi R.
March 2002 Virginia experiences severe drought,
many well failures
24
Water Year 2005
25
(No Transcript)
26
(No Transcript)
27
(No Transcript)
28
Land Surface Indices
Can capture information from PC1 and
PC2 using NDX1 PC1 CNTR NDX2 PC2
NW-SW Then PCs or NDXs can be used in regression
framework to predict future flow, e.g., summer
runoff
29
Flow prediction results
can we use the modes of variability to predict
summer streamflow?
30
Flow prediction results
31
Take away message

• The dream of a purely physical modeling based
  prediction system is unlikely to be realized due
  to uncertainties in data, parameters, physics and
  so forth.
• Statistical techniques can work hand in hand with
  dynamical ones to move prediction applications
  forward.

32
Forecast Calibration ESP
N