Streamflow Forecasting from Weather to Climate Scales
 Current US Practice, Challenges and Opportunities
Andy Wood
NCAR Research Applications Laboratory
Boulder, Colorado
Helmholtz Environmental Centre
Leipzig, Germany, Oct 13, 2014
1
Outline
Operational Seasonal Forecasting – western US
• objectives
• processes
Predictability Research
Opportunities for Improvement
The Arid Lands
Many droughts will occur; many seasons in a long series will be
fruitless; and it may be doubted whether, on the whole, agriculture
will prove remunerative.
John Wesley Powell, 1879
Report on the lands of the arid region of the
United States
NCAR
Boulder, Colorado
Precipitation,
1971-2000
3
Seasonal streamflow prediction is critical
One example:
Met. Water Dist. of
S. California (MWD)
MWD gets:
Surplus
1.25 MAF
OR
0.55 MAF
+$150M gap
from B. Hazencamp, MWD
?
4
How are forecasts created?
National Weather Service River Forecast Centers
Thirteen River Forecast Centers
Established in the 1940s for water
supply forecasting
Three primary missions:
1. Seasonal Water supply
forecasts for water management
2. Daily forecasts for flood,
recreation, water management
3. Flash flood warning support
NWS River Forecast Process
Weather and Climate
Forecasts
Hydrologic Model Analysis
hydrologic
expertise &
judgment
Forecast
precip / temp
model
guidance
River
Forecast
Modeling
System
Analysis &
Quality Control
Observed
Data
Outputs
Graphics
parameters
Calibration
River
Forecasts
Rules, values,
other factors,
politics
Decisions
Data Processing
•
•
•
•
•
•
Data collection
collation
Data assimilation
Quality control
Forcing construction (HAS)
Forecast Review
Observed forcings
Precipitation + Temperature
+ Freezing level + ET
•
•
•
•
•
•
Modeling
•
•
•
•
•
•
Dissemination
•
•
•
•
•
•
Model simulations
Hydrologic analysis
Hydraulic analysis
Bias & error adjustments
Regulation Integration
Collaboration
MPE
Multi-Precip Estimator
MM
Mountain Mapper Daily QC
Process methods vary
Interactive Processes
Threshold and spatial comparisons
Systematic estimation
Manual forcing and over-rides
Hourly and max/min data (temperature)
Quality Control (QC) and processing
procedures vary by region and RFC
Interactive methods
GFE
Output qualified for model use
Gridded Forecast Editor
Slide by H. Opitz, NWRFC
Forecast generation
Quality assurance
Dissemination
Web services
Coordination & collaboration
Basin monitoring & review
National Center Support
Forecast Forcings
WFO Weather Forecast Office
National
Model
Guidance
National Model guidance
available to WFO & RFC
WFO & RFC directly share gridded
forecasts via GFE exchange
Grids at 4km or 2.5km
GFE
Gridded Forecast Editor
RFC River Forecast Center
Slide by H. Opitz, NWRFC
RFCs employ grid and point
forcings to hydrologic model
Point or gridded data converted
to mean areal forcings
Data Assimilation
•
•
•
•
•
Data collation
Data assimilation
Quality control
Forcing construction (HAS)
Forecast Review
Modeling
•
•
•
•
•
•
Model simulations
Hydrologic analysis
Hydraulic analysis
Bias & error adjustments
Regulation Integration
Collaboration
NWS: Streamflow Forecasting is an iterative & interactive process
Analysis & Assessment
Bias adjustments and error corrections
Dissemination
•
•
•
•
•
•
Forecast generation
Quality assurance
Dissemination
Web services
Coordination & collaboration
Basin monitoring & review
Also: SNOW-17 Temperature
index model for simulating
snowpack accumulation and melt
Simulation and Analysis
Observed Hydrology
Sacramento Soil Moisture
Accounting Model
Slide by H. Opitz, NWRFC
Runtime Modifications River Forecast Centers
•
•
•
MOD capability has been available in the NWS >30
years
Generic MOD capability implemented within FEWS
Extend capability to other users outside of OHDcore models
Slide by H. Opitz,
NWRFC
Calibration Climatology
CHPS
Observed and Simulated not Tracking
MOD Interface
Result: Improved observed period simulation
Hydrologist render a Run-Time
modification to the SACSMA Model
and increases the lower zone
primary and supplemental states
Examples: Nooksack R
Typical situation during snowmelt: the simulation goes awry
 What can a hydrologist deduce from this simulation?
 As it is, blending simulation and obs gives an ‘unrealistic’ forecast
12
Examples: Nooksack R
One Solution -- Double the snowpack (WECHNG)
 Other approaches may also have been tried: lower temperatures, raise soil moisture, etc.
13
Examples: Nooksack R
The resulting simulation is better, hence the forecast is more confident
 Flows stay elevated, have diurnal signal of continued melt.
14
A manual, subjective process
Manual elements
Meteorological
analysis
Quality control of station data
Quality control of radar and radar parameters
WFO+HPC met.
forecast
Water Supply
Forecast
Hydrologic
simulation and flood
forecasting
Daily Flood Forecast
Process
Long-lead forecasting
Coordination with
USDA/NRCS
WFO forecast itself (though based on models)
RFC merge with HPC forecast (similar to WFO)
Sac./Snow17 model states and parameters
Bias-adjustment relative to obs. Flow
Input forcings (2nd chance at adjustment)
Model states as adjusted for flood forecasting
Choice of models (statistical / ESP )
Blend of models
Choice of meteorology: QPF, ENSO, None?
Merging with NRCS statistical forecasts
means, confidence limits (“10-90s”)
15
Water Supply Forecast Methods
Statistical Forecasting

Statistical Regression Equations

Primary NOAA/RFC forecast method from 1940’s to mid 1990’s.

Primary NRCS/NWCC forecast method

Historical Relationships between flow, snow, & precipitation (1971-2000)

Tied to a fixed runoff period (inflexible)
 Ensemble Simulation Model Forecasting

A component of a continuous conceptual model (NWSRFS)

Continuous real time inputs (temperature, precipitation, forecasts)

Accounts for soil moisture states (SAC-SMA) - drives runoff efficiency

Builds and melts snowpack (Snow-17) – output feeds SAC-SMA

Flexible run date, forecast period, forecast parameters.

Evolving toward ESP as primary forecast tool at NOAA/RFCs
NWS Community Hydrologic Prediction System
New forecasting platform at all RFCs
• New wrapper for old models…but...
• Facilitates research to operations
• gives more insight into the model used
Transition completed in 2011
ESP example -- Feb 1 forecast
Feb 1
Future
Medium chance of this
level snow or higher
Sno
w
Past
Low chance of this
level snow or higher
High chance of this
level snow or higher
Time
• ESP forecasts based on
• (1) initial conditions (e.g. snow pack, base flow, etc)
• (2) historical meteorological scenarios
©The COMET Program
Forecasts and Water Management
CBRFC ensemble flow
forecasts for Reclamation
water management
Average contribution to Lake Powell
Apr-Jul inflow:
Green RiverStakeholder
34%
Reclamation
Allocations
Colorado River 50%
MidtermSan Juan River 13%
Probabilistic
Model
Forecast impacts, e.g.:
if WY12 is wet enough, Lake
Mead hits surplus levels,
MWD gets water ~ $150M
Graphic from Dr. Katrina Grantz, Bureau of Reclamation
19
19
Water supply forecasting
Example flow volume prediction
increasing watershed moisture
reduces prediction spread
Runoff forecast for
this period
Statistical Forecast Equations:
Trial Lake SNOTEL
Predictor variables must make sense
Challenge when few observation sites exist within river basin
Challenge when measurement sites are relatively young
Fall & Spring precipitation is frequently used (why?)
Sample Equation for April 1:
April-July volume Weber @ Oakley =
+ 3.50 * Apr 1st Smith & Morehouse (SMMU1) Snow Water Equivalent
+ 1.66 * Apr 1st Trial Lake (TRLU1) Snow Water Equivalent
+ 2.40 * Apr 1st Chalk Creek #1 (CHCU1) Snow Water Equivalent
- 28.27
Source: NRCS
What are the sources of seasonal
streamflow predictability?
Opportunities for prediction
hydrological predictability
meteorological predictability
Hydrological Prediction: How
well can we estimate the
amount of water stored?
Accuracy in precipitation
estimates
Fidelity of hydro model
simulations
Effectiveness of hydrologic data
assimilation methods
Meteorological predictability:
How well can we forecast the
weather?
Opportunities: Which area
has most potential for
different applications?
Water Cycle (from NASA)
Seasonal predictor importance depends on hydroclimate
humid basin
uniform rainfall
no snow
small cycle driven by ET
cold basin
drier summers
deep snow
large seasonal cycle
April snowmelt
dominates May-June
runoff
Data & Methods
• automated system, large number of gages
• Basin Selection: Used GAGES-II, Hydro-climatic data
network (HCDN)-2009
– minimal human influence, long period of record
– large diversity of landscapes, energy- to water-limited
– ~670 rwatersheds
Assessing the sources of flow forecast skill
w=0.5
http://www.ral.ucar.edu/staff/wood/weights/
w=0.5
Snow-Driven Basin in
the Western US
• Wide seasonal variations
in influence of different skill
sources
• cold and/or dry forecast
period forecast skill
depends strongly on initial
condition influences
• warmer wet/snowmelt
forecast period forecast
skill depends strongly on
met. forecast skill
IHC: initial Hydrologic Conditions
SDF: Seasonal Climate Forecasts
Streamflow Forecast
Skill Elasticities
•
The % change in flow
forecast skill versus per %
change in predictor source
skill
•
Can help estimate the
benefits of effort to improve
forecasts in each area
Opportunities for Improvement
of Operational Forecasts
Improved modeling (ie, IHCs)
flexible
platform
31
Example: point
observations and
lumped model zones
did not capture
difference between
these two years
2010 June 1
2006 Snowbird SNOTEL trace
almost identical to 2010 trace
from June 1-10
However, distributed
satellite/model analysis
indicates south facing slopes
were melted out in 2006 
much lower runoff as a result
2006 June 1
From lumped conceptual models…
SNOW17 SWE
Sacramento
33
To finer resolution physically-based models
Distributed modeling
of water and energy
balance
34
Improved long-lead flow forecasts via climate prediction
Conditioned on IRI seasonal
forecast
Get the prediction
(A:N:B=40:35:25)
Divide historical (seasonal) total
into 3 tercile categories
Bootstrap 40, 35 and 25 sample of
historical years from wet, normal
and dry categories
Apply the ensembles in ESP instead
of historical weather sequences
(from Balaji Rajagopalan, CU)
CFSv2 precipitation forecast skill
• For seasonal time scales,
skill is lower
• some seasons may
have better skill,
varying little across
leads
East R at Almont, Co, precip
43
NWS has a new system for incorporating climate forecasts
ShortRange
HPC/RFC
single-valued
forecast
Day 1-5
ensembles
Medium- GFS/GEFS
ens. mean
Range
Day 1-14
ensembles
CFS/CFSv2
forecast
Day 15+
ensembles
LongRange
Merging
Joining
Blending
Other ensembles
(SREF, NAEFS…)
Under development / MMEFS
Other info
(climate indices…)
Calibrated
and
seamless
short- to
long-term
forcing
input
ensembles
Day 15+
ensembles
Planned
37
Is it time to start shifting the forecast paradigm?
Advantages of a manual process
•
•
•
‘handles’ data and model deficiencies in ways that current objective
approaches cannot
incorporates the knowledge of the expert forecaster in ways current
models and procedures may not
capable of producing ‘the best’ forecasts at times
Disadvantages
•
•
•
•
•
limits addition of new techniques that assume objectivity or automation
• data assimilation, post-processing, verification, ensembles
labor intensive – poor scaling limits opportunity for expanded services
cannot quantify uncertainty of an individual forecast
cannot quantify skill of forecasting system and set baseline for
systematic improvements – ie ‘benchmarking’
the ‘system’ is partly in the forecaster brain, thus system knowledge
can be lost as forecasters retire or move away
Such a shift would require broad institutional change
•
training, expertise, center staffing, partner interactions
38
Questions?
Questions?
Dr. Andy Wood
CBRFC
Development and Operations
Andy.Wood@noaa.gov
http://en.wikipedia.org/wiki/Lake_Powell
Precipitation Skill, CD 48
April 28, 2010
40
Satish Regonda, OHD
Streamflow Prediction Challenges:
forecast future weather (hours to seasons)
Fig. M. Clark (NCAR)
Historical Data
e.g., SNOW-17 / SAC
Historical Simulation
Past
Future
Forecasts
SNOW-17 / SAC
SWE
SM
Q
General State of Practice
• conduct ensemble simulation at fine-time step (e.g., sub-daily); aggregate to client needs
(e.g., daily to seasonal information)
• adjust results to compensate for known model biases (practice varies)
Motivation
Development & Performance
Predictability
Ensemble Forcing
Summary
Assessing regional variations
– Based on ~420 watersheds, 30 year hindcasts
Results are preliminary and subject
to change
Motivation
Development & Performance
Predictability
Ensemble Forcing
Summary
Assessing regional variations
– Based on ~420 watersheds, 30 year hindcasts
Results are preliminary and subject
to change
Motivation
Development & Performance
Predictability
Ensemble Forcing
Summary
Assessing regional variations
– Based on ~420 watersheds, 30 year hindcasts
Results are preliminary and subject
to change
A Snowmelt Basin
• Wide seasonal
variations in
influence of
different skill
sources
• Cold/Warm and
Dry/Wet
patterns are
clear
• Skill increases in
meteorology
produce nonlinear skill
changes in flow
Results are preliminary and
subject to change
Flow Forecast Skill Sensitivities
Assuming no skill in one source (model accuracy
for watershed states, or forecasts)…what are the
benefits of adding skill in the other source?
Flow Forecast Skill Sensitivities
A Rain-Driven Basin
• IC skill matters
most for 1 month
forecasts
• For longer
forecasts, met
forecast skill
dominates
Results are preliminary
and subject to change
Assuming no skill in one source (model accuracy
for watershed states, or forecasts)…what are the
benefits of adding skill in the other source?
Atmospheric Pre-Processor:
calibration
Off line, model joint distribution between single-valued forecast and
verifying observation for each lead time
Archive of observed-forecast pairs
PDF of Observed
Joint distribution
Sample Space
NQT
Observed
Joint distribution
Correlation (X,Y)
0
X
Forecast
PDF of Forecast
PDF of Fcst STD Normal
Y
Model Space
Observed
Y
PDF of Obs. STD Normal
NQT
X
Forecast
47
NQT: Normal Quantile Transform
Schaake et al. (2007), Wu et al. (2011)
Atmospheric Pre-Processor:
ensemble generation (1)
In real-time, given single-valued forecast, generate ensemble traces
from the conditional distribution for each lead time
Joint distribution
Y
Model Space
Conditional distribution
given xfcst
xfcst
X
x1
…
xn
Probability
Observed
1
NQT-1
Ensemble
members
for that
particular
time step
Forecast
0
x1 xi
xn
Ensemble forecast
Given single-valued forecast,
obtain conditional distribution
48
Schaake et al. (2007), Wu et al. (2011)
Abstract
Streamflow Forecasting from Weather to Climate Scales – Current US Practice, Challenges and
Opportunities
Several US agencies are charged with operationally producing seasonal streamflow forecasts, which
are a critical input to water management throughout the US, and particularly in the western US, where
flood-prone winters and dry summers necessitate careful decision-making over water allocation and
use. Water supply forecasts (WSFs) that predict the volume of snowmelt-driven runoff in the spring
and early summer, for example, are a critical product informing the regulation of major storage projects
such as Lakes Powell and Mead. This presentation focuses on the tradition of water supply
forecasting and describes an approach for diagnosing the sources of prediction skill. The talk will
highlight strengths and weaknesses of the current practice, which is highly manual and regionalized,
and potential opportunities for WSF improvement. The presentation also describes the current
practice in flood forecasting, and new developments toward a more centralized modeling and flood
prediction paradigm.
Improved Human Resources
The forecasting and water management enterprise
needs people with a strengths in:
•
•
•
•
•
•
watershed science
land surface modeling
atmospheric and/or climate science
probability and statistics
objective analysis (time-series, spatial)
programming (UNIX, R, Python, ‘traditional’, web,
database)
• technical communication
• systems engineering
• water resources planning, management and policy
50
Calibration Results
• Areas with seasonal
snow perform best
(using NSE
benchmark)
– More arid regions,
high plains perform
worst
– Need in-depth
examination of poor
performing basins
– Forcing data issues?
• ~670 basins
included in initial
calibration
Rain-Driven Basin in
the Southeastern US
•
For 6-month forecast,
seasonal variation in flow
forecast skill sensitivities
(met. forecast or IC) is
minimal
•
Effectiveness of skillful
met. forecasts diminishes
if initial conditions are not
well-captured
Analysis effort outside the forecast process is
critical to the forecast service
Mar 1 forecast
Mar 1
Future
Medium chance of
this level flow or
higher
Sno
w
Past
Low chance of this
level flow or higher
High chance of this
level flow or higher
Time
©The COMET Program
Apr 1 forecast
Medium chance of this
level snow or higher
High chance of this
level snow or higher
Sno
w
Past
Apr 1
Future
Low chance of this
level snow or higher
Time
©The COMET Program
Story line of an extreme year
Each hydrologic anomaly has a story line. As water year progresses:
• past wx/climate (hydrologic initial conditions) are an increasing
part of the plot
• future climate is a diminishing part of the plot
• extremes often involve pattern persistence …
but can be more complicated
Feb 1
e.g., storyline:
• big storm in Feb
• very wet April
• cool May/June
Future
Flow
Past
Wx/Climat
e signal
IC signal
IC
e.g., Wood & Lettenmaier, 2008 GRL
Time
Both frameworks can be skillful
ESP – model-based
PCR - statistical
Yet both have drawbacks
• raw ESP: fails to account fully for
hydrologic uncertainty
• may have bias
• overconfident
• PCR, ie, flow = f(SWE, acc. PCP)
• cannot provide confidence
bounds (heteroskedastic
error) in many places
• small samples
• used since mid-1990s
• Ad-hoc combination
Improved post-processing
• Statistical post-processing of forecasts can correct systematic bias and spread
problems (particularly ESP overconfidence and bias).
• Require long track record of consistent forecasts for training
• Would need to change the current operational ESP approach
April 1 Forecast of April-July streamflow volume, North Fork Gunnison
(SOMC2)
90th
50th
10th
Premise: We can improve shortterm water decisions and
increase long-term adaptation
capacity under climate change by
improving monitoring & forecasting
Improving how we use this
information
Report summarizes needs
monitoring
forecasting
information use
http://www.ccawwg.us/docs/ShortTerm_Water_Management_Decisions_Final_3_Jan_2013.pdf
‘ST Doc’ user needs
Other Categories
60