Post on 21-Dec-2015
Forecasting Streamflow and Reservoir Storage Summer of 2003
Richard Palmer, Andre Ball, Ani Kameenui,
Kasey Kudamik, Michael Miller,
Nathan Van Rheenen, Matthew WileyCEE
University of Washington
October 2003
Talk OverviewA. Background on Forecast Approach
B. Evolving Summer Forecasts
C. Accuracy of Forecast
D. Conclusions
Study Goals
Create six-month forecasts for municipal water supplies in the Puget Sound area using NCEP forecasts:– Water Supply– Water Demand– Storage in Reservoir– Decision Support System
Forecasting
– The herd instinct among forecasters makes sheep look like independent thinkers.
Edgar R. Fiedler
– If you have to forecast, forecast often.Edgar R. Fiedler
– An unsophisticated forecaster uses statistics as a drunken man uses lamp-posts - for support rather than for illumination.
Andrew Lang
Other Quotes
The future will be better tomorrow. Dan Quayle (1947 - )
Where a calculator on the ENIAC is equpped with 18,000 vacuum tubes and weighs 30 tons, computers in the future may have only 1,000 vaccuum tubes and perhaps weigh 1.5 tons. Popular Mechanics, March 1949
- More quotations on: [Computers]
Other Quotes
The best way to predict the future is to invent it. Alan Kay
The future belongs to those who prepare for it today.
Malcolm X (1925 - 1965)
– The future is here. It's just not widely distributed yet.
• William Gibson (1948 - )
Other Quotes
Enjoy present pleasures in such a way as not to injure future ones.
Seneca (5 BC - 65 AD)
The future ain't what it used to be. Yogi Berra (1925 - )
Study Domain
Auburn
Renton
N
Auburn
Renton
N
Cedar River
Green River
Tolt River
Sultan River
Models Used to Generate Forecasts• NCEP Meteorological Forecasts
• Distributed Hydrology, Soil-Vegetation Model
•Dynamic Systems Model
•Water Demand Forecasts
June Demand Forecast (6.24-12.24.03)
100
120
140
160
180
200
220
240
6/24/2003 7/24/2003 8/24/2003 9/24/2003 10/24/2003 11/24/2003 12/24/2003
de
ma
nd
, mg
d
Maximum Minimum25th percentile 75th percentileAverage Forecast using ave Tmax ('83-'02)Actual
Water Demand ForecastingPuget Sound Region
• Modelsshort (weekly-monthly)
and long (annual-decadal)-term
• Regions: Seattle, Tacoma, and EverettTacoma and Everett:
Municipal demandsSeattle: System-wide
demands
Purpose
A common characteristic of water resources planning is its failure to anticipate change. -D. Sewell, 1978
• Increase the accuracy of demand models for effective water resources planning and management.
• Provide information for monitoring and controlling demands during droughts, planning conservation programs, and supply and infrastructure changes.
• Create a framework for long-term forecasting while considering urban planning.
How well have we done?Water demand forecasts: 1968 an 1980 Seattle Water Plans
100
150
200
250
300
350
1966-67 1980 1990 2000projected years
mg
d o
r p
er c
ap g
d
1968 forecast gpd per capita
1968 forecast mgd ann. ave.
1980 forecast mgd ann. ave.
Actual gpd per cap
Actual mgd ann ave
Increase the accuracy of demand models for effective water resources planning and management.
Provide information for monitoring and controlling demands during droughts, planning conservation programs, and supply and infrastructure changes.
Recent HistorySPU Winter water use (1983-2003)
105
115
125
135
145
155
1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003
mg
d
Data Resources
• WATER related– Sources: Seattle Public Utilities; Tacoma Water; City of Everett– Daily water demands– Rate History; Number of users
• CLIMATE– National Climate Data Center (NCDC): SeaTac daily Tmax
and precipitation– National Centers for Environmental Prediction (NCEP):
downscaled climate ensembles
• HOUSEHOLD– Puget Sound Regional Council (PSRC)– Urban simulation group (UrbanSim)
Short-term Model DesignSeattle Region
• Data must be on a weekly time-step• Log-linear regression: Water Demand = Intercept**∙Ax∙Bx2∙Cx3∙Dx4∙Ex5∙Fx6∙Gx7
Ln(Water Demand) = Intercept** + x∙Ln(A) + x2∙Ln(B) + x3∙Ln(C) + x4∙Ln(D) + x5∙Ln(E) + x6∙Ln(F) + x7∙Ln(G)
Dependent variable System (SPU)-wide weekly averages
Independent variables A. Temperature (average weekly max) (Tmax)
B. Precipitation (weekly average)
C. Winter water use
D.* System user population
E. Water rate/price
F. Temperature (max) (one-week lag)
G. System-wide weekly average (one-week lag)
Model CalibrationSeattle Region: Summer
Summer Calibration Model: 1989-1998 (R2 88%)
125
175
225
275
1989 1989 1990 1991 1993 1993 1994 1995 1996 1996 1997 1998time
de
ma
nd
, mg
d
Actual
Predicted
Model ValidationSeattle Region: Summer
Summer validation: 1999-2003
R2 = 0.8483
125
145
165
185
205
225
245
125 145 165 185 205 225 245Actual demand, mgd
Fo
rec
as
ted
de
ma
nd
, m
gd
Model CalibrationTacoma Region: Summer
Tacoma summer calibration: 1990-1999 (R2 83%)
30
40
50
60
70
80
90
1990 1990 1991 1992 1993 1993 1994 1995 1996 1996 1997 1998 1999Time
Dem
and
, mg
d
ActualPredicted
Tacoma summer validation: 2000-2003
R2 = 0.8053
30
40
50
60
70
80
30 40 50 60 70 80 90Actual demand, mgd
Fo
rec
as
ted
de
ma
nd
, m
gd
Model ValidationTacoma Region: Summer
Everett summer calibration: 1990-2003 (R2 84%)
30
45
60
75
90
1990 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2002 2003
time
dem
and
, mg
d
ActualPredicted
Model ValidationEverett Region: Summer
Demand ForecastSeattle Region: April forecast
April Forecast: 4.29-10.22.03
100
120
140
160
180
200
220
240
4/29/2003 5/29/2003 6/29/2003 7/29/2003 8/29/2003 9/29/2003
de
ma
nd
, m
gd
MaxMin25th%75th%AverageActualForecast using ave Tmax ('83-'02)Forecast using real climate ('03)
Forecast Skill and ErrorAverage daily temperature maximum: past vs. present
0
5
10
15
20
25
30
1/1 1/29 2/26 3/26 4/23 5/21 6/18 7/16 8/13 9/10 10/8 11/5 12/312/31week
Tm
ax,
C
2003 Tmax 1983-2002 Tmax ave
Summer 2003 was an climate outlier
•Model is calibrated during less dramatic conditions
•Validated during warming (hence the drop in correlation)
Comparison of Forecast Skill
-2.25
-1.75
-1.25
-0.75
-0.25
0.25
0.75
4/29 5/13 5/27 6/10 6/24 7/8 7/22 8/5 8/19 9/2 9/16 10/1week
skill
met
ric
AprilJuneAugust
Forecast skill metric (Hamlet):
•Skill = 1 - [∑(forecast - observed)2/N / ∑(historical - observed)2/M ]
•Rewards precision, punishes spread
•Valuable metric during outlier years
Create a framework for long-term forecasting while considering urban planning.
– Using PSRC and UrbanSim information from household survey or Parcel Index Number databases
– Highly disaggregated database for modeling household or class specific water demands.
– Incorporate household variables such as: size, income, house age, house value, yard size, etc.
– Investigate benefits and drawbacks of disaggregated model and consider water resources during urban planning and land development (UrbanSim component).
Long-Term Forecasting
Long-term Model DesignSeattle Region
Seasonal (3) Water Data:Aggregated to 12 years
Seasonal (3) Water Data:Disaggregated sample set
Seasonal (3) Water Data:Spatially disaggregated &
aggregated to 12 years each
Seasonal (3) Water Data:matching PSRC accounts
+
++++
Regional climate andhousehold data
Regional climate andhousehold data
Regional climate andhousehold data Regional climate and select
PSRC household data
Decent R^2, poor p-values Unreliable household data
due to aggregation Limited data points (12)
Poor stat performance Numerous data points;
variable water use withstagnant hh & climate data
Terrific R^2 values (75%+), mediocrep-values, questionable coefficients
Unreliable household data due toaggregation
Limited data points (12) for eachaccount
Must match water accountto PSRC wave surveyhousehold data via address
Climate remains regionalbased on 3-seasons andyears of survey data
Disaggregated householdsover 6-years with householdspecific water use, size, andincome
Approaches to Forecasting SPU Demand
Long-term Model DesignCurrent work
Seattle Region
SPUhousehold
billing (1990-2002)
database
UrbanSimPSRC PINdatabase
OUTSIDE:AREA,AGE &VALUE
PEOPLE
INSIDE:PEOPLE
&UNITS
Matched by PIN(bimonthly)
ClimateData:
Tmax andPrecip
Seasonal regressions with random sampleof accounts (5-10,000); R2 values of 50+%
Overview of Meteorological Forecast Process
• National Centers for Environmental Prediction
NCEP Forecast
A set of 20 equally likely ensembles of paired precipitation and temperatures generated by GSM with slight variations in initial conditions
Downloaded from NCEP ftp site
Forecasts bias-corrected and downscaled
DHSVMDistributed Hydrology, Soil-Vegetation Model
• System is initiated with one year of previous conditions – Twenty assembles of paired precipitation and
temperatures are run.– Initial conditions are extremely important
(same future conditions are different with different initiations)
– Typically model underestimate summer flows
Streamflow Forecast
Systems Simulation Model
• Model calculates movement of water throughout system
• Integrates water supply, demands, fish flows and other operational considerations
• Lacks subtleties of actual operation
April Forecast Streamflows on Cedar River at Chester Morse
0
100
200
300
400
500
600
May-03 Jun-03 Jul-03 Aug-03 Sep-03 Oct-03
Av
era
ge
Mo
nth
ly F
low
(c
fs)
0
100
200
300
400
500
600
Median Ensemble Forecast Ave. Actual Hist. Ave
April 2003 Forecast: Total Seattle Reservoir Storage
0
5
10
15
20
25
30
35
40
45
50
4/29
/03
5/13
/03
5/27
/03
6/10
/03
6/24
/03
7/08
/03
7/22
/03
8/05
/03
8/19
/03
9/02
/03
9/16
/03
10/0
1/03
10/1
5/03
Date
To
tal S
tora
ge
Bil
lio
n G
allo
ns
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
45.00
50.00
median Ensembles Forecast Average Observed Storage
Low Reservoir Conditions
April 2003 Forecast: Resolving Discrepancies from Observed Storage
0
5
10
15
20
25
30
35
40
45
50
4/29
/03
5/13
/03
5/27
/03
6/10
/03
6/24
/03
7/08
/03
7/22
/03
8/05
/03
8/19
/03
9/02
/03
9/16
/03
10/0
1/03
10/1
5/03
Date
To
tal S
tora
ge
Bil
lio
n G
allo
ns
ObservedStorage
frcst inputs/frcst demands
act inputs/frcst demands
act inputs/ actdemands
FISH actinputs/ actdemands
.75 morainereturn
Low Reservoir Conditions
Adjusted April 2003 Forecast: Total Seattle Reservoir Storage
0
5
10
15
20
25
30
35
40
45
50
4/29
/03
5/13
/03
5/27
/03
6/10
/03
6/24
/03
7/08
/03
7/22
/03
8/05
/03
8/19
/03
9/02
/03
9/16
/03
10/0
1/03
10/1
5/03
Date
To
tal S
tora
ge
Bil
lio
n G
allo
ns
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
40.00
45.00
50.00
median Ensembles Forecast Average Observed Storage
Low Reservoir Conditions
April Retrospective Forecast Three Month Total Flow Observed vs. Forecast
y = 1.0773x - 38.822
R2 = 0.8468
0
200
400
600
800
1000
1200
1400
1600
1800
0 200 400 600 800 1000 1200 1400 1600
Forecasted Flow (cfs)
Ob
se
rve
d F
low
(c
fs)
April Retrospective Forecast Six Month Total Flow Observed vs. Forecast
y = 1.1435x - 111.51
R2 = 0.6654
0
500
1000
1500
2000
2500
0 200 400 600 800 1000 1200 1400 1600 1800
Forecasted Flow (cfs)
Ob
se
rve
d F
low
(c
fs)
Conclusions
NCEP ensemble forecasts, combined with hydrologic model, produced good summer forecasts for 2003.
Typically, NCEP ensemble forecasts, combined with hydrologic model, provides does useful information (exceptions noted).
Forecasts ranked by ENSO provides some insight into forecast quality