Review of Extended Streamflow Prediction of the
National Weather Service River Forecast System
Joseph A. Pica
Portland State University
CE 505 Read & Conference : NWSRFS ESP
The purpose of this project is to review the technique of extended streamflow prediction(ESP) and provide examples of its current use by the National Weather Service(NWS). ESP's role in the NWS River Forecast System(NWSRFS) is outlined. A sample ESP run is presented and interpreted, then existing and potential products resulting from such runs are mentioned. The ESP Analysis and Display Program(ESPADP) will be shown as the latest interactive tool for evaluating ESP runs and producing useful information for NWS customers. Finally, future developments in ESP are discussed. They include an ESP verification procedure and the ability to objectively incorporate long range climate forecasts into the ESP simulation.
In the early 1970's, the California-Nevada River Forecast Center(RFC) and the State of California first used the ESP procedure for water supply forecasting. The NWS Hydrologic Research Laboratory began its own program to develop an ESP program in 1975. This project called for the development and testing of a procedure to predict streamflow volume over periods of time ranging from a week to an entire season. Typically, these are periods beyond which the NWS has precipitation and temperature forecast ability. In addition, statistical evaluation of the predicted volume was to be included in the procedure (Day 1985).
The ESP procedure which was developed involved running deterministic hydrologic model simulations over the time period for which the volume forecast was desired. Deterministic hydrologic models require precipitation, temperature, and potential evaporation input for each time step of a simulation. Today, the NWS forecasts precipitation for three days and temperature for five days. ESP input includes these short-term forecasts and uses historical precipitation, temperature, and potential evaporation time series for the rest of the simulated period. This assumption that past meteorological conditions are representative of what may occur in the future is the basis of the ESP technique. Conditional simulations for each year of available historical record are made by running the model with the initial conditions being the model state variables on the starting day of the simulations. This ensemble of simulations can then be analyzed statistically. The most familiar output is a frequency distribution of the volume generated. Other output information includes maximum flows , minimum flows, mean flows, number of days to maximum/minimum flows, and number of days until a criterion is reached.
Error in the ESP procedure is due to climate variability, model/calibration error, and data errors. Historical temperature, precipitation, and potential evaporation time series are estimates of future conditions. Deterministic hydrologic models are only approximations of the physical systems they represent and can not be perfectly calibrated. Differences can also be accounted for by inaccuracies in the initial model states and bias in the observed streamflow from the rating of flows.
ESP's Role in NWSRFS
When the NWS began the NWSRFS project in 1979, it was decided that ESP would be an integral part of NWSRFS. NWSRFS is a collection of modules including hydrologic and hydraulic models and techniques, data manipulation methods, analytical and statistical tools, and computer programs for all of the above. The three main systems in NWSRFS are the calibration system, operational forecast system (OFS), and ESP. Each system utilizes the same hydrologic models. In the calibration system, precipitation, temperature, and potential evaporation time series are created from historical data. Model parameters are selected which enable these models to best simulate the observed historical streamflow. Next, the calibrated model is defined in the OFS. With real-time hydrometeorological data and short-term forecast information, the model is run daily to generate short-term streamflow forecasts and maintain model states. ESP provides the long-term forecast mechanism in NWSRFS. Conveniently designed, the calibration system and OFS lay the groundwork for making an ESP simulation. ESP runs the same models with the initial model states from the OFS and the historical time series generated in the calibration system to generate an ensemble of conditional simulations. A probabilistic long-term streamflow forecast is made from the statistical analysis of these simulations. Since 1984, ESP has been included in the official releases of NWSRFS (Day 1985).
Sample ESP Run and Products
Appendix A contains the summary for the April 1, 1997, ESP batch run for the inflow to Dworshak Reservoir on the North Fork of the Clearwater River, Idaho. The purpose of this run was to compute the April-July volume forecast. The first summary page for the simulation shows the observed, historically simulated, and conditionally simulated volumes for each year of meteorological data(1950-1988). General statistical information is shown following each years' data and is summarized on the following page in a frequency table. The ESP April-July volume forecast on April 1 for Dworshak inflow was 5.4 million acre-feet with the 75% and 25% exceedance volumes at 5.2 and 5.6 million acre-feet, respectively. The 75% exceedance statistic means that the probability of the streamflow volume being greater than 5.2 million acre-feet is 75%. The third page of the ESP simulation is the exceedance probability plot which is meant to illustrate the inability of the model to simulate the variability in the observed streamflow. For Dworshak inflow, the historical simulations at lower volumes are higher than the observed data, while the simulations at the highest volumes are slightly lower than the observed data. The conditional simulation indicates that this year's volume forecast is higher than any in the calibrated period of record. Other batch runs could be made for Dworshak Reservoir inflow indicating other statistical information, but it would require runs specifically designed to output those variables for a specific time period.
Currently, several RFCs are using ESP runs in making water supply forecasts and other products. The Colorado Basin RFC produces its standard water supply forecast using ESP, with an example shown in Appendix B. This RFC also sends its ESP traces to Denver Water for use in management of their reservoirs. Graphical volume forecast products from the Arkansas-Red Basin RFC are also shown in Appendix B. The first product indicates the exceedance probabilities of volume for Pueblo Reservoir on the Arkansas River, Colorado, for the period from April 15 to May 14, 1997. The second product indicates several April-September raw ESP forecasts versus the average volume and the actual water supply volume forecast. The Northwest RFC is in the process of initializing basins in Western Washington and Oregon in the ESP system. For the past several years, Dworshak Reservoir inflow forecasts have been performed by the NWRFC to evaluate the ESP procedure and to provide an independent, alternative volume forecast to official water supply forecast procedures. At the North Central RFC, a project is being done with a local government in Iowa which couples the ESP maximum flow forecast with a geographic information system to create a map of the probability of the extent of flooding. Today's tendency is for future products in ESP to be graphical and to be available to the public on the World Wide Web.
The ESP Analysis and Display Program is a graphical user interface which has been developed to allow forecasters more flexibility and ease in examining ESP runs. The specification of each of the individual statistics and time periods in the ESP batch run is no longer required. However, a general ESP batch run must still be completed with the conditional simulations being stored to a file for use by ESPADP. In the ESPADP interface, the user can specify the basin of interest, applicable time period, type of plot, statistics, and type of distribution. The applicable time period must be contained in the window specified in the ESP batch run. Types of plots available include an exceedance probability plot, a trace ensemble plot, and a probability interval plot. Samples of these plots as well as menu screens from ESPADP are shown in Appendix C. Statistics include those already mentioned: accumulated volume, mean flows, maximum flows, minimum flows, days to maximum/minimum flows, and days to a reach a criteria. For each statistic, normal, log-normal, and empirical distributions can be shown. The above options have all been available in the batch run in tabular form, if not graphical.
New options in ESPADP allow time series to be subjectively weighted. Menus allow the user to input weights for an individual years' trace or input the climatological forecasts by month for precipitation and temperature. These options might be used if similarity does or does not exist between the meteorological conditions of the current year and past years. Future options will include objective methods of incorporating long range climate forecasts and ESP verification information. Another option which is to become available in ESPADP is the objective adjustment of the conditional simulations based on the bias between historically simulated and observed streamflows.
ESP verification is intended to give an indication of the quality of the basin calibration and the resulting skill in ESP forecasts. For deterministic forecasts, this is done by judging how well the forecast matches the observed streamflow for a few events. Probabilistic(ESP) forecasts provide a forecast distribution and do not have a single value against which to compare the observed streamflow. Instead, a verification technique has been developed which generates ESP traces for historical dates so that ESP forecasts can be reconstructed for those dates. Each year is assumed independent and is excluded from the data base when its traces are generated. Traces are one year in length and generated for each historical year. Several types of analysis have been suggested for comparing these traces to observed streamflow with assumptions having to be made about the conditional streamflow distribution and/or the forecast. For example, one suggestion is to compute the root mean square error assuming that the forecast is the conditional mean. More work still needs to be done to incorporate these analytical tools and implement verification in ESPADP (Day 1992).
ESP and Climate Forecasts
As mentioned previously, ESPADP allows for the subjective weighting of time series or entry of climate patterns. More knowledge is continuing to be acquired about global climate patterns. As the skill in long range climate forecasts increases, the value of incorporating these forecasts into ESP will need to be quantified. Work is proceeding toward the automatic, objective entry of these long range climate forecasts into ESP simulations (Ingram 1995).
The technique of ESP has not changed significantly since it was introduced in the 1970s. However, the use of ESP has expanded with technology. Graphical user interfaces and other NWSRFS applications have led to improvement in calibration, operational forecasting, and ESP forecasting of basins. Currently, work is being done to enhance ESP forecasting by providing verification information to the calibrator and users. In the future, coupling the ESP procedure with long range climate forecasts may continue to provide for improved forecasts.
Day, G.N., "Extended Streamflow Forecasting Using NWSRFS". Journal of Water Resources Planning and Management. ASCE, 111(2), 1985, pp. 157-170.
Day, G.N., Brazil, L., McCarthy, C.S., and Laurine, D.P., "Verification of the National Weather Service Extended Streamflow Prediction Procedure". 28th Conference and Symposium on Managing Water Resources During Global Change. Reno, Nevada, November, 1992.
Ingram, J.J., Hudlow, M.D., and Fread, D.L., "Hydrometeorological Coupling for Extended Streamflow Predictions". American Meteorological Society Conference on Hydrology. Dallas, Texas, January 15-20, 1995.
Appendix A. Dworshak Reservoir Inflow ESP Run
FORECAST PERIOD: 3/31/1997 HR 24 PST TO 7/31/1997 HR 24 PST
DATA TYPE: RQIM DATA TYPE: SQIN
OBSERVED HISTORICAL CONDITIONAL
YEAR IN ACFT IN ACFT IN ACFT
1950 4033944.0 3822667.5 5367845.5
1951 2813877.3 2420716.5 5023380.5
1952 2976403.3 2542157.8 5270430.0
1953 2626064.0 2378813.5 5521924.0
1954 3501919.8 2568428.5 5519729.0
1955 2988064.8 2622009.3 5454243.5
1956 3907757.0 3799167.5 5172734.5
1957 3165882.3 2780466.8 6076959.5
1958 2774030.5 2801860.5 6038300.5
1959 3147655.5 3099145.3 5260600.5
1960 2708891.8 2948237.0 5596764.5
1961 2895140.0 2801910.8 5733974.5
1962 3030769.0 2945538.8 5341840.0
1963 1799436.3 2224735.3 5417343.5
1964 3401339.0 3631368.5 5712018.0
1965 3308037.0 3914234.5 5368249.5
1966 2149913.3 2358093.5 4827566.0
1967 2691557.0 2824496.8 5070133.0
1968 1972155.8 2306428.5 5294553.5
1969 2877149.3 3449055.0 5284922.0
1970 2619220.3 2971245.0 5520191.0
1971 4052114.3 4050469.8 5712626.0
1972 4677637.0 5125293.5 5413742.0
1973 1342208.5 1497248.1 4769484.5
1974 4657367.0 4081446.5 5274191.5
1975 3221260.8 3279472.0 5149320.0
1976 3374385.0 3804729.5 5036868.0
1977 1242480.1 1435692.8 5075921.5
1978 2407803.5 2874075.0 5422376.5
1979 2696872.8 2626689.8 5343793.5
1980 2289668.0 2651142.0 5781139.0
1981 2152103.5 2432422.5 6005621.5
1982 3328429.0 3290329.0 5243112.0
1983 2122701.0 2234951.5 5128926.0
1984 2762367.3 2823679.0 5606539.0
1985 2948136.8 2520672.8 5300330.5
1986 2031423.5 2648314.5 5217168.5
1987 1510372.1 1735114.4 5134703.0
1988 1752965.6 2153361.8 5469743.5
MEAN 2819474.3 2883997.3 5383572.5
ST DEV 810903.6 749484.8 300505.6
MINIMUM 1242480.1 1435692.8 4769484.5
MAXIMUM 4677637.0 5125293.5 6076959.5
FITTING DISTRIBUTION: LOG-NORMAL
OBSERVED HISTORICAL CONDITIONAL
EXCEEDANCE VALUE VALUE VALUE
PROBABILITY IN ACFT IN ACFT IN ACFT
.95 1704048.5 1832927.3 4903919.0
.90 1887923.0 2011411.1 5004352.0
.75 2240615.3 2349374.3 5176835.0
.50 2709632.5 2791281.8 5375208.0
.25 3276827.0 3316310.0 5581182.0
.10 3888987.3 3873526.3 5773546.5
.05 4308626.5 4250716.0 5891789.5
DWORSHAK RES INFLOW
SEGMENT : DWRI1 + FORECAST PERIOD: 3/31/1997 HR 24 PST TO 7/31/1997 HR 24 PST OUTPUT VARIABLE : ACCUMULATED VALUE + OUTPUT ID : A FITTING DISTRIBUTION:LOG-NORMAL 6272986. .........+...+.....+.......+......+........+.....+.....+....+....+.....+.....+........+......+.......+.....+...+......... . * * * C . . * C . . * * *C . . *** C . 5318354. + *** C + . * * C X . . *C * H . . * * H . . + O+H . 4508999. + O H + . OH . . OH . . + XH . . + XH . 3822813. + X X XOH + . H . . X HO . . X H + . . +H+ + . 3241051. + XX + . +X . . +H+ . . + ++O . . X+ O . 2747824. + ++ + O + . ++ + H O . . XX H O . . XH O . . X XH + O . 2329656. + X X H O + . X X H O+ . . X H + + . . H O+ . . H O+ . 1975125. + H O+ + . H O . . H O . . H O + . . H O X + . 1674548. + H O + . H O . . H O . . H O X + . . H O . 1419713. + O X + . O . . O + . . O . . O + . 1203659. .........+...+.....+.......+......+........+.....+.....+....+....+.....+.....+........+......+.......+.....+...+......... .995 .99 .98 .95 .90 .80 .70 .60 .50 .40 .30 .20 .10 .05 .02 .01 .005 ACFT EXCEEDANCE PROBABILITY LEGEND FITTED EMPIRICAL DISTRIBUTION DISTRIBUTION HISTORICAL SIMULATION H X CONDITIONAL SIMULATION C * OBSERVED O +
Appendix B. Sample Products from the Colorado Basin and Arkansas-Red Basin RFCs
1. Colorado Basin Final Water Supply Forecast, May, 1997 ZCZC SLCESPSLR TTAA00 KSLR 070000 :National Weather ServiceABV GRAND JUNCTION 148 124 162 170 65 50 170 GUNNISON BASIN 184 138 157 190 70 40 170 ABV CISCO (TOTAL) 176 133 160 180 70 45 170 SAN JUAN ABV BLUFF (TOTAL) 206 161 98 230 75 40 150 C. McCarthy/J. Smith/D. Van Cor CBRFC (801-524-5130) NNNN
:Colorado Basin River Forecast Center :Salt Lake City Utah :May Final Forecast May 6, 1997 :** FINAL FORECASTS APPEAR APPROXIMATELY 7 DAYS AFTER PRELIMINARY FORECASTS ** .B SLC 970801 M DH24/DC9705061800/DVM03/QCVFEZ5/QCVFEZF/QCVFEZT :FLOOD CONTROL RESERVOIR UNREGULATED INFLOW FORECASTS :1 MAY THROUGH 31 JULY 1997 (units: 1000's ACRE-FEET) :RESERVOIR MOST :ID NAME PROBABLE MAXIMUM MINIMUM
HODA3 :LAKE MEAD :/ 11000/ 13525/ 8670 GLDA3 :LAKE POWELL :/ 10733/ 13078/ 8523 NVRN5 :NAVAJO RES :/ 977/ 1202/ 802 VCRC2 :VALLECITO RES :/ 227/ 274/ 179 BMDC2 :BLUE MESA RES :/ 895/ 1110/ 715 GRNU1 :FLAMING GORGE :/ 1634/ 2114/ 1174 .END .B SLC M DH24 /DC9705061800/ .B1 DY970531 /QCMFEZ5/DRE+1/QCMFEZ5/DRE+2/QCMFEZ5 : MONTHLY DISTRIBUTION OF FLOW (1000's A-F) : OBS FORECAST OUTLOOK : Jan Feb Mar Apr %AVG May Jun Jul Ap-Jul %AVG GLDA3 :LAKE POWELL 458 424 1049 1267 126% :3580/4750/2403/ 12000 155% SAPC2 :BLUE MESA 29 23 43 105 140% : 310/ 400/ 185 1000 143% MPSC2 :MORROW POINT : / / / 1100 143% CLSC2 :CRYSTAL UNREG** 36 31 63 136 134% : 400/ 535/ 230/ 1300 144% GBRW4 :FONTENELLE 43 33 83 98 101% : 285/ 535/ 332/ 1250 147% GRNU1 :FLAMING GORGE 59 52 239 166 98% : 480/ 750/ 404/ 1800 151% VCRC2 :VALLECITO 7 5.5 13.6 23 115% : 85/ 100/ 40/ 250 130% NVRN5 :NAVAJO UNREG 25 28 149 153 94% : 435/ 405/ 135/ 1130 146% VCRC2 :vallecito chg stg 5 4 -10 5 : 27/ 52/ 7/ CHUN5 :Azotea tunnel flo 0 0 9 15 : 32/ 32/ 3/ NVRN5 :NAVAJO REGLTD * 20 24 150 133 : 376/ 321/ 125/ .END
: * REGULATED FORECASTS ARE DERIVED FROM ANALYZING THE RECORD : AND COMPARING IT TO AVERAGES. THESE FORECASTS HAVE AN : ADDITIONAL SOURCE OF ERROR BECAUSE THE RESERVOIR AND TUNNEL : OPERATORS MAY SUBSTANTIALLY DEVIATE FROM THE ESTIMATED : REGULATION.
: ** UNREGULATED CRYSTAL INFLOW COMBINES BLUE MESA UNREGULATED : INFLOW PLUS THE SIDE INFLOW TO BOTH MORROW POINT AND CRYSTAL PRECIPITATION SUMMARY - % OF AVERAGE BY MONTH - WY 1997 RIVER BASIN: OCT NOV DEC JAN FEB MAR APR GREEN ABV FLAMING GORGE 102 112 253 160 55 70 130 ABV GRN RVR,UT(TOTAL) 125 127 212 150 55 50 150 COLORADO
2. Arkansas-Red Basin Water Supply Products
Appendix C. ESPADP SCREENS
1.Control Screen - User preferences, basin selection, and general display/adjustment selections
2. Exceedance Probability Plot
3. ESP Trace Ensemble
4. ESP Probability Interval Plot
5. Double Plot Option
6. Climatology Adjustment Menu
7. Weighting of Yearly Time-Series Menu