Research
Overview
Currently, three forecast approaches are used to develop streamflow forecasts: a) the Ensemble Streamflow Prediction (ESP, formerly Extended Streamflow Prediction) method; b) ensemble forecasts downscaled from several climate models (the NCEP Coupled Forecast System, NCEP CFS, and the NASA/GMAO Seasonal-to-Interannual Prediction Project (NSIPP) Tier 2 forecasts); and c) statistical downscaling of the CPC official seasonal climate outlooks.
An overview (rather than technical) paper describing the components of the system has been submitted to BAMS (2005/11/18), and more detailed papers on the methods are planned. Questions may also be directed to Dr. Andy Wood, aww@u.washington.edu.
Methods
General
The current streamflow forecasts were made using the Variable Infiltration Capacity hydrologic model, applied for daily simulation of western U.S. hydrology at 1/8 degree grid resolution. The forecast signal of each ensemble is dependent on two elements: a spin-up hydrologic simulation, using recent observed weather, to estimate current conditions; and the meteorological forecasts. The initial state (the most important components of which are snowpack and soil moisture variables) is used as a starting point for the meteorological forecasts (sequences of daily precipitation, minimum and maximum temperature) for each grid cell in the simulation domain. Runoff and baseflow from the hydrologic model cells are routed through a grid-based stream network to produce streamflow at major locations in five basins/sub-regions simulated: Columbia R. (PNW), California, Colorado R., Great Basin and upper Rio Grande R.
Spin-up
The spin-up period varies in each forecast, but is generally 1 to 2 years, from a previous summer to the forecast date (when the hydrologic state of the model is saved, and adjusted by the assimilation of observed SWE.)
Because we do not have real-time access to the dense COOP meteorological station network for the last 2-3 months of the period, spin-up forcings (e.g., daily precip) for the model are estimated from a variety of sources, listed in the table below. Efforts are continually underway to develop improved real time access data during the spin-up period.
Domain |
Spin-up details |
|
US |
Up to 2-3 months prior: NCDC COOP station
based gridded dataset (2-3000 stations in domain), created using
methods described in Maurer et al. (2002). |
|
Canada Portion |
Entire spinup: gridded percentiles/anomalies from the reduced set of index stations applied to long term distributions/means of Environment Canada station based gridded dataset (~120 stations in domain), created using Maurer et al. (2002) methods. |
The final step in determining initial conditions is the adjustment of the inital snow water equivalent state via an assimilation of current SNOTEL and ASP (Automatic Snow Pillow, from Environment Canada) snow water equivalent observations. This step helps correct mis-estimation of snowpack due to hydrologic model bias and forcing errors. DETAILS
ESP-based approach
The ESP method involves using recent meteorology to "spin-up" or initialize the snow and moisture states of a hydrologic model, then using in turn a set of meteorology forecasts taken from previous years (in this case from 1959-1998), beginning on the same day that the forecast is initialized. The result forms an unconditional ensemble of forecasts based on the assumption that the daily climate during the forecast period could be the same as it was in any of the previous years for the same calendar period. Two conditional forecast results are also created, formed by: a) restricting the climate forecasts to those previous years sharing the ENSO state (El Nino) projected for the current winter-spring forecast period; and b) restricting the climate forecasts to those previous years sharing the current winter-spring ENSO and PDO (positive) states. The three streamflow forecasts ensembles are compared to climatology, the monthly streamflow distributions of streamflows that occurred during the period on which the ESP is based (reflecting the initial conditions associated with each year, rather than the current state).
Unconditional forecast met data: 1960-1999, each 365 or 366-day met.
trace starts on day 1 of forecast
ENSO classification is based on the CPC monthly ONI index. PDO positive years are those
after 1976; PDO negative years are those before 1974.
NSIPP/NCEP-CFS based approach
Forecasts initialized: NCEP - 1 month prior to forecast start
month, e.g., Jan 1 used for Jan 25 forecast; NSIPP - 2 months prior,
e.g., Dec 1 used for Jan 25 forecast (due to lag in availability).
Forecast period: 9 months (NCEP) and 7 months (NSIPP)
following material is out of date
. The initial state is used as the starting point with 20 (NCEP) or 9 (NSIPP) meteorological forecasts derived from climate model forecast simulations of precipitation and temperature. Climate model forecasts are available at the end of the first week of the month, so beginning of month forecasts use the climate forecasts from the previous month. The climate model outputs are monthly, at T62 (~1.9 degree) spatial resolution for NCEP and 2 x 2.5 degrees for NSIPP, so a number of steps are required to translate them into a suitable spatial and temporal format for input to the daily 1/8 degree VIC simulations. These steps are as follows:- bias-correction of monthly climate model-scale outputs using a probability-mapping approach: expression of the outputs as percentiles with respect to the climate model climatology, then substituting the associated values from the observed record;
- simple interpolation of monthly forecast anomalies to 1/8 degree scale;
- application of 1/8 degree anomalies to 1/8 degree climatological means to create monthly 1/8 degree values; and
- disaggregation of monthly values to daily values by random resampling of observed 1/8 degree month-long daily precipitation and temperature patterns, with rescaling/shifting to preserve the forecast signal.
These methods are described in detail in Wood et al. (2002).
CPC outlook-based approach
Forecast period: 12 months
Probability of exceedence (POE) forecasts for average monthly temperature and total precipitation in each of 102 climate divisions within the US are translated into a 30-member ensemble of monthly T and P for each climate division by the following procedure. The statistical forecast parameters given in the POE forecast format are used to generate random distributions of 3-month-average P & T, which are then translated to anomalies (precip percent of normal and temperature difference from normal). Pending development of a disaggregation method or the release of disaggregated forecasts, these are treated as monthly anomalies. Second, using a statistical method called the Schaake Shuffle (Clark et al., 2004), the 30-member anomaly distributions are inter-associated (temporally, spatially, and cross-variable) to produce a 30-member ensemble of monthly T and P anomaly sequences. Finally, the monthly, climate-division scale sequences are temporally and spatially disaggregated using the methods described in Wood et al. (2002), and the resulting forcings are used to drive 30 hydrologic forecast simulations. The overall approach and results have not been described in any publication, yet.
References
Maurer, E.P., A.W. Wood, J.C. Adam, D.P. Lettenmaier, and B. Nijssen, 2002, A Long-Term Hydrologically-Based Data Set of Land Surface Fluxes and States for the Conterminous United States, J. Climate 15, 3237-3251.
Wood, A.W., Maurer, E.P., Kumar, A. and D.P. Lettenmaier, 2002. Long Range Experimental Hydrologic Forecasting for the Eastern U.S. J. Geophysical Research, VOL. 107, NO. D20, October.
Clark, M., S. Gangopadhyay, L. Hay, B. Rajagopalan and R. Wilby, 2004, The SCHAAKE Shuffle: A method for reconstructing space-time variability in forecasted precipitation and temperature fields, J. of Hydrometeorology, 5, 243-262, February.
Wood, A.W. and D.P. Lettenmaier, 2006, A testbed for new seasonal hydrologic forecasting approaches in the western U.S., Bull. Amer. Met. Soc. (in review).