United States Department of Agriculture
Natural Resources Conservation Service
National Water and Climate Center Go to Accessibility Information
Skip to Page Content
National Water and Climate Center

Water Supply Forecasting -- A Short Primer

Introduction

A water supply forecast is a prediction of streamflow volume that will flow past a point on a stream during a specified season, typically in the spring and summer. The NRCS, in cooperation with the National Weather Service, issues water supply forecasts for over 750 points in the western United States near the first of the month between January and June each year and at other times as requested.

The basis of water supply forecasting lies in the fact that most of the annual streamflow in western North America originates as snowfall that has accumulated in the mountains during the winter and early spring. This snowpack serves as a natural reservoir, storing water during the winter and releasing it during the spring and summer snowmelt season. The delay between when the snow falls and when it melts is what makes it possible for hydrologists to make predictions of snowmelt runoff.

In some areas, however, snowmelt is not as dominant, and therefore forecasting is more difficult. For example, on the west side of the Cascade Mountains, the east slope of the Rocky Mountains, and in parts of the Southwest, winter, spring, and/or summer rainfall can supply a significant amount of the streamflow volume. Since the source of this runoff occurs in the future at the time forecasts are issued, and it cannot easily be predicted, forecast uncertainty is higher in these areas than in dominantly snowmelt basins.

Models

Most water supply forecasts are made using statistical models. These are equations that express a mathematical relationship between the predictor variables (snowpack, precipitation, antecedent streamflow, etc.) and the seasonal streamflow volume of interest. Statistical models have the advantage that they are relatively simple and straightforward to calibrate and use, and they are usually acceptably accurate. The disadvantages are that they require long historical records (preferably greater than 20 years), and they do not represent all known physical processes that affect streamflow.

An alternative to statistical models is simulation models. These models attempt to represent, to a greater or lesser extent, all of the main physical processes affecting the movement of water within a watershed and the generation of streamflow. They operate on a continuous basis using a daily or shorter time step. The main advantage of simulation models is that, by explicitly accounting for physical processes, they have a more complete description of what is happening in the watershed and can potentially make more accurate streamflow predictions, especially under unusual circumstances. Other advantages of simulation models are that they can be run year-around and can produce other outputs besides seasonal streamflow volumes, such as full hydrographs and other hydrograph-based quantities. The disadvantages of simulation models are that they require significantly more input data than statistical models, are more difficult and time consuming to calibrate, require more complex output interpretation, and require more database and software infrastructure. Although the use of simulation models is limited at present, they nevertheless have much potential, and their use will increase in the future.

Uncertainty

Regardless of the model used, no forecast is perfect, so each forecast has uncertainty. The sources of forecast uncertainty are unknown future weather, model error (due to mathematical form and/or incomplete process descriptions), and data error (erroneous values or incomplete site coverage).

To express this uncertainty, the forecast is presented not as a single value but as a range of values, each with a specific probability of occurrence. The wider the spread among these values, the more uncertain the forecast. As the season progresses, forecasts generally become more accurate.

See “Interpreting Water Supply Forecasts” (http://www.wcc.nrcs.usda.gov/factpub/intrpret.html) for information on understanding forecast uncertainty and the range of forecast values for each point.

Data

Snowpack measurements from the NRCS’s SNOTEL system and manual snowcourses, as well as measurements from similar remote data networks in California, British Columbia, and Alberta, provide the key data source for making snowmelt runoff predictions. Measurements of other variables are used as well, such as precipitation, antecedent streamflow, groundwater, temperature, and large-scale climate indices. These other data are obtained from various federal and state agencies.

Statistical Model Development

The first step in developing statistical forecasting models is for the hydrologist to select a set of candidate predictor variables. Data sites are selected by examining a map and obtaining data from those sites in and near the basin of interest. The predictor variables considered for each forecast issue month include current snow water equivalent, precipitation for past months, streamflow for past months, large-scale climate teleconnection indices (e.g., Southern Oscillation Index) that have a proven relationship to the basin, and other variables that have predictive capabilities, such as fall groundwater levels or spring temperature.

Once the data are assembled, the hydrologist then does a preliminary screening, based on simple correlation analysis, to select the variables that have a potentially useful predictive relationship and to discard those that have a negligible relationship.

With these candidate variables, the statistical models are then developed. Generally, separate equations are developed for each forecast issuance date using only those variables available at that time. Month-to-month consistency in the forecasts is ensured by maintaining a high degree of similarity in data site and variable usage throughout the forecasting season.

Most statistical models used by the NRCS have been developed using the methods described by Garen (1992) (paper available at http://www.wcc.nrcs.usda.gov/publications/wsfpubs.html). These methods include principal components regression, jackknife (cross-validation) testing, and a variable combinations search algorithm.

Forecast Operations

Forecasting hydrologists increasingly rely on data reported by real-time or near real-time automatic reporting data networks. On the first working day of the month, enough data are generally available from these networks for the hydrologist to begin executing the forecast models and reviewing the forecasts. Some data are less timely than this, particularly precipitation data collected by cooperative weather observers and some streamflow and reservoir data. It may be necessary to wait another day or two to compute forecasts for basins that use these data.

The hydrologist reviews each forecast for reasonableness and for spatial and temporal consistency, making adjustments as necessary, in consultation with NRCS Water Supply Specialists. Graphical and GIS-based visualization tools assist in this process. Once the hydrologist is satisfied with the forecasts, they are shared with the National Weather Service River Forecast Centers for their review. Further adjustments are made as necessary during this “coordination” process to arrive at mutually agreed-upon forecasts.

In some selected basins and in response to special requests, mid-month forecasts are issued. These are usually generated from the following first-of-month forecasting models, for which the snow water equivalent and precipitation variables have been extrapolated from current observations to the first of the following month, based on selected assumptions for the intervening period.

Simulation models are also executed for selected basins. Forecasts are prepared using the Ensemble Streamflow Prediction technique in which multiple future streamflow scenarios are generated. Each scenario is based on the current watershed state and a different historically observed climate input series to represent the (future) forecast period. These streamflow scenarios can be summarized statistically to obtain forecasts and associated uncertainty of not only seasonal streamflow volumes but also other hydrologic quantities such as peak flow, date to recede below a threshold flow, etc. Simulation models are being used as an alternative procedure for seasonal streamflow volume forecasts as well as the basis for additional forecast products.

Conclusion

Water supply forecasts in the western United States have been produced since the relationship between winter snowpack and spring and summer streamflow was noted a century ago by the “father” of snow surveying and water supply forecasting, Dr. James Church, who was a professor at the University of Nevada. The program has evolved over the decades, developing a widespread network of snow measurement sites and more and more sophisticated measurement equipment, data analysis techniques, and forecasting models. This evolution and program development continues, exploiting new capabilities made possible by advancements in computing power, the Internet, GIS, and hydrologic science.