92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Tuesday, 24 January 2012: 2:45 PM
Improving Short Term Streamflow Forecasts by Combining a Statistical and Distributed Hydrological Model
Room 350/351 (New Orleans Convention Center )
Kenneth T. Waight, AWS Truepower, LLC, Troy, NY; and G. E. Van Knowe

Real-time operational hydrological forecasts are needed for hydroelectric power generation, water resource planning and warnings of possible water-related disasters. There are several modeling approaches to real-time hydrological forecasts, each having strengths and weaknesses. MESO has developed a multi-model streamflow forecasting system for flood and hydroelectric forecasting in the southwestern Washington Lewis River Basin. The modeling system consists of a distributed hydrological model and a statistical model used in combination to produce hourly streamflow forecasts (Figure 1) for several locations in the basin.

Fig. 1. A schematic of the hourly hydrological forecasting system.

The distributed model is the Distributed Hydrology Soil Vegetation Model (DHSVM). It is a physically-based, spatially distributed hydrology model that was developed at the University of Washington. It has been widely applied and supported both (1) operationally for streamflow prediction for hydropower and (2) in a research capacity to examine the effects of changes in vegetation, forest management, etc. on streamflow. DHSVM explicitly represents the effects of diverse topography and heterogeneous subsurface conditions on the downslope redistribution of subsurface moisture that provides a dynamic representation of the spatial distribution of soil moisture, snow cover, evapotranspiration, and runoff. Multi-sensor precipitation estimates (integrating radar-derived precipitation and rain gauge data) are used to provide input into the distributed model before the beginning of the forecast period. Also, a mesoscale atmospheric model, the Mesoscale Atmospheric Simulation System (MASS), provides bias-corrected precipitation forecasts which drive the distributed model during the forecast period.

The statistical method used in the 0 - 7 day component of MESO's forecasting system is the K-Nearest Neighbor (K-NN) approach based upon nonparametric resampling. K-NN resampling is a statistical approach to downscaling that constructs an analog of the current pattern to be used in the forecasts from a linear combination of past patterns. This methodology has shown good forecasting skill, especially for the longer term forecasts. In the K-NN method, observed precipitation and streamflow data are used to provide initial conditions. Precipitation from two sources is used to determine the precipitation for the future conditions. One source of precipitation data is from the MASS; the other source is from NOAA's "reforecasted" dataset. The model used to create NOAA's "reforecasted" dataset utilizes data, quality control methods and modeling techniques on historical data sets that are not available in real time.

In order to assess the value of the combined distributed and statistical model as compared to a stand-alone distributed or statistical model, the results of hourly forecasts made using the three approaches (distributed, statistical and combined) for the 2011 water year will be presented.

Supplementary URL: