A Multi-Model Hydrologic Ensemble for Seasonal Streamflow Forecasting in the Western U.S. Theodore J. Bohn, Andrew W. Wood, Ali Akanda, and Dennis P. Lettenmaier Department of Civil and Environmental Engineering, Box , University of Washington, Seattle, WA American Geophysical Union Fall Meeting, December 2005 ABSTRACT Since 2003, the Variable Infiltration Capacity (VIC) macroscale hydrology model has been applied in real time over the western U.S. for experimental ensemble hydrologic prediction at lead times of six months to a year. VIC hydrologic initial conditions are produced from gridded station observations during a two-year runup period prior to the forecast date; and hydrologic forecast ensembles are driven by climate forecasts from several sources, including NCEP and NASA climate model outputs, CPC official seasonal outlooks and, as a baseline forecast, Extended Streamflow Prediction (ESP). We are now in the process of expanding this approach to include forecasts made from a Bayesian combination of the results from a suite of land surface models. Our initial set of LSMs includes VIC, the NWS grid-based Sacramento model (HL-RMS) and the NCEP NOAH model. All three LSMs are implemented on the 1/8 degree grid used by the North American Land Data Assimilation System (N-LDAS). Here we present preliminary results from several river basins in the Western US, focusing on both retrospective deterministic simulations and retrospective ESP-based ensemble forecasts and forecast error properties. We compare linear regression and Bayesian methods of combining model results, and investigate seasonal and geographic variations in forecast skill. Our data set includes 20+ years of 1-year, ESP-based, 25-member ensemble forecasts for each model, using both April 1 and October 1 as starting dates, from the Salmon River, ID, the Feather River, CA, and the San Juan River, UT. 5 Results: Split-Sample CONCLUDING REMARKS While this study is still in its preliminary stages, evidence so far suggests that: An ensemble of land surface models can make accurate estimates of hydrological variables more consistently and under a wider range of conditions than individual models. Ensembles of land surface models may be able to help us more accurately estimate hydrological variables in regions where there are few observations. Future plans: We are in the process of testing the multi-model ensemble’s performance in the context of Extended Streamflow Prediction (ESP) ensembles, in which an ensemble of historical meteorological inputs is used to drive a model. We expect the multi-model ensemble to be more stable in the face of difference meteorological forcings, increasing the relative influence of initial conditions on the forecast. We are in the process of comparing ensemble performance as a function of start date. Note: See the author for a list of references. 1 Models The LSMs in our ensemble all share the same basic structure, consisting of grid cells containing a multi- layer soil column overlain by one or more “tiles” of different land covers, including vegetation with and without canopy and bare soil. Water and/or energy fluxes are tracked vertically throughout the column from the atmosphere through the land cover to the bottom soil layer. The figure to the right illustrates these features as implemented in the VIC (Variable Infiltration Capacity) macroscale land surface model (Liang et al., 1994). Basins used in this Study 3 Method Model Inputs and Outputs Land Surface Parameters VIC LDAS parameters (Maurer et al, 2002) (note: the version used here did not match the version of the LDAS forcings; hence VIC’s output does not match the observations as closely as it should. NOAH NLDAS (Mitchell et al, 2004) SAC NLDAS (Mitchell et al, 2004) Forcing degree LDAS forcings (Maurer et al, 2002) Disaggregated to 3-hourly time step NOTE: no further calibration of the models was performed. 2 4 Salmon River, Feather River, Colorado River, Monthly Model Weights Timeseries, Monthly Std DevMonthly MeansMonthly RMSE Monthly Means Monthly Std Dev Monthly RMSE For a given hydrological variable Y, the Bayesian Model Averaging method (Raftery et al., 2005) yields not only an optimal estimate X ens but also a probability distribution p(Y|X ens ) about that estimate, which allows us to estimate the uncertainty of the ensemble forecast. This approach requires the data set to be unbiased and normally-distributed. However, streams in the Western U.S. tend to have 3-parameter log-normal (LN3) distributions, often having a different distribution during each season or month. Therefore, in order to form our ensemble forecasts of streamflow, we must transform the data into the Gaussian domain, following the procedure below: Multi-Model Ensemble Process Flow The final product is an ensemble distribution for each point in time. We chose the 10 th and 90 th percentiles of this distribution, denoted as Q10 and Q90 above, to represent the lower and upper bounds of uncertainty in our plots. Salmon River Feather River Colorado River Models at a glance VIC Physically-based horizontal soil layers Energy balance Elevation bands NOAH Physically-based horizontal soil layers Energy balance No elevation bands Sacramento/SNOW17 Conceptually-based soil storages No energy balance Elevation bands Degree-day snow melt scheme No explicit vegetation Potential Evapotranspiration must from NOAH Our test consisted of two parts: 1. training the ensemble on even-numbered years between 1956 and 1994 (1954 and 1992 in the case of the Feather River) and 2. validating the ensemble on odd-numbered years. The plots above show that no single model performs consistently better than the others; performance varies with season. For the Salmon River, VIC tends to correctly capture summer/autumn baseflow and the timing of the spring snowmelt peak, but often underestimates the magnitude of the peak. Meanwhile, SAC and NOAH tend to miss baseflow entirely and have earlier snowmelt peaks than the observations. NOAH in particular substantially underestimates snowmelt, which may be due to overestimation of evaporation. From the monthly RMSE, it is clear that the ensemble estimate of stream flow is generally at least as good as the best individual (bias-corrected) model at any given time. The ensemble forecast accuracy tends to decrease as flow rate increases.