Multi-Year Forecast of Colorado River Flows Using a Bayesian Dynamic Linear Model

Abstract:

Multi-year forecasts of river flow are important for efficient water resources management, particularly in the Colorado River Basin (CRB) given its numerous reservoirs and multi-year storage capacity. On the CRB, forecasts of the current year’s spring season (year 1) and next year’s spring season (year 2) flow are used in reservoir management. Forecasts for year 1 are generally skillful and both statistical and physical modeling systems provide skillful ensemble streamflow predictions (ESP) using seasonal climate and snow forecasts that are used by reservoir managers. However, efficient management requires skillful forecasts of year 2, which are currently lacking and thus, climatology is used instead, leading to sub-optimal reservoir management decisions. Skillful year 2 forecasts continue to be a challenging problem on the CRB. We employ a Bayesian Dynamic Linear Model (BDLM) with large-scale climate and land covariates and demonstrate the potential for skillful forecasts in both years 1 and 2. Motivated by prior studies, we selected indices of El Niño Southern Oscillation (ENSO), Pacific Decadal Oscillation (PDO), and Atlantic Multidecadal Oscillation (AMO) as predictors. In addition, we used the prior year’s flow and runoff efficiency (RE) to capture the hydrologic characteristics of the River Basin. A Bayesian Dynamic Linear Model offers an attractive alternative to a standard linear regression model, for it enables use of time varying coefficients that capture the non-stationarity of the teleconnections, in addition to providing robust estimation of uncertainties in the forecasts via posterior distributions. During the fitting period of 1907-1979 the model captures the temporal variability very well, yielding a correlation between the historic flows and posterior mean from the BDLM of 0.99 for both year 1 and year 2 forecasts. In the hindcasts for the period of 1980-2016, the BDLM ensembles exhibit higher skill relative to climatology in ~50% of hindcast years, with maximum rank probability skill scores (RPSS) of 0.95 and 0.91 for the year 1 and year 2 forecasts, respectively. The BDLM was also able to capture the multi-year variability moderately well. Our results indicate the potential use of large-scale teleconnections for an improved year 1 and 2 forecast of CRB flows and efficient water management.