A teleconnection-based forecasting model for natural streamflow in Idaho, USA
A Principal Component Analysis of the selected climatic indexes was used to identify the monthly indexes from one year that described the most variation in total QU in the following year. Based on this analysis, we constructed multivariate linear regression models to forecast annual QU at the start of the water year, using a stepwise model building technique to maximize the adjusted r2 between the forecast and observed QUs with Mallow's CP as a stopping rule in order to reduce collinearity among the predictor variables. Historical data from 1977 through 2004 were used as a training dataset to construct the model, with data from 1950-1976 and 2005-2008 withheld from model development for use in model verification. Using a bootstrapping procedure in which four years were randomly removed from the training data, we performed the stepwise regression model building procedure and computed model goodness-of-fit statistics for the removed years. This approach successfully predicted the annual QU to a maximum ±6.5% error throughout the 1977-2004 training record and a maximum ±10.5%, ±5.8%, ±5.0% error throughout the 1950-1960, 1961-1976, 2005-2008 verification records respectively. The 1950-1960 verification period occurred during construction of Lucky Peak Reservoir, which may account in part for weaker model results. Despite a relatively large number of predictor variables in the constructed models, the model development techniques employed ensure independence among the predictor variables. Ongoing work seeks to explore the physical linkages underlying for the observed statistical correlations and the applicability of the employed methods in other geographic settings within and outside of the interior Pacific Northwest.