Long Lead-Time Forecasting of U.S. Streamflow Using Partial Least Squares Regression

Glenn A. Tootle, University of Tennessee - Knoxville
Ashok Singh, University of Nevada, Las Vegas
Thomas C. Piechota, University of Nevada, Las Vegas
Irene Farnham, S.M. Stoller Corporation

Abstract

Pacific and Atlantic Ocean sea surface temperatures (SSTs) were used as predictors in a long lead-time streamflow forecast model in which the partial least squares regression (PLSR) technique was used with over 600 unimpaired streamflow stations in the continental United States. Initially, PLSR calibration (or test) models were developed for each station, using the previous spring-summer Pacific (or Atlantic) Ocean SSTs as predictors. Regions were identified in the Pacific Northwest, Upper Colorado River Basin, Midwest, and Atlantic states in which Pacific Ocean SSTs resulted in skillful forecasts. Atlantic Ocean SSTs resulted in significant regions being identified in the Pacific Northwest, Midwest, and Atlantic states. Next, streamflow stations were selected in the Columbia River Basin, Upper Colorado River Basin, and Mississippi River Basin and a PLSR cross-validation model (i.e., forecast) was developed. The results of the PLSR cross-validation model for each station varied with linear error in probability space scores of +9.5 to +51.0% where 10% is considered skillful forecasts using Pacific and Atlantic SSTs.