Application of Partial Least-Squares Regression in Seasonal Streamflow Forecasting

The application of partial least-squares regression (PLSR) in seasonal streamflow forecasting was investigated using snow water equivalent, precipitation, temperature from automatic Snow Telemetry sites, and previous flow conditions as input variables. The forecast performance of PLSR models was compared to principal components regression (PCR) models as well as to the Natural Resources Conservation Service (NRCS) official forecasts in three Rio Grande watersheds including the Rio Grande Headwater Basin, Conejos River Basin in Colorado, and Rio Grande Basin above Elephant Butte Reservoir, New Mexico. The results indicated that using a correlation-weighted precipitation index is a relatively effective method in both improving forecast accuracy and developing relatively parsimonious regression models. In comparison of PLSR and PCR, similar forecast accuracies were obtained for both methods in jackknife cross validation and the test period (2003-2007) although PLSR has higher calibration coefficient of determination (R-2) and can reach its minimal prediction error with a smaller number of components than PCR. The comparison with NRCS official forecasts showed that the application of PLSR in seasonal streamflow forecasting is promising. This approach could be combined into NRCS's operational forecasting environment for possible forecast improvement.