Statistical Ensemble Seasonal Streamflow Forecasting in the South Saskatchewan River Basin by a Modified Nearest Neighbors Resampling

Abstract

An ensemble seasonal streamflow forecasting model is developed for two watersheds in the South Saskatchewan River Basin of southern Alberta, Canada. The ensembles are generated from a mean forecast by using a modified K -nearest neighbor algorithm. The mean forecasts are produced by a robust M-regression model that uses snow water equivalent and large-scale climate information as predictors where the best combination of predictors is automatically selected by the generalized cross-validation criterion. It is shown that skillful forecasts of the April–September flow can be obtained as early as the beginning of December preceding the runoff year, thus extending the current forecast lead time by up to two months. An assessment of the potential economic value of the forecasts shows that with the same set of predictors, ensemble forecasts offer superior economic value for a wide range of end-users as compared to conditional median forecasts.