Enhancing the Predictability of Seasonal Streamflow With a Statistical-Dynamical Approach

Seasonal streamflow forecasts facilitate water allocation, reservoir operation, flood risk management, and crop forecasting. They are generally computed by forcing hydrological models with outputs from general circulation models (GCMs) or using large-scale climate indices as predictors in statistical models. In contrast, hybrid statistical-dynamical forecasts (combining statistical methods with dynamical climate predictions) are still uncommon, and their skill is largely unknown. Here we conduct systematic forecasting of seasonal streamflow using eight GCMs from the North American Multi-Model Ensemble, 0.5-9.5 months ahead, at 290 stream gauges in the U.S. Midwest. Probabilistic forecasts are developed for low to high streamflow using predictors that reflect climatic and anthropogenic influences. Results indicate that GCM forecasts of climate and antecedent climatic conditions enhance seasonal streamflow predictability; while land cover and population density predictors decrease biases or enhance skill in certain catchments. This paper paves the way for novel forecasting approaches using dynamical GCM predictions within statistical frameworks. Plain Language Summary Streamflow forecasts several months ahead of a season are important for water management and the prevention of risks related to floods and hydrological droughts. However, existing methods for producing seasonal streamflow forecasts are often complex and computationally intensive. Here we provide a systematic evaluation of a statistical-dynamical approach to streamflow forecasting in several hundred river catchments across the U.S. Midwest. We assess whether global climate model forecasts can be used as predictors in statistical models to produce skillful forecasts of river flow, up to 10 months ahead. Results indicate that forecasts of rainfall and temperature, antecedent climatic conditions, as well as information on population density and land cover, can be used effectively to forecast streamflow at seasonal time scales. By including information on the future antecedent climatic conditions, streamflow forecasts can be enhanced months ahead. Information on human influences, in contrast, helps reduce the biases in the streamflow forecasts. These results pave the way for statistical-dynamical forecasting in catchments around the world and suggest that process-driven combinations of different predictors can be used to produce skillful streamflow forecasts in different seasons, for both high flows (i.e., floods) and low flows (i.e., representative of hydrological droughts).