Differing roles of base and fast flow in ensemble seasonal streamflow forecasting: An experimental investigation

Abstract

Seasonal streamflow forecasting, which is important for water resources management, is generally a challenging task in hydrology. This paper focuses on the consideration of base and fast flow for the forecasting of seasonal streamflow. Taking the three tributaries of the Pearl River in South China as case study, five forecasting experiments, which forecast the average daily flow in the three-month period by using the average daily flow in the antecedent month, are devised to elucidate the different roles of base, fast, and total flow. The recursive digital filter (RDF) model is employed for the separation of base flow from fast flow; the Bayesian joint probability (BJP) model is applied to generate ensemble forecasts under cross validation; and the Schaake shuffle is used to establish temporal structure to link ensemble members when adding up forecasts of base flow and fast flow. Forecasts in the different experiments are evaluated in terms of bias, reliability, and skill, and the relationship between correlation coefficient and forecast skill is illustrated. The results show that unbiased and reliable seasonal forecasts are generally obtained in the experiments. Correlation coefficients of base flow are overall higher than those of total flow while correlation coefficients of fast flow are lower. Forecast skill tends to increase with correlation coefficient. Meanwhile, fast flow can lead to extremely large total flow in the antecedent month, which can substantially enlarge the forecasts and diminish the forecast skill when the corresponding observation is not as large. By contrast, base flow tends to always contribute to forecast skill. The separation of base flow from fast flow, which accounts for the different dynamical processes of fast and base flow, can reduce the impact of fast flow and improve the forecast skill of total flow.