A soil moisture index as an auxiliary ANN input for stream flow forecasting

This study tests the short-term forecasting improvement afforded by the inclusion of low-frequency inputs to artificial neural network (ANN) rainfall-runoff models that are first optimized by using only fast response components, i.e. using stream flow and rainfall as inputs. Ten low-frequency ANN input candidates are considered: the potential evapotranspiration, the antecedent precipitation index (APIi, i=7, 15, 30, 60, and 120 days) and a proposed soil moisture index time series (SMIA, for A=100, 200, 400 and 800 mm). As the ANNs considered are for use in real-time lead-time-L forecasting, forecast performance is expressed in terms of the persistence index, rather than the conventional Nash–Sutcliffe index. The APIi are the non-decayed moving average precipitation series, while the SMIA are calculated through the soil moisture accounting reservoir of the lumped conceptual rainfall-runoff model GR4J. Results, based on daily data of the Serein and Leaf rivers, reveal that only the SMIA time series are useful for one-day-ahead stream flow forecasting, with both the potential evapotranspiration and the APIitime series failing to improve the ANN performance.