Long-lead daily streamflow forecasting using Long Short-Term Memory model with different predictors

Study region: The Yangtze River basin.Study focus: Machine learning techniques have been widely applied to streamflow prediction at short lead times, and increasing numbers of predictors are introduced to improve predictive skills. However, the effect different predictors and their combinations at long lead times remains to be explored. In this study, the Long Short-Term Memory (LSTM) model is used to forecast the daily streamflow over the Yangtze River during flood seasons at lead times of 1-30 days, by using different combinations of streamflow, precipitation, soil moisture, and evapotranspiration as predictors. The effects of different predictors on daily streamflow and high/low flows, together with the difference between short and long lead times, are investigated.New hydrological insights for the region: The experiment using antecedent streamflow as predictor has a high (0.58-0.99) Kling-Gupta efficiency (KGE) at 1-7days lead times, but KGE decreases rapidly as the lead time increases. Adding precipitation predictor increases the KGE of daily streamflow by 0.09-0.25 and corrects negative/positive biases in high/low flows at long lead times (> 20 days). Soil moisture and evapotranspiration predictors have added values to daily streamflow forecasting at long lead times by further improving the KGE by 0.04-0.11, but exert minor influences on streamflow extremes. Thus, how to incorporate multiple hydrometeorolog-ical predictors in LSTM model to improve the forecasts of extremes at long leads is still a challenge.