Daily Streamflow Forecasts Based on Cascade Long Short-Term Memory (LSTM) Model over the Yangtze River Basin

Abstract

Medium-range streamflow forecasts largely depend on the accuracy of meteorological forecasts. Due to large errors in precipitation forecasts, most streamflow forecasts based on deep learning rely only on historical data. Here, we apply a cascade Long Short-Term Memory (LSTM) model to forecast daily streamflow over 49 watersheds in the Yangtze River basin for up to 15 days. The first layer of the cascade LSTM model uses atmospheric circulation factors to predict future precipitation, and the second layer uses forecast precipitation to predict streamflow. The results show that the default LSTM model provides skillful streamflow forecasts over most watersheds. At the lead times of 1, 7, and 15 days, the streamflow Kling–Gupta efficiency (KGE) of 78%, 30%, and 20% watersheds are greater than 0.5, respectively. Its performance improves with the increase in drainage area. After implementing the cascade LSTM model, 61–88% of the watersheds show increased KGE at different leads, and the increase is more obvious at longer leads. Using cascade LSTM with perfect future precipitation shows further improvement, especially over small watersheds. In general, cascade LSTM modeling is a good attempt for streamflow forecasts over the Yangtze River, and it has a potential to connect with dynamical meteorological forecasts.