LAUSR

Recent advances in stream-flow prediction using Recurrent Neural Networks (RNN) or Long Short-Term Memory (LSTM) have prompted their exploration in the field of hydrology. Predicting hydrological time series is of great importance, but is very challenging as it is influenced by many complex factors such as dynamic spatio-temporal feature correlation and nonlinear meteorological-hydrological. In this context, we propose a framework based on dynamic spatio-temporal attention (DSTA) to predict the stream-flow of a hydrological station over several days. It consists of three main modules: a spatial module, a temporal module, and a trend input module. Experiments were carried out on real data sets from four watersheds in the mainstream of the Yangtze River and proved that our method is superior to the six baseline methods.