Spatio-Temporal Attention LSTM Model for Flood Forecasting

In order to reduce the loss caused by flood, a large number of researches based on data, algorithms, machine learning and other technical means are used to realize flood forecasting. It will be a kind of flexible research method to realize the flood prediction of small and medium-sized rivers through intelligent models such as neural network. The area of small and medium-sized river basins is relatively small. Precipitation, soil moisture, evaporation and other factors can affect the timely total runoff prediction. However, not all the hydrological features is always valuable for flood forecasting, even at some time, noise of the factors will have larger interference on forecast process. Therefore, dynamic extraction of key feature vectors from various hydrological information plays an important role in flood forecasting. This paper proposed a flood forecasting model (STA-LSTM model) by using long short-term memory model (LSTM) and attention mechanism. We take the Lech river basin in Europe as the experimental basin and the results show that STA-LSTM performs well and has high research value with comparison of support vector machine (SVM), fully connected network (FCN) and original LSTM