A stream prediction model based on attention-LSTM

The small- and medium-sized watersheds have complex and varied hydrogeological features, boundary conditions, and human activities. There are nonlinear interactions between these factors, which leads to great challenges in predicting the stream of the river. Since not all factors are positively correlated with flood forecasting, and irrelevant factors tend to bring a lot of noise, it is necessary to give more attention to the absolute action factors. In this paper, we forecast the flow values over the next 12 hours, using an Attention-LSTM prediction model with an attention mechanism based on long-term and short-term memory networks that consider past stream data, past weather data, and weather forecasts data. We use data from Tunxi watershed, China, and evaluate the model with root mean square error (RMSE), Nash-Sutcliffe efficiency coefficient (NSE), and coefficient of determination (R²). The forecast results of the Attention-LSTM model are compared with the prediction results of two traditional machine learning models and an LSTM model. The experimental results show that the Attention-LSTM model has a higher score, and provided a new method for flood forecasting.