Medium-long-term prediction of water level based on an improved spatio-temporal attention mechanism for long short-term memory networks

Abstract

River water level usually given by nonlinear and nonstationary time series and affected by numerous complex spatial and temporal factors. But not all input factors are positively correlated with flood forecasting, and uncorrelated inputs tend to bring a lot of noise, resulting in lower prediction accuracy. Attention mechanism is an effective solution to feature filter. This study presents a spatio-temporal attention LSTM (STA-LSTM). Meanwhile, attention visualization makes models interpretation providing the basis. In this work, the attention-weight matrix is generated independently by the spatial attention module and the temporal attention module to focus on the more valuable feature factors in time and space, respectively. Spatio-temporal attention is a combination of spatial attention and temporal attention to assign scores to input sequences. Forecasting the water level of Hanchuan in the next 6 h based on the observations in the past 6 h in the middle and lower reaches of the Hanjiang River, China. The experimental results indicate that the mean absolute error (MAE) of STA-LSTM is 0.74835, root mean square error (RMSE) is 0.90197, mean absolute percent error (MAPE) is 3.29924, and coefficient of determination (R2) is 0.94138. This study improves the accuracy of the water level prediction method, which makes it easier for decision-makers to plan evacuations in advance.