The importance of short lag-time in the runoff forecasting model based on long short-term memory

Abstract

It is still very challenging to enhance the accuracy and stability of daily runoff forecasts, especially several days ahead, owing to the non-linearity of the forecasted processes. Here, we hypothesize that short lag-time has a significant impact on forecasting results. Thus, we incorporate short previous time steps into long short-term memory (LSTM) and develop the Self-Attentive Long Short-Term Memory (SA-LSTM). In SA-LSTM, the self-attention mechanism is used to model interdependencies within short previous time steps. SA-LSTM is evaluated at eight runoff datasets. The experimental results demonstrate that, compared with state-of-art benchmark models, SA-LSTM achieves the best performance. The RMSEs of SA-LSTM are at least 2.3% smaller than that of the second best model at the seventh day. The NSEs and NSE_In of SA-LSTM are at least 4.6% and 6.4% higher than those of the second best model at the seventh day. Furthermore, SA-LSTM also surpasses the baseline methods for base, mean and peak flows. The superiority of SA-LSTM can be attributed to its exploitation of information in short lag-time.