The Importance Of Short Lag-time Based On Long Short-term Memory And Runoff Forecasting Model Research

Rivers are an important part of surface water.Forecasting runoff accurately is always an important research content in hydrology.However,the high-accuracy runoff forecasting is still quite difficult because of the non-linear characteristics of runoff.Therefore,for daily runoff forecasting,we add short lag-time input to the model,and combine the self-attention mechanism and long short-term memory(LSTM)to construct a self-attentive long short-term memory(SA-LSTM)model and a LSTM based rainfall runoff model with attentive long and short lag-time(LSTM-ALSL)in this paper.Selected eight stations in the Mississippi River Basin and four stations in the Model Parameter Estimation Experiment(MOPEX)as the research objects,the daily runoff forecasting with runoff input and the daily runoff forecasting with rainfall runoff input are carried out.Through comparing and analyzing the overall forecasting accuracy,peak flow and base flow forecasting accuracy of different models,the conclusions are obtained as follows:(1)In daily runoff forecasting with runoff input,the SA-LSTM model improves the accuracy of daily runoff forecasting effectively by adding a short lag-time input to the LSTM and combining the self-attention mechanism to obtain information of short lag-time.At all eight stations in the Mississippi River Basin,the accuracy of SA-LSTM model is the highest among all models.The Nash–Sutcliffe efficiency coefficient(NSEs)of SA-LSTM model are averagely 14.2% higher than that of LSTM model.The root mean square error(RMSEs)and mean square error(MAEs)of SA-LSTM model are12.8% and 14.7% lower than that of LSTM model on average.The SA-LSTM model also significantly improves the accuracy for peak flow and base flow forecasting.It shows that the short lag-time input effectively improves the accuracy of forecasting.(2)In daily runoff forecasting with rainfall runoff input,the LSTM-ALSL model adds a short lag-time input to the LSTM model.It applies the self-attention mechanism to both the short lag-time input and the long lag-time input,and combines the original long lag-time input for forecasting.This method improves the accuracy of daily runoff forecasting effectively.At the four stations in MOPEX dataset,the accuracy of the LSTM-ALSL model is the highest among all models.The NSEs of LSTM-ALSL model are averagely 5.7% higher than that of LSTM model.The RMSEs and MAEs of LSTM-ALSL model are at least 8.1% and 10.6% lower than the LSTM model,respectively.At the same time,LSTM-ALSL model also significantly improves accuracy for peak flow and base flow forecasting.It shows that the short lag-time input effectively improves the accuracy of forecasting.