Comparison And Analysis Of Runoff Simulation Between SWAT Model And LSTM Model

In the context of frequent occurrence of global hydrological events,runoff simulation is of great significance for the rational formulation of watershed water resources allocation plans.This paper uses the SWAT model(Soil and Water Assessment Tool)and LSTM(Long Short Term Memory)model to realize the monthly runoff and daily runoff simulations of the headwater catchment of the Yellow River basin from 2008 to 2017.The SWAT model uses DEM,soil data,land use and meteorological data,a hydrological model was constructed using ArcSWAT,and the model parameter sensitivity analysis and parameter calibration and verification were completed,and the hydrological simulation process from rainfall to runoff output was realized.The LSTM model has completed the selection of hyperparameters and model training based on the historical runoff data,based on the Keras deep learning framework and Python,and realized the prediction of runoff.The conclusions of this paper are as follows:(1)Both the SWAT model and the LSTM model can achieve a good simulation of runoff.Monthly runoff simulation and daily runoff simulation,R2 is between 0.53 and 0.98,and NSE is between 0.4 and 0.98.(2)The SWAT model has advantages over the LSTM model in monthly runoff simulation.Among them,the simulation accuracy R2 and NSE from 2014 to 2017 are 0.72 and 0.61,respectively,which are improved by 0.04 and 0.01 compared to the LSTM model.The average R2 and average NSE under different prediction lengths using the cross-validation method are 0.79 and 0.64,which are improved by 0.11 and 0.09 compared with the LSTM model.Analysis of the reason why the SWAT model's monthly runoff simulation accuracy is higher.On the one hand,due to the stronger correlation between the SWAT model runoff simulation results and precipitation,using correlation analysis,the Pearson correlation coefficient between SWAT runoff simulation results and precipitation is 0.82 The 0.79 of the LSTM model makes its simulation results more accurate than the LSTM model in the year of abundant water,making the overall simulation accuracy higher.On the other hand,because the LSTM model has fewer monthly runoff simulation training samples,the LSTM model has a lower monthly runoff simulation accuracy than the SWAT model.Therefore,using the monthly runoff data from 2002 to 2013 as the training set,the training samples are expanded to twice the original,and compared with the original training set.The results show that the simulation accuracy R2 and NSE of the expanded training set are 0.73 and 0.71,respectively.Compared with the original training set,it was improved by 0.05 and 0.11 respectively.(3)The LSTM model has advantages over the SWAT model in daily runoff simulation.Among them,the simulation accuracy R2 and NSE from 2014 to 2017 are 0.98 and 0.98,which are 0.22 and 0.33 higher than the SWAT model,respectively.Using cross-validation,the average R2 and average NSE at different prediction lengths were 0.98 and 0.98,respectively,which were improved by 0.18 and 0.27 compared to the SWAT model.The reason may be that the LSTM model has sufficient daily runoff simulation training samples.The results show that the LSTM model is more sensitive to the training sample size.When the training samples are sufficient,the LSTM model runoff simulation accuracy will be greatly improved.