The Simulation Of Hydrological Process Based On Machine Learning In Typical Cryospheric Watersheds

As a black-box model,machine learning models have no clear hydrological mechanism compared to physical based hydrological models,and can be simulated without the need for watershed related physical parameters,greatly reducing the parameters of the model.It has strong learning ability,association ability,and can approximate arbitrary linear relationships.It can more accurately and quickly simulate watershed hydrological processes,and obtain relatively accurate results.It is suitable for Cryospheric Watersheds without data.Therefore,machine learning models are widely used in hydrological research,but the simulation of hydrological process based on machine learning in the cryospheric watersheds which are located at high-altitude areas are seldom mentioned.Based on meteorological and hydrological data,CN05 dataset,and CMIP6 data,this paper conducts simulation studies on the hydrological processes of two typical cryosphere basins,namely,the Yarkant River basin and the Shule River basin.Using trend analysis,M-K catastrophe analysis,and HILL estimation methods,it reveals the climatic indicators that affect the hydrological process,and uses principal component regression analysis to explore the relationship between extreme climate and runoff;Hydrological process simulation based on six machine learning methods is combined with precision evaluation indicators(NSE,RMSE,R)and hydrological process frequency curve to comprehensively analyze the simulation effect of the model;Based on the Taylor diagram and interannual variability skill score,the climate change simulation capabilities of the two watersheds were evaluated,and the future climate change patterns in the study area under the SSP1-2.6,SSP2-4.5,and SSP5-8.5 scenarios were estimated,quantifying the uncertainty of climate prediction.The main conclusions are as follows:(1)The temperature,precipitation,and runoff in the Yarkant River basin show a significant upward trend;The temperature and runoff in the Shule River basin showed a significant upward trend,while the precipitation showed a stable upward trend.The abrupt changes in temperature and runoff in the Yarkant River basin occurred in 1998,while the abrupt changes in temperature and runoff in the Shule River basin occurred in 1993 and 1998,respectively.There was no abrupt change in precipitation in the two basins.The runoff thresholds for the Yarkant River basin and the Shule River basin are 148.3 mm and 112.8 mm,respectively.Extreme precipitation in the Yarkant River basin has a significant impact on runoff changes,followed by extreme temperature.The extreme temperature in the Shule River basin has a significant impact on runoff changes,followed by extreme precipitation.(2)The performance of double-layer LSTM in Yarkant River Basin is much better than that of other models,and the performance of double-layer LSTM in Shule River Basin is similar to that of other models.Double-layer LSTM is more suitable for hydrological process simulation in cryosphere basin.The loss function is used to evaluate the parameterization scheme of model.It is found that the simulation effect of LSTM model in the study area is mainly affected by the optimizer,followed by learning attenuation rate and initial learning rate in Yarkant River Basin.The importance of initial learning rate is second in Shule River Basin.Combined with the abrupt change of runoff and input data of model,the results suggest that climatic factors show different impacts on hydrological processes during two periods in two basins.However,the total amount of precipitation and extreme precipitation shows greater impacts on the hydrological process in the study area than temperature during whole study period.(3)Compared with historical observation data from 1961-2014,the future climate of the two watersheds showed an upward trend under the SSP1-2.6,SSP2-4.5,and SSP5-8.5scenarios.Before the middle of this century,the change trend of the three scenarios was not significantly different.Over time,the trend of climate change under the high forcing scenario was much greater than that under the medium and low forcing scenario.Among the 22 climate scenarios,four models,namely,INM-CM4-8,IITM-ESM,MRI-ESM2-0,and NESM3,are better at simulating the climate of the two watersheds.The uncertainty of temperature prediction in the Yarkant River basin is the main factor in the short-term model uncertainty,and over time,it is taken over by scenario uncertainty.The uncertainty of temperature and precipitation in the Shule River basin mainly stems from model uncertainty.