Study Of Runoff Simulation And Transfer Learning Model Based On Deep Learning

The runoff during the hydrological process is affected by the interaction of complex nonlinear physical mechanisms,involving multiple disciplines such as hydrology,meteorology,and ecology.In the context of varying global climate,accelerating Earth’s water cycle,and changing regional runoff patterns,it has been one of the difficult and hot topics in the field of hydrology to efficiently use the limited data to accurately simulate the flow in the hydrological process.In recent years,artificial intelligence,especially deep learning,has developed rapidly and shown great potential in other fields such as medicine and remote sensing.However,in the field of hydrology,the application of deep learning is facing challenges in terms of limited dataset and imbalanced gauge distribution.To address these issues,this study focuses on deep learning and uses multiple methods such as distributed hydrological models,deep learning models,and various statistical methods to simulate and analyze hydrological streamflow at different basin scales in China,the United States,the United Kingdom,and Chile.Meanwhile,aiming to enhance the model’s ability in streamflow simulation,this study improves and innovates the deep learning model,and further develops a deep learning framework applicable for hydrological research.The main results of this thesis are as follows:Based on deep learning and CAMELS(the Catchment Attributes and MEteorology for Large-sample Studies)series datasets,this study establishes the LSTM(Long short-term memory)models for basins in the U.S.,U.K.,and Chile respectively;while in the upper Min River basin in China,the simulation results from SWAT(Soil and Water Assessment Tool)model are used to analyze attribution of the runoff changes and to establish the LSTM model for Min River basin.The regional meteorological and hydrological conditions in the four countries/regions show distinct differences and the streamflow simulations in different regions at diverse scales shows:LSTM model performs differently in the simulation of extreme flow events,the model generally underestimates the high flow and overestimates the low flow in most basins,while the LSTM overall has a good performance with the averaged ensemble NSE(Nash-Sutcliffe efficiency coefficient)of 0.742.The LSTM with SWAT-extended data of Min River basin in China also shows good performance with NSE of 0.666.As compared to basins with arid climate and glacier cover,the LSTM model is more accurate in simulating wet climate abundant rainfall basins.Based on the LSTM model and transfer learning strategy,this study optimizes the model structure and develops the Transfer Learning model(TL)and transfer framework.Using CAMELS dataset as the "source data",the transfer framework applied various basins in China,the U.K.and Chile show that,compared with the LSTM model established in each region,the TL improves the simulation result with the ensemble NSE overall increased by 5.9%(ensemble NSE increased by 10.2%、4.3% and 3.5% in China,Chile and the U.K.,respectively).On the whole,the TL effectively improved the model performance in regional streamflow simulation without increasing the amount of input target area,which verifies the excellent applicability of the TL in different regional and makes the technical aspects ready for the efficient use of global data.Based on the LSTM model and TL for basins in China,the U.K.,and Chile,this study compared their application under the condition of a small amount of target data,and the results show that the TL and its framework have significant advantages.In the "one-year training model" where the input data are relatively scarce,the TL shows its advantages from the "pre-training model" and network structure,and compared with the corresponding LSTM,the optimal TL improves the ensemble NSE by about 10.8%(NSE increased by 6.9%,21.4%and 4.2%in China,Chile,and the U.K.,respectively),ranging from 0.033 to 0.128.In terms of the ability of extreme-event simulation,the TL also significantly mitigated the negative bias of the FHV and the positive bias of the FLV.Finally,this study clarifies the characteristics of important components in the transfer framework based on multiple experimental settings and runoff simulation results for diverse regions.The transfer framework consists of the Transfer Learning model,source data,and target data.The global application results of the transfer framework in diverse scale regions show that the optimal TL model option for the target region is relatively stable,which would save a lot of time and calculation for practical applications.The quality of the "source data" is critical to the performance of TL and its’ framework.The TL performance tends to increase with fluctuation as the number of basins in the "source data" increases,especially the initial increase of 50 basins would effectively improve the performance.As the length of the time series in the "source data" increases,the diversity information also improves the performance of the TL.However,there is a theoretical upper limit to the improvement of the TL performance by the quality of the "source data".Through comparing the difference between the process-based model and deep learning model in providing effective information for the target region under the transfer framework,this study further proved that the information derived from the deep learning model is more efficient than that provided by the process-based model.The Transfer Learning model and its framework proposed in this thesis have data compatibility and flexible model options.The successful application in the basins of China,the U.K.,and Chile demonstrates that the commonalities exist in the global hydrological dynamics,providing a scientific,reliable,and convenient technical support for the efficient application of deep learning in the field of hydrology worldwide.The advantage of TL effectively alleviates the limitations of small datasets and regional characteristics differences on a global scale,solving the problem of streamflow simulation in data-scarce regions,with wide practical value.In future studies,exploration of the globalized data synergy based on the framework of deep learning can be expected to provide technical and theoretical support for solving problems in hydrology such as scale effect and streamflow simulation in ungauged areas.