Temporal-spatial Evolution Characteristics And Future Trends Of Potential Evaporation In Three-river Headwaters Region

Evaporation is a vital component of the meteorological-hydrological processes,which is an important way of transporting water from the earth's surface to the atmosphere in the water cycle.Potential evaporation(PE)is an important parameter for evaluating regional evaporation capacity.Regarded as China's main source of water,Three-River Headwaters Region(TRHR)is also an important ecological barrier in West China.Therefore,an in-depth study on the evolution characteristics and future trend of PE in this region can help to better understand the future meteorological-hydrological conditions in this region,and further better protect the ecological environment of this region.This study collects the observed evaporation data recorded at 14 sites in the TRHR from 1960 to 2014.Firstly,using the IDW(Inverse Distance Weighted)method and Man-Kendall test,the spatial distribution and temporal changes of PE in this ecologically sensitive area are analyzed,and the results indicate that: 1)the central part of this region has the lowest PE,while the northeastern and southwestern parts have relatively higher PE;2)the annual mean PE of this region has a quite weak correlation with longitude and latitude,but has a relatively stronger correlation with altitude;3)at the seasonal scale,PE increases from spring to summer,decreases sharply from summer to autumn,and continues to decline from autumn to winter;4)from 1960 s to 2000 s,PE firstly shows a downward trend,reaches the lowest value in1980 s,and then begins to rise after that;5)only 8 sites have significant changing trends in PE,among which the numbers of sites showing increasing and decreasing trends are the same.Furthermore,in this study,BP(Back-Propagation)neural network and LSTM(Long Short-Term Memory)model are used to predict the changes of PE in the TRHR for the next 30 years.The first 80% of the data from 1960-2014 are used as the training/validation set and the last 20% as the test set,and the results indicate that:the NSE(Nash-Sutcliffe efficiency coefficient)value of BP neural network is 0.89,and the MAPE(Mean Absolute Percentage Error)value is 10.76%;while the NSE value of LSTM model is 0.90,and the MAPE value is 9.87%.Based on the two trained models,the long-term PE in the TRHR is predicted,and the results indicate that: the curve of the BP neural network has a relatively stable feature with only a little randomness,which can hardly extract effective features;in contrast,the LSTM model has higher randomness and can extract more effective features.With reference to the changing rate of PE,the values are 5.26mm/30 a for the BP neural network and 252.9mm/30 a for the LSTM model.