In the cold regions of western China, underlying surfaces are mainly (87.7%) composed of grassland, meadow and desert. However, the hydrological functions of these landtypes are unclear and lacking in adequate measured data. Based on the 4-year (2009-2012) observation work at point scale, small watershed scale and simulation results from Soil-Vegetation-Atmosphere Transfer (SVAT) system, it concludes that the alpine desert should be the primary runoff production area, which takes part 12% of high-cold region. While the alpine grassland and meadow (taking about 64%) contribute to the watershed runoff a little, its ecological function is more evident than its hydrological function. Combined with other research results in the literature, runoff coefficient for different landsacpes can be sorted as: glacier>cold desert>swamp meadow>hill slope shrub>meadow>alpine grassland>forest. If the vegetation belt moves upward under the global warming, the runoff coefficient will decrease in the alpine watershed of China.

More than 1.4 billion people depend on water from the Indus, Ganges, Brahmaputra, Yangtze, and Yellow rivers. Upstream snow and ice reserves of these basins, important in sustaining seasonal water availability, are likely to be affected substantially by climate change, but to what extent is yet unclear. Here, we show that meltwater is extremely important in the Indus basin and important for the Brahmaputra basin, but plays only a modest role for the Ganges, Yangtze, and Yellow rivers. A huge difference also exists between basins in the extent to which climate change is predicted to affect water availability and food security. The Brahmaputra and Indus basins are most susceptible to reductions of flow, threatening the food security of an estimated 60 million people.

无资料区的径流模拟问题是国内外水文研究的难点之一。基于相似流域的参数移植法是常用的解决方法之一,但如何判断相似流域是制约此类方法发展的难点。本文以滇池流域为例,采用自组织映射神经网络(SOM)和层次聚类分析(HCA)联合模式,选取16个流域物理特征为指标进行子流域分类,以确定相似流域。运用无分层的K-means分类的SOM法将整个滇池流域划分为7类具有水文属性的子流域组,分类情景与HCA基本一致,两者实现相互验证。采用HBV水文模型模拟子流域径流过程,并选择部分子流域进行组内参数移植交叉检验。结果显示,HBV模型可较好的模拟滇池流域径流过程;此外,子流域交叉检验结果优良,表明同组内参数可以相互移植。本文不仅为解决滇池流域无资料问题提供了可靠手段,而且由于SOM实现了高维流域特征可视化展示,有助于管理者全面、深入的把握滇池流域水文属性的空间分布特征,为进行水资源管理提供指导。

Runoff prediction in ungauged basins (PUB) is one of the difficult research areas in hydrological studies. Parameter replacement using data from similar basins is one of the common methods in dealing with the PUB problem. When basins are similar in physical properties, their hydrological behaviors are assumed to be also similar and thus the hydrological model parameters can be transferred from the donor basin to the target basin. However, it is hard to determine whether a donor basin is indeed similar to a target basin and therefore it is not always clear whether the parameters can be transferred between the basins. Existing research often focus on river basin PUB problem, with inadequate attention on lake basins that contain a number of river streams. This study addresses the PUB question using Lake Dianchi Basin as an example. Lake Dianchi Basin has a complicated river network as well as serious PUB problems. Self-organizing maps (SOM) and hierarchical clustering analysis (HCA) were jointly used to identify analogy basins based on 16 physical attributes, including area, length, slope, drainage density, Ke, mean elevation, average precipitation, six land use types and three soil types. SOM method with K-means cluster was applied to classify similar sub-basins into distinct groups and Davis-Bouldin index was used to determine the optimal group numbers. After 1000 iterations the 43 sub-basins were classified into seven groups (I-VII). This SOM-based classification result is the same as the result of HCA except for two sub-basins. Among the seven groups, group I, IV, and VII contains most of the sub-basins and the other four groups contain no more than three sub-basins each. Different groups have different characteristics and the classification result provides a guidance for local management of the lake basin. For instance, group I is located in high elevation area where the density of streams and infiltration rate of the soil are both low therefore the area is flood-prone, thus the local government should pay more attention on flood control in such area. HBV model was used to simulate the runoff process and for sub-basins where the simulation went well, their parameters were used in the cross-basin test. The cross-basins test was applied to test whether or not hydrological model parameters could be transferred between two sub-basins in the same group. Six stations in three groups were selected as examples and sub-basins in each two sub-basin pair are from the same group. The result shows that the HBV model performs well in the runoff simulation of Lake Dianchi Basin (R2≥0.718, NSE≥=0.495). The cross-basin test result is also very promising (R2≥0.654 and NSE≥0.472) — it proves that ungauged sub-basins could borrow the model parameters of gauged basins in the same group. Thus, this research provides a solution for solving the PUB problem in the Lake Dianchi Basin. This research provides a basis for solving the problem of lack of data for runoff modeling for the basin. Meanwhile, SOM visualizes multi-dimensional properties of the basin, which is useful for practitioners in water resource management to comprehensively understand the spatial distribution of hydrological characteristics of Lake Dianchi Basin.

论文对比分析了1980—2016年基于站点插值降水数据CMA(China Meteorological Administration)和APHRODITE(Asian Precipitation-Highly-Resolved Observational Data Integration Towards Evaluation)、卫星遥感降水数据PERSIANN-CDR(Precipitation Estimation from Remotely Sensed Information using Artificial Neural Network-Climate Data Record)和GPM(Global Precipitation Measurement)、大气再分析数据GLDAS(Global Land Data Assimilation System)以及区域气候模式输出数据HAR(High Asia Refined analysis)在雅鲁藏布江7个子流域的降水时空描述,利用国家气象站点数据对各套降水数据进行单点验证,并以这6套降水数据驱动VIC(Variable Infiltration Capacity)大尺度陆面水文模型反向评估了各套降水产品在雅鲁藏布江各子流域径流模拟中的应用潜力。结果表明：① PERSIANN-CDR和GLDAS年均降水量最高(770~790 mm),其次是HAR和GPM(650~660 mm),CMA和APHRODITE年均降水量最低(460~500 mm)。除GPM外,其他降水产品在各子流域都能表现季风流域的降水特征,约70%~90%的年降水量集中在6—9月份。② 除PERSIANN-CDR和GLDAS外,其他降水产品皆捕捉到流域降水自东南向西北递减的空间分布特征。其中,HAR数据空间分辨率最高,表现出更详细的流域内部降水空间分布特征。③ 与对应网格内的国家气象站降水数据对比显示,APHRODITE、GPM和HAR降水整体低估(低估10%~30%),且严重低估的站点主要集中在下游(低估40%~120%)。PERSIANN-CDR和GLDAS整体表现为高估上游流域站点降水(高估28%~60%),但低估下游流域站点降水(低估11%~21%)。④ 在流域径流模拟上,当前的6套降水产品在精度或时段上仍无法满足水文模型模拟的需求。⑤ 通过水文模型反向评估,6套降水产品中区域气候模式输出的HAR在流域平均降水量和季节分配上更合理。

The gauge-based precipitation data from the National Climate Center, China Meteorological Administ-ration (CMA), Asian Precipitation-Highly-Resolved Observational Data Integration Towards Evaluation (APHRODITE), Precipitation Estimation from Remotely Sensed Information using Artificial Neural Network-Climate Data Record (PERSIANN-CDR), Global Precipitation Measurement (GPM), Global Land Data Assimilation System (GLDAS), High Asia Refined analysis (HAR) are compared with each other and evaluated by the precipitation data from 16 national meteorological stations during 1980-2016 in the Yarlung Zangbo River and its sub-basins. The potential utilities of these multiple precipitation datasets are then systematically evaluated as inputs for the variable infiltration capacity (VIC) macroscale land surface hydrologic model. The results show that: 1) PERSIANN-CDR and GLDAS contain the largest precipitation estimates among the six datasets with mean annual precipitation of 770-790 mm, followed by the HAR and GPM (650-660 mm), while CMA and APHRODITE contain the lowest precipitation estimates with mean annual precipitation of 460-500 mm. All the products can detect the large-scale monsoon-dominated precipitation regime in the Yarlung Zangbo River and its sub-basins with 70%-90% of annual total precipitation occurring in June-September except the GPM. 2) The general spatial pattern of the annual mean precipitation fields is roughly in agreement among the six datasets, with a decreasing trend from the southeast to the northwest in the Yarlung Zangbo River Basin except the PERSIANN-CDR and GLDAS. 3) Relative to the data from the national meteorological stations, APHRODITE, GPM, and HAR generally underestimate precipitation by 10%-30%, while PERSIANN-CDR and GLDAS overestimate precipitation from stations in upstream sub-basins by 28%-60% and underestimate precipitation from stations in downstream sub-basins by 11%-21%. 4) The six precipitation datasets cannot satisfy the needs of hydrological simulation in term of accuracy or period in the basin. 5) HAR precipitation data—output of regional climate model—show more reasonable amount and seasonal pattern among the six datasets in the upper Brahmaputra according to the inverse evaluation by VIC hydrological model.

. The degree of belief we have in predictions from hydrologic models will normally depend on how well they can reproduce observations. Calibrations with traditional performance measures, such as the Nash-Sutcliffe model efficiency, are challenged by problems including: (1) uncertain discharge data, (2) variable sensitivity of different performance measures to different flow magnitudes, (3) influence of unknown input/output errors and (4) inability to evaluate model performance when observation time periods for discharge and model input data do not overlap. This paper explores a calibration method using flow-duration curves (FDCs) to address these problems. The method focuses on reproducing the observed discharge frequency distribution rather than the exact hydrograph. It consists of applying limits of acceptability for selected evaluation points (EPs) on the observed uncertain FDC in the extended GLUE approach. Two ways of selecting the EPs were tested – based on equal intervals of discharge and of volume of water. The method was tested and compared to a calibration using the traditional model efficiency for the daily four-parameter WASMOD model in the Paso La Ceiba catchment in Honduras and for Dynamic TOPMODEL evaluated at an hourly time scale for the Brue catchment in Great Britain. The volume method of selecting EPs gave the best results in both catchments with better calibrated slow flow, recession and evaporation than the other criteria. Observed and simulated time series of uncertain discharges agreed better for this method both in calibration and prediction in both catchments. An advantage with the method is that the rejection criterion is based on an estimation of the uncertainty in discharge data and that the EPs of the FDC can be chosen to reflect the aims of the modelling application, e.g. using more/less EPs at high/low flows. While the method appears less sensitive to epistemic input/output errors than previous use of limits of acceptability applied directly to the time series of discharge, it still requires a reasonable representation of the distribution of inputs. Additional constraints might therefore be required in catchments subject to snow and where peak-flow timing at sub-daily time scales is of high importance. The results suggest that the calibration method can be useful when observation time periods for discharge and model input data do not overlap. The method could also be suitable for calibration to regional FDCs while taking uncertainties in the hydrological model and data into account.\n

以渭河流域为例,从流域水文循环的角度出发,在SWAT(Soil and Water Assessment Tool)分布式水文模型和Palmer干旱指数(Palmer Drought Severity Index)原理的基础上提出了干旱分析模型SWAT-PDSI,对渭河流域干旱的时空演变规律和发生频率进行了分析。研究结果表明：①经率定和验证的SWAT模型能够较好地模拟渭河流域的水文变化过程;②利用SWAT-PDSI对典型干旱事件(1995年干旱)的评估结果显示,该模型能较好地反映渭河流域干旱的时空差异和变化规律;③渭河流域、渭河干流和泾河流域均表现为变干的趋势,而北洛河流域表现为变湿的趋势,但均未通过95%的置信水平检验;④渭河流域多数子流域的SWAT-PDSI多年平均值处于-1~1,说明该流域多数地区处于正常状态;⑤渭河流域北部的北洛河流域和泾河流域的上游地区易发生干旱,发生中等以上、严重以上和极端干旱事件的频率最高。

Drought is a complex natural hazard that is difficult to define and assess. By considering the hydrological cycle of river basins, a drought evaluation model SWAT-PDSI was constructed based on the Palmer Drought Severity Index(PDSI) and simulated results of SWAT (Soil and Water Assessment Tool) in the Weihe River Basin. Moreover, characteristics of the spatiotemporal distribution and frequency of drought hazard in the basin were analyzed. Results showed that: (1) The calibrated and validated SWAT model can be used to predict the hydrological process in the Weihe River Basin since it results in similar simulated trend of change as the observed data; (2) The SWAT-PDSI model based on SWAT and PDSI well describes the characteristics of drought in the Weihe River Basin as verified by the drought indices in 1995; (3) Temporally, the Weihe River Basin, Weihe River mainstream, and Jinghe River Basin showed a drying trend and the Beiluohe River Basin showed a wetter trend, but these were not statistically significant at the 0.05 confidence level; (4) Spatially, the annual average of SWAT-PDSI was between -1 and 1 in most subbasins; (5) High frequency drought areas were mainly distributed in the upstream of the Beiluohe River Basin and the Jinghe River Basin.

The accuracy of five soft computing techniques was assessed for the prediction of monthly streamflow of the Gilgit river basin by a cross-validation method. The five techniques assessed were the feed-forward neural network (FFNN), the radial basis neural network (RBNN), the generalised regression neural network (GRNN), the adaptive neuro fuzzy inference system with grid partition (Anfis-GP) and the adaptive neuro fuzzy inference system with subtractive clustering (Anfis-SC). The interaction between temperature and streamflow was considered in the study. Two statistical indexes, mean square error (MSE) and coefficient of determination (R-2), were used to evaluate the performances of the models. In all applications, RBNN and Anfis-SC were found to give more accurate results than the FFNN, GRNN and Anfis-GP models. The effect of periodicity was also examined by adding a periodicity component into the applied models and the results were compared with a statistical model (seasonal autoregressive integrated moving average (Sarima)) to check the prediction accuracy. The results of this comparison showed that periodicity inputs improved the prediction accuracy of the applied models and, in all cases, the soft computing models performed much better than the Sarima model. The periodic RBNN and Anfis-SC models increased the MSE accuracy of Sarima by 25.5-24.7%.

Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient-based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O(1). Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.

. Accurate river streamflow forecasts are a vital tool in the fields of water security, flood preparation and agriculture, as well as in industry more generally. Traditional physics-based models used to produce streamflow forecasts have become increasingly sophisticated, with forecasts improving accordingly. However, the development of such models is often bound by two soft limits: empiricism – many physical relationships are represented empirical formulae; and data sparsity – long time series of observational data are often required for the calibration of these models. Artificial neural networks have previously been shown to be highly effective at simulating non-linear systems where knowledge of the underlying physical relationships is incomplete. However, they also suffer from issues related to data sparsity. Recently, hybrid forecasting systems, which combine the traditional physics-based approach with statistical forecasting techniques, have been investigated for use in hydrological applications. In this study, we test the efficacy of a type of neural network, the long short-term memory (LSTM), at predicting streamflow at 10 river gauge stations across various climatic regions of the western United States. The LSTM is trained on the catchment-mean meteorological and hydrological variables from the ERA5 and Global Flood Awareness System (GloFAS)–ERA5 reanalyses as well as historical streamflow observations. The performance of these hybrid forecasts is evaluated and compared with the performance of both raw and bias-corrected output from the Copernicus Emergency Management Service (CEMS) physics-based GloFAS. Two periods are considered, a testing phase (June 2019 to June 2020), during which the models were fed with ERA5 data to investigate how well they simulated streamflow at the 10 stations, and an operational phase (September 2020 to October 2021), during which the models were fed forecast variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) Integrated Forecasting System (IFS), to investigate how well they could predict streamflow at lead times of up to 10 d. Implications and potential improvements to this work are discussed. In summary, this is the first time an LSTM has been used in a hybrid system to create a medium-range streamflow forecast, and in beating established physics-based models, shows promise for the future of neural networks in hydrological forecasting.\n

. Long Short-Term Memory (LSTM) networks have been applied to daily discharge prediction with remarkable success.\nMany practical applications, however, require predictions at more granular timescales.\nFor instance, accurate prediction of short but extreme flood peaks can make a lifesaving difference, yet such peaks may escape the coarse temporal resolution of daily predictions.\nNaively training an LSTM on hourly data, however, entails very long input sequences that make learning difficult and computationally expensive.\nIn this study, we propose two multi-timescale LSTM (MTS-LSTM) architectures that jointly predict multiple timescales within one model, as they process long-past inputs at a different temporal resolution than more recent inputs.\nIn a benchmark on 516 basins across the continental United States, these models achieved significantly higher Nash–Sutcliffe efficiency (NSE) values than the US National Water Model.\nCompared to naive prediction with distinct LSTMs per timescale, the multi-timescale architectures are computationally more efficient with no loss in accuracy.\nBeyond prediction quality, the multi-timescale LSTM can process different input variables at different timescales, which is especially relevant to operational applications where the lead time of meteorological forcings depends on their temporal resolution.\n

冻土覆盖率高的小流域的径流形成受温度因素控制明显，普通水文模型不适用，而常规冻土水文模型因需要较多的气象观测要素而难以应用。考虑冻土流域产流机制，利用青藏高原腹地风火山小流域2017—2018年逐日降水、气温、径流观测数据，以降水、气温为输入，径流为输出，基于长短期记忆神经网络（LSTM）建立了适用于小流域尺度的冻土水文模型，并利用2019年观测数据进行验证。模型得益于LSTM特殊的细胞状态和门结构能够学习、反映活动层冻融过程和土壤含水量变化，具有一定的冻土水文学意义，能很好地模拟冻土区径流过程。模型训练期R²、NSE均为0.93，RMSE为0.63 m³·s^-1，验证期R²、NSE分别为0.81、0.77，RMSE为0.69 m³·s^-1。同时，为了验证模型可靠性，将模型应用于邻近的沱沱河流域，模型训练期（1990—2009年）R²、NSE均为0.73，验证期（2010—2019年）R²、NSE分别为0.66、0.64，模拟结果较好。考虑到未来气候变化，通过模型对风火山流域径流进行了预测：降水每增加10%，年径流增加约12%；气温每升高0.5 ℃，年径流增加约1%；春季融化期、秋季冻结期径流增幅明显，而由于蒸发加剧、活动层加深，径流在8月出现了减少。模型经训练后依靠降水、气温作为输入能较好地模拟、预测青藏高原冻土区小流域径流，为缺少土壤温度、水分等观测数据的冻土小流域径流研究提供了一种简单有效并具有一定物理意义的方法。

The Qinghai-Tibet Plateau， known as the Third Pole， with 42.4% permafrost coverage， is sensitive to climate change. Runoff generation of small-scale watershed with high permafrost coverage is significantly controlled by temperature factor， which makes ordinary hydrological model unsuitable for this area， while lack of measured data such as soil temperature and moisture makes common permafrost hydrological model difficult to be applied. Moreover， increase in air temperature will result in permafrost degradation， which fundamentally changes the hydrogeological conditions in permafrost regions and finally changes the runoff process in permafrost watershed. Thus， the air temperature is a key factor in permafrost runoff modeling. LSTM （long short-term memory） is a special recurrent neural network with a more detailed internal processing unit， which contains cell state and gate structures， helping it effectively use long-distance time series information in hydrology. In this study， we developed a permafrost hydrological model at small-scale watershed based on LSTM neural networks with the consideration of runoff generation mechanism in permafrost. And it was applied in Fenghuoshan watershed， a tributary of the source region of the Yangtze River with 100% permafrost coverage， located at the central of Qinghai-Tibet Plateau. In the LSTM permafrost hydrological model， the precipitation and air temperature are employed as model inputs， while the runoff is regarded as the output. The daily precipitation， air temperature， and runoff observation data from year 2017 to 2018 were employed to train the model， and the dataset of year 2019 was used for model validation. Benefiting from the special cell state and gate structures of LSTM， the model is capable of learning and reflecting freeze-thaw processes and soil moisture seasonal variation in the active layer， with cell state evolution of some LSTM neurons consistent with these processes. It gives the model a certain permafrost hydrological significance and the high performance of permafrost runoff simulation. The values of R2， NSE and RMSE were 0.93， 0.93， 0.63 m3·s-1 during training period， 0.81， 0.77， 0.69 m3·s-1 during validation period， respectively. Besides， the model performed well in all periods within the year， including spring flood period， summer recession period， summer flood period， autumn recession period and winter freezing period. The model was also applied in the Tuotuohe watershed， which is close to Fenghuoshan watershed. The values of R2， NSE were 0.73， 0.73 during training period， 0.66， 0.64 during validation period， respectively. The model result was comparable to the results of CRHM model and WEB-DHM-SF model， which demonstrates it was reasonable and reliable. And the model was employed to predict runoff changes of Fenghuoshan watershed under 10 different climate change scenarios， those were 10% or 20% increase in precipitation with 0 ℃， 1.0 ℃， 2.0 ℃ increase in air temperature and 0.5 ℃， 1.0 ℃， 1.5 ℃， 2.0 ℃ increase in air temperature with precipitation unchanged. It shows that every 10% increase in the precipitation will result in approximately 12% increase in the annual runoff， while every 0.5 ℃ increase in the air temperature will result in approximately 1% increase. The thaw of underground ice induced by increase in air temperature contributes little to the runoff increase. However， it significantly changes the runoff process through altering the freeze-thaw processes in the active layer， which has a different influence on the runoff during different periods. Increase in the air temperature will result in significant increase in the runoff during spring thaw and autumn freeze period， while the runoff decreases in August due to the increased evaporation and deepened active layer caused by the increase in the air temperature. Meanwhile， increase in the air temperature prolongs the thaw period and shortens the freeze period， which will change the runoff compositions. This illustrates the temperature-induced variable source area runoff generation process， namely the runoff generation in permafrost region not solely determined by soil moisture but controlled by temperature conditions. The results show that the trained model can be employed to simulate and predict runoff of small permafrost watershed with only precipitation and air temperature as inputs， which are easier available in permafrost areas. It provides a simple and effective method， with a certain physical meaning， for permafrost watershed lacking observation data such as soil temperature and moisture.

. Rainfall–runoff modelling is one of the key\nchallenges in the field of hydrology. Various approaches exist, ranging from\nphysically based over conceptual to fully data-driven models. In this paper,\nwe propose a novel data-driven approach, using the Long Short-Term Memory\n(LSTM) network, a special type of recurrent neural network. The advantage of\nthe LSTM is its ability to learn long-term dependencies between the provided\ninput and output of the network, which are essential for modelling storage\neffects in e.g. catchments with snow influence. We use 241 catchments of the\nfreely available CAMELS data set to test our approach and also compare the\nresults to the well-known Sacramento Soil Moisture Accounting Model (SAC-SMA)\ncoupled with the Snow-17 snow routine. We also show the potential of the LSTM\nas a regional hydrological model in which one model predicts the discharge\nfor a variety of catchments. In our last experiment, we show the possibility\nto transfer process understanding, learned at regional scale, to individual\ncatchments and thereby increasing model performance when compared to a LSTM\ntrained only on the data of single catchments. Using this approach, we were\nable to achieve better model performance as the SAC-SMA + Snow-17, which\nunderlines the potential of the LSTM for hydrological modelling applications.\n

论文基于2003—2014年水文资料,采用长短期记忆神经网络(Long-Short Term Memory,LSTM),构建了汉江上游安康站日径流预测模型,评价了不同输入条件下日径流预测的精度。结果表明：当预见期为1 d时,在仅以安康站前期日径流量作为输入的条件下,LSTM模型在训练期和检验期的效率系数分别达到0.68和0.74;如再将流域前期面雨量和上游石泉站前期日径流量加入LSTM网络作为输入变量,安康站日径流量预测效果将更好,训练期和检验期的效率系数最高可达到0.83和0.84,均方根误差也有显著削减,且对主要洪峰流量的预测能力也有一定提高。此外,LSTM可以有效避免过拟合等问题,具有较好的泛化性能。但当预见期从1 d延长至2、3 d时,LSTM的预测精度显著降低。