High and cold mountainous area runoff simulation method for improving VI-glacier model by combining LSTM (Long Short Term Memory)

Abstract

The invention provides a method for simulating runoff in a cold and cold mountainous area by combining LSTM (Long Short Term Memory) to improve a VI-glacier model. The method is used for solving the technical problems that runoff components in the cold and cold mountainous area are diverse and each runoff source cannot be directly distinguished. The method comprises the following steps: firstly, determining model input data including DEM data, vegetation data, soil data, meteorological data, hydrological data and glacier boundary data; a GPM rainfall product is adopted to drive meteorological data; secondly, introducing a degree-day factor model to quantify the glacier ablation process, coupling the degree-day factor model with the VIC model, and constructing a VI-glacier model; then, an SCE-UA automatic optimization algorithm and an artificial trial and error method are combined to calibrate the VICs-glacier model, and an optimal parameter value is obtained; and meanwhile, error correction is carried out on a simulation result of the VI-glacier model through the LSTM model, so that runoff correction is realized. The model provided by the invention can relieve the uncertainty of water resources caused by glacier changes, and the future runoff can be predicted more accurately.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of hydrological runoff simulation, and in particular to a high-cold mountainous area runoff simulation method combining LSTM to improve a VIC-glacier model. BACKGROUND

[0002] In view of the multi-source nature of high-cold mountainous area runoff and the difficulty in directly distinguishing various runoff sources, it is particularly important to distinguish these runoff components. Accurate observation or simulation of glacial meltwater is crucial for understanding hydrological cycle and managing water resources. However, few distributed hydrological models are designed for high-cold regions and take into account the complexity of runoff generation, and most existing hydrological models ignore glacial melting. The VIC model (Variable Infiltration Capacity Model) is a macro hydrological model based on a distributed object developed by Liang Xu. Due to its advantages of considering water balance and energy balance, being compatible with storage runoff and excess infiltration runoff, and considering the influence of soil unevenness within the sub-grid on runoff, the VIC model is widely used in runoff simulation of basins. Hydrological models with glacial melting process can provide an effective tool for simulating the hydrological response of climate change in glacialized basins. The Degree-Day Factor Model (DDFM) is derived from the demand for estimation of ice and snow melt in glaciology and hydrology research. Since the VIC model does not have a glacial module, scholars have coupled the Degree-Day Factor Model (VIC-glacier) to build a glacial runoff model, and have achieved significant results in many basins. Su Ting et al. simulated the runoff change of TRSR from 1984 to 2018 by using the VIC-galcier model and studied the contribution of each component of runoff (rainfall, snowmelt and glacial runoff) to the total runoff and the reason for the change in runoff. Gai Huanghe et al. built a VIC model and a VIC-glacier model for the Lhasa River. Compared with the VIC model, the VIC-glacier model has better runoff simulation effect in the study area, with a correlation coefficient of daily runoff simulation value and observation value close to 0.8 and a Nash efficiency coefficient above 0.75. Wang Ning et al. built a VIC-glacier model for the northwest three high mountain source catchment areas of the Tarim River to simulate the flow and flood, which showed good simulation effect. Ren et al. combined the energy balance glacial meltwater scheme with the variable infiltration capacity hydrological model to build a VIC-glacier model and applied it to the Muji Basin in the eastern Pamirs. This model has good application potential for large basins with limited meteorological observations. Runoff simulation in areas rich in glacial resources can effectively alleviate the uncertainty of water resources caused by glacial change and more accurately predict future runoff trends.

[0003] However, due to the nonlinearity, complexity, time-varying nature and parameter factors of hydrological models (such as vegetation index, soil porosity index), although models such as VIC model consider the internal mechanism of hydrology and use parameter calibration method to improve the performance of the model, the simulation accuracy is still different from the true value, and the simulation error may come from various reasons, such as inaccurate input data, structural limitations of the model itself or complexity of the basin characteristics. The output error correction, that is, the correction of the error between the prediction results of the model and the observed values, is the most direct way to reduce the prediction error, so we consider using the error correction method to reduce the uncertainty brought by the VIC model and improve the accuracy of the model. Common error correction methods include time series analysis, recursive least squares, Kalman filtering, etc. Among them, the autoregressive model (AR) in time series analysis is simple to operate, has high prediction accuracy for stationary time series, and has been widely used in runoff simulation and prediction at different time scales. For example, Xu et al. compared the advantages and disadvantages of K-nearest neighbor algorithm, traditional error autoregressive algorithm and feedback simulation algorithm, and proved that the three methods can effectively correct the prediction results. Shirisha et al. conducted runoff prediction in three basins of Banha, Harsoor and Kadaikulam, and used fuzzy error prediction correction model to correct the runoff, which achieved good results. In recent years, with the rapid development of deep learning technology, LSTM model has been gradually introduced into hydrological runoff simulation and flood prediction due to its excellent performance in time series prediction. By learning the error pattern in historical flow data through LSTM model, we can correct the simulation results to improve the overall simulation accuracy of VIC-glacier model. Xu et al. established a Long Short-Term Memory (LSTM) model in the middle reaches of the Yellow River to predict floods, and compared the prediction accuracy of floods at different prediction periods. In this study, LSTM model and AR model are used to correct the simulation results of VIC-glacier model to improve the model accuracy, and the correction accuracy of the two models is compared and verified. SUMMARY

[0004] In order to solve the technical problem that the runoff components in alpine mountainous areas are diverse and each runoff source cannot be directly distinguished, the application provides an alpine mountainous area runoff simulation method combining LSTM to improve VIC-glacier model, which couples VIC model and glacier degree-day model to distinguish glacier runoff, and uses automatic calibration of SCE-UA and LSTM model error correction method to further optimize the coupled model, so as to improve the accuracy of the model and effectively alleviate the uncertainty of water resources caused by glacier change, and more accurately predict the future runoff trend.

[0005] In order to achieve the above object, the technical scheme of the present application is as follows:

[0006] A method for improving the runoff simulation of alpine mountainous areas by combining LSTM and VIC-glacier model, the steps of which are as follows:

[0007] Step 1: Collect data: the input data includes DEM data, vegetation data, soil data, meteorological data, hydrological data and glacier boundary data; and the GPM precipitation product is used to drive the meteorological data;

[0008] Step 2: Establish the VIC-glacier coupled model: introduce the degree-day factor model to quantify the glacier ablation process, and couple the degree-day factor model with the VIC model to build the VIC-glacier model;

[0009] Step 3: Parameter calibration of the VIC-glacier model: combine the SCE-UA automatic optimization algorithm with the artificial trial and error method to calibrate the VIC-glacier model and obtain the optimal parameter value;

[0010] Step 4: Correct the simulation results of the VIC-glacier model by the LSTM model to realize runoff correction.

[0011] Preferably, the DEM data comes from the SRTM plan of the geographic spatial data cloud, the spatial distribution rate is 90m x 90m, and 215 0.08333° x 0.08333° calculation grids are established in the study area by comprehensively considering the spatial resolution of precipitation data and underlying surface remote sensing data and the running efficiency; the vegetation data is the global 1km land cover data released by the University of Maryland, and the vegetation in the study area is mainly open shrublands and grasslands according to the proportion of different vegetation types in the grid; the soil data comes from the 5' resolution soil database provided by FAO; the meteorological data comes from the station observation of China Meteorological Administration, including daily minimum temperature, daily maximum temperature and daily average wind speed, and the meteorological data in the unit grid is calculated by using the inverse distance weighted interpolation method; the hydrological data comes from the daily runoff data of Tongguzilake hydrological station from 2010 to 2018 and the daily runoff data of Uruwat hydrological station from 2010 to 2018 provided by Tarim River Basin Administration; the glacier boundary data comes from the second glacier inventory dataset.

[0012] Preferably, the calculation method of glacier ablation amount is as follows:

[0013] M = DDF x PDD (1);

[0014]

[0015] Where M is the glacier melt water equivalent; PDD is the positive accumulated temperature of the period, and DDF is the degree-day factor; T t is the daily average temperature of day t, H t is a logical variable, which is 1 when T t > 0, and 0 when T t ≤ 0.

[0016] Preferably, the method for parameterizing the VIC-glacier model is as follows: first, the initial values of the model parameters are adjusted using the trial-and-error method; then, the SCE-UA algorithm is used for automatic optimization and calibration to minimize the error between the model output and the actual observation data; during the iterative calculation process of the SCE-UA algorithm, the initial values determined by the trial-and-error method are appropriately adjusted according to the results to improve the accuracy of the parameters.

[0017] Preferably, the method for error correction of the simulation results of the VIC-glacier model by the LSTM model is as follows: the LSTM error correction method is used to simulate the residual error sequence of the measured value and the model simulation value, and the precipitation data is used as the input; specifically, the residual error sequence of the simulated runoff of the VIC-glacier model and the observation value is extracted, the LSTM neural network is introduced to train the spatio-temporal nonlinear characteristics of the residual error sequence, the Adam optimizer and the early stopping strategy are used to prevent overfitting, and the predicted residual error is dynamically superimposed on the output of the VIC-glacier model to realize the correction of the runoff.

[0018] Preferably, a plurality of performance indicators are used to evaluate the accuracy and robustness of the simulated runoff, including the Nash-Sutcliffe efficiency coefficient (NSE), the mean absolute error (MAE), the root mean square error (RMSE), the Kling-Gupta efficiency coefficient (KGE), the median absolute error (MdAE), and the maximum absolute error (MaxAE); the calculation formulas are as follows:

[0019] The calculation formula of the Nash-Sutcliffe efficiency coefficient is as follows:

[0020]

[0021] The calculation formula of the mean absolute error is as follows:

[0022]

[0023] The calculation formula of the root mean square error is as follows:

[0024]

[0025] The calculation formula of the Kling-Gupta efficiency coefficient is as follows:

[0026]

[0027] The calculation formula of median absolute error is as follows:

[0028] MdAE = median (|Q i,s -Q i,o |) (7);

[0029] The calculation formula of maximum absolute error is as follows:

[0030] MaxAE = max (|Q i,s -Q i,o |) (8);

[0031] In the formula, Q i,o refers to the measured flow series data; Q i,s refers to the flow series data simulated by the model; refers to the measured monthly and daily average surface runoff; r is the Pearson correlation coefficient between the simulation value and the observation value; and β is the bias ratio. is the variation rate, and CV represents the coefficient of variation.

[0032] Preferably, the expression of the LSTM error prediction is:

[0033] ε t+k = f{P, Q, ε t+k-1 , ε t+k-2} (12);

[0034] In the formula, ε t is the flood forecast error at the current time, ε t-1 represents the flood forecast error at the previous time; P is the rainfall at the corresponding time; Q is the previous flood forecast flow at the corresponding time; and ε t+k represents the error predicted by the model at the future time.

[0035] The beneficial effects of the application are:

[0036] 1) The degree-day factor model is introduced to describe the glacier ablation process, which is coupled with the VIC model to construct the VIC-glacier model; the VIC-glacier model is particularly suitable for glacier-dominated rivers by strengthening the mechanism constraint of glacier runoff.

[0037] 2) The SCE-UA automatic optimization algorithm is combined with the artificial trial and error method to calibrate the VIC model. First, the artificial trial and error method is used to adjust the initial value of the model parameter; then, the SCE-UA algorithm is used for automatic optimization calibration to minimize the error between the model output and the actual observation data. In the iterative calculation process of the SCE-UA algorithm, the initial value determined by the artificial trial and error method is appropriately adjusted according to the results to improve the parameter accuracy.

[0038] 3) The VIC-glacier model combined with the LSTM error correction technology has multiple advantages, such as capturing long-term correlation, extracting local time features, and optimizing uncertainty estimation, which can provide more reliable prediction, thereby effectively alleviating the uncertainty of water resources caused by glacier change. This enables decision-makers to take timely flood control measures and more accurately predict future runoff trends. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 Figure 1 is a general map of the study area; (a) is an elevation map of different altitudes in Xinjiang, (b) is a schematic diagram and glacier coverage map of the Kalakashi River Basin, and (c) is a schematic diagram and glacier coverage map of the Yulong Kashi River Basin.

[0041] Figure 2 Figure 2 is a technical roadmap of the present application.

[0042] Figure 3 Figure 3 is a flow concentration schematic diagram of the VIC-glacier model of the present application.

[0043] Figure 4 Figure 4 is a schematic diagram of the internal structure of LSTM.

[0044] Figure 5 Figure 5 is an error prediction structure diagram.

[0045] Figure 6 Figure 6 is a conceptual diagram of the VIC-glacier-LSTM model.

[0046] Figure 7 Figure 7 is the observed daily runoff and simulated (SIM) daily runoff of the VIC-glacier model and the VIC model in the calibration period (2010-2015) and the verification period (2016-2018) of two basins; (a) and (b) are Yulong Kashi River; (c) and (d) are Kalakashi River.

[0047] Figure 8 Figure 4 shows the monthly variation of glacier runoff for the upstream of Hotan Glacier; (a) and (b) show the monthly average flow of glacier runoff and non-glacier runoff for Yurungkash River and Kaxkash River, respectively; (c) and (d) show the percentage of monthly glacier runoff supply for Yurungkash River and Kaxkash River, respectively.

[0048] Figure 9 Figure 5 shows the seasonal variation of glacier runoff percentage for Yurungkash River and Kaxkash River; (a) and (c) show the seasonal variation of glacier runoff percentage for Yurungkash River; (b) and (d) show the seasonal variation of glacier runoff percentage for Kaxkash River.

[0049] Figure 10 Figure 6 shows the runoff simulation results of LSTM error corrected VIC-glacier model for Yurungkash River and Kaxkash River during the training period and the test period; (a) and (b) show the runoff simulation results for Yurungkash River; (c) and (d) show the runoff simulation results for Kaxkash River.

[0050] Figure 11 Figure 7 shows the comparison of runoff prediction process curves of LSTM error corrected VIC-glacier model and AR error corrected VIC-glacier model for Yurungkash River and Kaxkash River; (a)-(f) show the comparison results for Yurungkash River, wherein (a) and (b) are 1-day prediction period, (c) and (d) are 3-day prediction period, and (e) and (f) are 6-day prediction period; (g)-(l) show the comparison results for Kaxkash River, wherein (g) and (h) are 1-day prediction period, (i) and (j) are 3-day prediction period, and (k) and (l) are 6-day prediction period.

[0051] Figure 12 Figure 8 shows the comparison of monthly average flow between the error corrected model and the measured value; (a) is for Yurungkash River; (b) is for Kaxkash River.

[0052] Figure 13 Figure 9 shows the Bootstrap model interval prediction results; (a) is for 1-day prediction period; (b) is for 3-day prediction period; (c) is for 6-day prediction period. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0054] To explore the contribution of glacier meltwater to runoff in alpine mountainous areas and improve the accuracy of runoff prediction, the present application proposes to build a VIC-glacier coupled model for the Yulong Kashi River and the Kekeshe River (the upstream of the Hetian River basin). The model aims to simulate the glacier-fed runoff process and correct errors through long short-term memory (LSTM) model and autoregressive (AR) model to improve its performance. The performance of the two error correction methods is evaluated through the performance results under different prediction periods. In addition, the present application quantifies the contribution of glacier meltwater to total runoff and explores the related uncertainties, providing support for the comprehensive management of water resources and ecological protection in Xinjiang and other glacier-dominated high mountain basins.

[0055] Figure 1 A research area overview map is given. The two tributaries of the upstream of the Hetian River, the Kekeshe River in the west and the Yulong Kashi River in the east, both originate from the glacier area of the Kunlun Mountains and are important water sources and lifelines of the oasis in southern Xinjiang. The Kekeshe River originates from the north slope of the Karakoram Mountains, near the eastern glacier group of Mount Jogle, with a geographical position roughly between 34°50'N-36°50'N and 77°30'E-79°50'E. The river is 808 km long, and the Wuluwati hydrological station is located at the outlet of the basin, with a control area of about 19983 km 2 , and the average annual runoff at the outlet is 22.26 billion m 3 . The Yulong Kashi River originates from the Mushi Mountain glacier area on the north slope of the middle section of the Kunlun Mountains, with a geographical position roughly between 35°20'N-36°50'N and 79°20'E-81°30'E. The river is 504 km long, and the Tongguziluoke hydrological station is located at the outlet of the basin, with a control area of about 14575 km 2 , and the average annual runoff at the outlet is 22.64 billion m 3 . The Yulong Kashi River and the Kekeshe River are called the Hetian River after they join at the Koshilash, and finally join the Tarim River at the Xiaota hydrological station in Aksu, becoming a key water source for the ecology and agriculture of the Tarim Basin. The hydrological characteristics of the two tributaries are obviously controlled by seasonal precipitation and high mountain snowmelt. Snowmelt in summer is the main source of supply, and the river runoff is less in winter, showing obvious seasonal runoff characteristics. The basin is located in a typical continental arid climate zone, with an average annual temperature of about 12°C, little rain, cold and dry, and in extreme drought state all year round.

[0056] According to the second glacier inventory data, the number of glaciers in the Hetian area is the most in the whole region (5640), and the area and ice storage are also the largest (6812.67 km 2 , 632.66 km 3 ), accounting for 27.25%, 30.11% and 29.36% of the total glaciers in Xinjiang, respectively. Among them, there are 1331 glaciers in the Yulong Kashi River basin, with a total area of about 3000 km 2The reserves are 410.32 km². 3 Among them, 17 glaciers are longer than 10 km; the Karakash River basin has a total of 1987 glaciers, with a total area of ​​approximately 2000 km². 2 The reserves are 156.09 km². 3 The glaciers of the two rivers exhibit diverse distribution types (see Table 1). The Hotan River source area is the region with the largest and most concentrated distribution of modern glaciers on the main peak of the western section of the Kunlun Mountains, and glacial meltwater is the main source of water supply for the Hotan River's endorheic basin. The average glacier coverage in the southern mountainous region of the Tarim Basin is 7.0%, with the Yurungkash River having the highest coverage, followed by the Karakash River. Therefore, selecting the Yurungkash River basin and the Karakash River basin as typical glacial runoff basins in the high-altitude and cold mountainous areas of Xinjiang is highly appropriate.

[0057] Table 1. Statistical distribution of glacier types in the Yurungkash River and Karakash River

[0058]

[0059]

[0060] The input data for the VIC-glacier model includes DEM data, vegetation data, soil data, meteorological data, hydrological data, and glacier boundary data. DEM data comes from the SRTM project of Geospatial Data Cloud, with a spatial distribution of 90m×90m. Taking into account factors such as the spatial resolution of precipitation data and underlying surface remote sensing data, as well as operational efficiency, 215 computational grids of 0.08333°×0.08333° were established for the study area. Vegetation data comes from the global 1km land cover data released by the University of Maryland. Based on the proportion of different vegetation types in the grid, the vegetation in the study area is mainly open shrublands and grasslands. Soil data comes from the 5′ resolution soil database provided by FAO. Meteorological data comes from observations at stations of the China Meteorological Administration, including daily minimum temperature, daily maximum temperature, and daily average wind speed. The inverse distance weighted interpolation method was used to calculate the meteorological data within the unit grid. Hydrological data comes from the daily runoff data of the Tongguziluoke hydrological station and the Uruwati hydrological station from 2010 to 2018 provided by the Tarim River Basin Authority. Glacier boundary data comes from the Second Glacier Inventory dataset.

[0061] Remote sensing precipitation products, reanalysis precipitation data and other precipitation products can provide high spatial and temporal resolution and continuous distribution of precipitation information, and are an important data source for precipitation information in areas with lack of data. Due to the sparse distribution of stations in alpine mountainous areas and the fact that most of them are located in valley areas, the precipitation monitoring capacity is weak, and the error of grid precipitation obtained by traditional hydrological station data interpolation is large, which cannot accurately reflect the water resource cycle process in inland river basin. Therefore, the present application considers using GPM precipitation product as the meteorological driving data of VIC model. GPM is a new generation of precipitation product after TRMM, which not only inherits the method of TRMM precipitation data, but also improves the observation accuracy, spatial and temporal resolution and detection ability. IMERG (Integrated Multi-satellite Retrievals for GPM) is a three-level product of GPM, which can provide global 0.1°, 1d precipitation data, including three products Early-Run, Late-Run and Final-Run. The version used in the present application is IMERG V6 Final-Run, which can provide daily precipitation data from June 2000 to now. The data set integrates various passive microwave and infrared observation data, and has been proved to perform well in mountainous areas. In order to improve the local accuracy, the quantile mapping method is used to correct the bias of global precipitation measurement (GPM) data by using the observation data of regional ground weather stations. The corrected gridded data set is directly applied to the VIC-glacier model. This method helps to retain the original spatial heterogeneity and can represent the elevation-dependent precipitation pattern, especially suitable for the glacier supply source area with limited ground weather station coverage. By incorporating high-resolution gridded input data and elevation sensitivity correction, the model can better represent the spatial variability of precipitation in this complex terrain basin.

[0062] Figure 2 The flowchart shows the basic technical route of the present application. The flowchart shows that starting from the glacier coverage area, vegetation, digital elevation model (DEM) and climate related data, the model is gradually established, optimized and corrected by machine learning method, and finally the precise prediction or simulation of the runoff of the basin is realized.

[0063] Firstly, the percentage of glacier coverage, vegetation coverage, digital elevation model (DEM) data, soil parameter data and temperature gradient are important input parameters in the whole study. The percentage of glacier coverage is used to measure the distribution of glaciers in the study area; while the vegetation coverage may help to study the response of the ecosystem or the relationship with glacier melting; the digital elevation model (DEM) provides important basic data for the study of terrain, combined with elevation data, the distribution of glaciers in different elevation zones can be further analyzed; the temperature gradient directly affects the dynamic change of the glacier over time, so it is also an indispensable parameter in the model. Among them, the precipitation parameter which has the greatest impact on climate data, selects GPM precipitation product as the driver, which can effectively reflect the water cycle process in inland river basin. Based on these data, the VIC model and degree-day model are coupled together, and the automatic calibration method of SCE-UA is used to optimize the model. On this basis, the residual sequence of the coupled model simulation runoff and observation value is extracted, the LSTM neural network is introduced to train the spatio-temporal nonlinear characteristics of the residual, the Adam optimizer and the early stopping strategy are used to prevent overfitting, and the predicted residual is dynamically superimposed on the mechanism model output to realize the correction of runoff; In order to verify the correction advantage of LSTM, the autoregressive model AR(p) is constructed as a comparison model, and the performance difference of the two types of correction models is compared quantitatively from the NSE, MAE, RMSE and other indicators, and the performance of the VIC-glacier model in the alpine mountainous area is improved.

[0064] The VIC variable infiltration capacity curve model is a spatially distributed hydrological model with strong physical mechanism based on the SVATS idea, which is easy to couple with meteorological model and considers the interaction of climate, terrain, soil properties and vegetation. The VIC model is divided into upper, middle and deep soil layers through soil layering and vegetation cover information to respond to short-term precipitation events and long-term hydrological balance. Its energy and water balance mechanism considers evapotranspiration and surface energy exchange, realizes the double constraint of water and energy, and improves the accuracy of the model. The model subdivides the evapotranspiration process into transpiration, surface evaporation and soil water evaporation, and the water use characteristics of different vegetation types are included in the calculation, so that the VIC model is more close to the actual situation in the simulation of water flow.

[0065] The application of VIC model includes land surface model operation and confluence module operation, and the appropriate confluence model is coupled with VIC land surface model according to the research needs to analyze the regional adaptability of the model. The confluence module developed by Lohmann et al. is used to confluence the grid runoff output by VIC land surface model to the grid control hydrological station at the outlet of the basin, and the confluence calculation is completed. The confluence process diagram of VIC model is as follows Figure 3 In the confluence module, the D8 algorithm is used to process the water flow direction of the basin grid.

[0066] In view that the VIC model does not contain a calculation module of the glacier ablation process, and the applicability in the glacier supply basin is not high, the degree-day factor model is introduced to describe the glacier ablation process, which is coupled with the VIC model to construct the VIC-glacier model. The traditional glacier hydrological model often relies on the simulation of complex physical processes, but the data requirement is large, and it is difficult to apply in the high mountain area with complex terrain and the area with scarce data. In order to obtain reliable melt water estimation in the case of lacking complex meteorological data or ground observation, the researchers proposed a simplified model based on air temperature data, namely the degree-day factor model. The DDFM establishes a linear relationship between the cumulative air temperature and the melt water amount by introducing the degree-day factor parameter, so as to realize the rapid and effective estimation of the melt water amount in the environment with scarce data. Coupling the two models together can effectively combine the climate data processing capability of the VIC model and the glacier melting prediction capability of the degree-day model, so as to realize the dynamic and accurate simulation and prediction of the glacier module of the runoff. The basic mechanism of the degree-day factor model is based on the positive correlation between temperature and snowmelt. Temperature rise will increase the melting rate of ice and snow, thereby producing more melt water. The DDFM takes the cumulative temperature of air temperature higher than a certain threshold as the melting driving force, and the degree-day factor quantitatively correlates the cumulative temperature with the actual melt water amount. The calculation of the glacier ablation amount is shown in formula (1) and (2).

[0067] M = DDF x PDD (1);

[0068]

[0069] Where M is the glacier ablation water equivalent; PDD is the positive accumulated temperature of the period, DDF is the degree-day factor; T t is the daily average temperature of day t, H t is a logical variable, which is 1 when T t > 0, and 0 when T t ≤ 0.

[0070] In the VIC-glacier model framework, the glacier melt water generated by the temperature index method is added to the surface runoff component of the VIC model. Specifically, based on the degree-day factor of ice, the daily melt water amount is calculated by using the gridded air temperature. The melt water amount is then introduced into the surface soil of each model grid cell covered by glaciers as an additional runoff input. In each simulation time step, the melt water amount is simultaneously introduced into the surface and underground runoff components with rainfall and glacier melt water components. This integration mechanism ensures that the glacier melt water directly participates in the generation process of surface and underground runoff, thereby affecting the water balance at the basin scale.

[0071] The SCE-UA algorithm (Shuffled Complex Evolution-University of Arizona) is a global optimization algorithm designed for parameter calibration of complex, multi-parameter, and nonlinear models. Its core idea is to balance global search and local optimization by mixing multiple evolutionary strategies to avoid falling into local optimal solutions. In the VIC model, the SCE-UA algorithm is used to automatically calibrate key parameters (such as soil permeability, saturated hydraulic conductivity, vegetation coverage coefficient, etc.) to improve the simulation accuracy of the model. The algorithm first randomly generates multiple complexes, each containing a possible combination of parameters, then generates new solutions through genetic evolution operations such as crossover and mutation, and periodically mixes the optimal solutions of different complexes, finally realizes convergence in iterations to obtain the optimal solution. Through SCE-UA algorithm calibration, the VIC model can achieve high-precision hydrological simulation under various environmental conditions.

[0072] The global search capability and convergence speed of the SCE-UA automatic optimization algorithm can be used to find the optimal solution, while the experience and understanding of the trial-and-error method can help determine the appropriate initial value of the parameters, thereby accelerating the convergence speed of the algorithm. To ensure the accuracy and simulation efficiency of model calibration, the present invention combines the SCE-UA automatic optimization algorithm with the trial-and-error method to calibrate the VIC model. First, the trial-and-error method is used to adjust the initial value of the model parameters; then, the SCE-UA algorithm is used for automatic optimization calibration to minimize the error between the model output and the actual observation data. During the iterative calculation process of the SCE-UA algorithm, the initial value determined by the trial-and-error method is adjusted appropriately according to the results to improve the parameter accuracy. Table 2 provides a detailed description of the model parameter calibration.

[0073] Table 2 Parameters required for VIC-glacier model calibration and manual adjustment range

[0074]

[0075] The present application adopts multiple performance indicators to evaluate the accuracy and robustness of simulated runoff, including Nash-Sutcliffe efficiency coefficient (NSE), mean absolute error (MAE), root mean square error (RMSE), Kling-Gupta efficiency coefficient (KGE), median absolute error (MdAE) and maximum absolute error (MaxAE). Among them, the Nash-Sutcliffe efficiency coefficient (NSE) reflects the overall performance of the model; the root mean square error (RMSE) and the mean absolute error (MAE) are used to evaluate the error size of the whole time series - the root mean square error (RMSE) measures the deviation between the observed value and the simulated value, and the mean absolute error (MAE) quantifies the average amplitude of absolute error, both of which are sensitive to extreme values. The Kling-Gupta efficiency coefficient (KGE) can capture the consistency of simulation and observation time series in trend, magnitude and variability by considering correlation, bias ratio and variability ratio at the same time. The median absolute error (MdAE) is based on the median of absolute error, which can provide a robust typical model performance measure and is less affected by outliers; while the maximum absolute error (MaxAE) represents the maximum single error, reflecting the ability of the model to deal with extreme conditions. The calculation formula is as follows:

[0076] The calculation formula of Nash-Sutcliffe efficiency coefficient is as follows:

[0077]

[0078] The calculation formula of mean absolute error is as follows:

[0079]

[0080] The calculation formula of root mean square error is as follows:

[0081]

[0082] The calculation formula of Kling-Gupta efficiency coefficient is as follows:

[0083]

[0084] The calculation formula of median absolute error is as follows:

[0085] MdAE = median (|Q i,s -Q i,o |) (7);

[0086] The calculation formula of maximum absolute error is as follows:

[0087] MaxAE = max (|Q i,s -Q i,o |) (8);

[0088] In the formula: Q i,oQ i,s is the observed runoff series data; is the observed monthly and daily average surface runoff; r is the Pearson correlation coefficient between the simulated and observed values, indicating the consistency of the time pattern; and β is the bias ratio; is the variation rate, which is used to quantify the relative difference of the average value, and CV represents the coefficient of variation.

[0089] The existing watershed hydrological model, which is usually derived from the model parameter values of the measured rainfall and runoff data, can only reflect the average situation of the formation of the watershed runoff, and if it is used to predict the future runoff, the error is inevitable and difficult to control. At the same time, considering the distributed hydrological model of the hydrological mechanism, some simulation errors may be generated due to the structural limitations of the model itself or the complexity of the watershed characteristics, therefore, it is necessary to use error correction techniques to correct the prediction error. The basic principle of error correction is based on the autocorrelation of runoff error, and the known simulation error is used to correct the future error, and the corrected error is used to further correct the simulated runoff at the same time, so as to correct the system error of the physical model. The simulation results of the VIC-glacier model are corrected by using the LSTM model and the AR model, and the correction accuracy of the two models is compared and verified.

[0090] The runoff time series has autocorrelation, and the simulation error at time t can be predicted according to the error at each time, so as to realize the real-time correction of the simulated runoff at time t. The AR model has the core characteristics of linearity, autoregression and low complexity, and reveals the internal law of the sequence through linear combination of historical data, and has explanation and calculation efficiency, but is limited by the linear hypothesis and stationary requirement. In the present application, the simulation error of the VIC-glacier model is corrected by the autoregressive (AR) method. According to the time sequence change, the difference between the simulated runoff and the observed runoff of the VIC-glacier model constitutes a time series X t ,

[0091] X t = Q sim,t -Q obs,t (9)

[0092] In the formula: X t is the runoff simulation error time series, which should meet the stationary and non-white noise premise; Q sim,t and Q obs,t are the simulated and observed runoff, respectively.

[0093] Firstly, the runoff simulation error X t is subjected to Ljung-Box white noise test, and if X tThe fact that the data is not white noise indicates a correlation in the sequences, allowing for further analysis. The Augment Dickey-Fuller (ADF) test was used to analyze X. t Perform a stationarity test; if X t If the sequence is not stationary, the difference method is used to stabilize the data. The stationary X sequence is not white noise. t Follows the AR(p) model:

[0094]

[0095] In the formula: c is a constant; p is the order of the autoregressive model; For model parameters; ε t The noise is white noise with a mean of 0 and a variance of c. Considering both model fit and complexity, the optimal order p is determined by the Bayesian Information Criterion (BIC).

[0096] The predicted runoff after error correction is as follows:

[0097] Q′ sim,t =Q sim,t -X t (11)

[0098] In the formula Q′ sim,t This is the predicted runoff after error correction.

[0099] In the AR model settings, the same rate period as the VIC model was selected as the training period (January 2010 - December 2015), and the validation period was selected as the testing period (January 2016 - December 2018). Three forecast periods of 1 day, 3 days, and 6 days were set for runoff forecasting.

[0100] Existing watershed hydrological models are typically constructed based on the inversion of measured rainfall and runoff data, reflecting only the average runoff generation situation of the watershed. If used for future runoff forecasting, errors are not only unavoidable but also difficult to control. Therefore, employing error correction techniques to correct forecast errors is a necessary remedial measure. The greatest advantage of using LSTM models lies in their powerful time series modeling capabilities, enabling them to capture long-term dependencies and complex nonlinear characteristics in historical information, resulting in more accurate and robust error correction. This is particularly applicable in simulation and forecasting systems. By learning error patterns in historical flow data through LSTM models, error correction can be applied to simulation results, thereby improving the overall simulation accuracy of the VIC-glacier model. This invention employs the LSTM error correction method, using the residual sequence of measured and model-simulated values, along with precipitation data, as input to simulate the residual sequence. Figure 4 This is a schematic diagram of the internal structure of an LSTM cell.

[0101] In LSTM error prediction, the transmission of information between neural networks is controlled by gating. This allows it to not only memorize past information but also, relying on its core block structure, determine the usefulness of information, selectively forgetting useless details. Through this filtering, the model is trained using existing forecast data and flow errors to establish a mathematical relationship between prediction errors and measured rainfall, predicted flow, and other data, thereby reducing prediction errors and improving accuracy. Taking a model that uses current and previous time-to-time values ​​to predict the next time-to-time value as an example, its error prediction structure is as follows: Figure 5 As shown, the functional relationships between the corresponding error prediction variables are as follows:

[0102] ε t+k =f{P,Q,ε t+k-1 ,ε t+k-2} (12);

[0103] In the formula, ε t ε represents the flood forecast error at the current moment. t-1 ε represents the flood forecast error at the previous moment; P is the rainfall at the corresponding moment; Q is the previous flood forecast discharge at the corresponding moment; ε t+k This represents the error in the model's predictions for future times.

[0104] Therefore, this invention introduces a deep learning model, the LSTM model, to learn error patterns in historical traffic data. This allows for error correction of the simulation results, thereby improving the overall simulation accuracy of the VIC-glacier model. For example... Figure 6 As shown, the residual sequence of measured values ​​and model simulation values, along with precipitation data, were used as inputs to simulate the residual sequence. In the model settings, the same rate period as the VIC model was selected as the training period (January 2010 - December 2015), and the validation period was used as the testing period (January 2016 - December 2018). Three forecast periods of 1 day, 3 days, and 6 days were set for runoff forecasting. The LSTM model had 50 hidden layer units, a Rule activation function, and used the Adam optimizer with a learning rate of 0.01 for training, with 200 training epochs.

[0105] The results of the calibration and validation of the VIC-glacier model at the Tongguziluoke and Uluwati hydrological stations are as follows: Figure 7 As shown in the figure. Among them, 2009 was the model warm-up period, 2010-2015 was the calibration period, and 2016-2018 was the validation period.

[0106] from Figure 7As can be seen from Table 3 and Table 4, the VIC-glacier model coupled with the glacier module performs significantly better than the VIC model in simulating the runoff of the two rivers. In the Yulong Kashi River, the VIC-glacier model significantly improves the accuracy and reliability of runoff simulation compared to the VIC model. During the calibration period, the VIC model only achieves an NSE of 0.28, a KGE of -0.54, and MAE, RMSE and MaxAE of up to 82.48 m3 / s, 177.47 m3 / s and 931.53 m3 / s, respectively, indicating poor overall fitting and serious underestimation of peak runoff. In contrast, the NSE of the VIC-glacier model is 0.84, the KGE is 0.89, and the MAE, RMSE and MaxAE are significantly reduced to 31.60 m3 / s, 60.45 m3 / s and 413.25 m3 / s, respectively, which can more accurately capture the characteristics of runoff. During the validation period, the NSE of the VIC model decreased to -0.26, and the KGE decreased to -0.42, with large errors persisting; while the VIC-glacier model maintained an NSE of 0.82 and a KGE of 0.69, with stable error indices, showing strong robustness in runoff simulation and effectively reducing the impact of uncertainty. In the Kekeshe River, the VIC-glacier model also performs well. During the calibration period, the NSE of the VIC model is 0.10, and the KGE is -0.17, with relatively large MAE and RMSE. The NSE of the VIC-glacier model is 0.84, and the KGE is 0.78, with significantly reduced error indices (MAE 27.94 m3 / s, RMSE 41.34 m3 / s), showing excellent simulation performance. During the validation period, the NSE of the VIC model is -0.18, with expanding errors; while the NSE of the VIC-glacier model is 0.76, and the KGE rises to 0.84, with stable error indices, and the overall trend is closer to the actual flow.

[0107] Although the VIC-glacier model performs well overall, there are still deviations in some years. For example, in the calibration period of the Yulong Kashi River, the simulation results slightly underestimate the flood peak; in the validation period, both the flood peak and the minimum flow are underestimated. Similar situations also exist in the Kekeshe River, which may be related to uncertainty factors such as model parameters, low estimation of meteorological factor driving items, etc. However, the VIC-glacier model, by strengthening the mechanism constraint of glacier runoff, is particularly suitable for glacier-dominated rivers. The newly added glacier calculation module significantly improves the performance and robustness of the model, and performs better than the VIC model in the simulation of the two rivers, and can better capture the long-term variation characteristics of the flow.

[0108] Table 3 Daily-scale performance indicators of simulated flow of Yulong Kashi River during calibration (2010-2015) and validation (2016-2018) periods

[0109]

[0110]

[0111] Table 4. Daily-scale performance indicators of simulated flow of the Karakash River during calibration (2010-2015) and validation (2016-2018) periods.

[0112]

[0113] The monthly average glacial runoff, non-glacial runoff, and the proportion of glacial runoff for the two tributaries from 2010 to 2018 are shown in the table below. Figure 8 .Depend on Figure 8 It can be seen that the mechanism of glacier runoff change includes: (1) The annual changes of glacier runoff and non-glacier runoff in both basins show a single peak, and the peak is reached in August; (2) Glaciers begin to melt in May and begin to increase as part of runoff replenishment. The glacier runoff of Yulongkash River shows a steep increase trend, while the glacier runoff of Karakash River gradually increases; Glaciers are widely distributed in Yulongkash River Basin, and the summer meltwater volume is absolutely dominant; Glaciers in Karakash River Basin are less than those in Yulongkash River, and the glacier replenishment runoff is slightly lower than that of Yulongkash River; (3) The proportion of glacier runoff in the two tributaries reaches its maximum in July and August, with Yulongkash River reaching more than 90% and Karakash River reaching 80%, indicating that glacier melting is the core driving force of summer flood peak; (4) Glacier replenishment runoff dominates from May to September, with the glacier contribution rate of Yulongkash River reaching 60% to 90% and the glacier contribution rate of Karakash River reaching 70% to 80%. Runoff in other months is mainly precipitation runoff.

[0114] Determining the contribution of different components to the total flow is a challenge. This invention uses the VIC-glacier coupling model to separate glacial runoff components and quantify their contribution to the total flow. Figure 9The data shows that glacial melt in the upper reaches of Hotan contributes the most in summer, followed by autumn. In spring, glacial and precipitation contributions are roughly equal, mainly due to the melting of snow in mountainous areas as temperatures rise; however, the melting of snowmelt and glaciers is still in its early stages. Summer is the season with the highest runoff, reaching its peak during this period. This phenomenon is closely related to the highest summer temperatures and the maximum melting of glaciers and snow. Besides temperature, complex and variable extreme weather in summer, such as torrential rains, also accelerates glacial melt, leading to a significant increase in water volume. In autumn, as temperatures gradually decrease, runoff begins to decrease significantly relative to summer, as glacial meltwater enters its later stages, and river flow returns to a lower level. Winter runoff is the lowest of the year, and is only supplied by precipitation, which exists in the form of snow. This is mainly due to the cold weather in the high-altitude mountains of Xinjiang, which limits the contribution of precipitation and snowmelt to river flow; most river sources are frozen, causing flow to drop to its lowest level.

[0115] Glacial melt in the Yurungkash River contributes 78.5% to its annual flow, and in the Karakash River, it contributes 62%. This trend is consistent with the overall trend reported by Wang Ning et al. in their flood simulation of three basins in northwestern Tarim River. However, the glacial recharge is slightly underestimated, which may be related to factors such as model parameter settings, precipitation input, or the way glacial dynamic processes are handled. The VIC-glacier model uses the degree-day factor method to calculate glacial ablation, and the glacial ablation process is affected by factors such as temperature, glacier coverage, and glacier exposure. If the degree-day factor is too low or fails to fully reflect regional glacial differences, the contribution of glacial meltwater will be underestimated, leading to an underestimation of the total runoff simulation results. Future research can further reduce this underestimation bias by improving precipitation data input and optimizing glacial ablation algorithms.

[0116] from Figure 7 It can be seen that the VIC-glacier model coupled with the degree-day model has a good agreement with the actual flow process curve, and the simulation accuracy is greatly improved. However, the model also has several instances where the simulation results are not good: there is still a significant underestimation at the peak runoff point, both during the calibration and validation periods; the spring runoff simulation is slightly underestimated in some years; and there is a significant overestimation before the flood season in some years. In order to improve the forecasting capability of the VIC-glacier model in high-altitude and cold mountainous areas, this invention performs error correction on the simulation results. Figure 10Table 5 and Table 6 show the prediction results of the VIC-glacier model combined with LSTM error correction technique for the two rivers, and evaluate the performance of the model at different prediction periods (1 day, 3 days and 6 days). The results show that the coupled model has high short-term prediction accuracy after error correction. Taking the East Branch of the Yulong River as an example, the NSE values of 1-day and 3-day predictions are 0.92 and 0.89 respectively, and the RMSE and MAE values are relatively low. This indicates that the model can effectively capture the runoff trend in a short time range. However, as the prediction period is extended to 6 days, the model performance decreases, with the NSE value dropping to 0.82, and the RMSE and MAE values significantly increasing. This shows that the prediction error will grow in a longer prediction period, especially during the flood peak event, the deviation between the predicted value and the observed value becomes more obvious. This phenomenon is also confirmed in the time series and scatter plots, where the points of long-term prediction deviate more and more from the ideal line of 1:1, especially at the peak of summer runoff, the model tends to produce more conservative predictions. Overall, the VIC-glacier model combined with LSTM error correction technique shows strong short-term prediction ability for the Yulong River, further proving its applicability in high-altitude cold regions.

[0117] Table 5 Performance comparison of LSTM error correction VIC-glacier model in training and testing periods of Yulong River

[0118]

[0119] Table 6 Performance comparison of LSTM error correction VIC-glacier model in training and testing periods of Yarkant River

[0120]

[0121]

[0122] Figure 11The test results of the VIC-glacier model after LSTM error correction and the VIC-glacier model after AR error correction are presented in the training and testing phases at various forecast periods. The results show that the predicted values ​​of all models exhibit a strong correlation with the observed values. Notably, the LSTM error-corrected model performs better, especially in capturing peak flow and characterizing the overall runoff trend, with its simulation results showing a higher degree of agreement with the observed values. In the scatter plot, the data points of the LSTM error-corrected model are more closely clustered near the 1:1 line, indicating a significant improvement in accuracy, particularly under high runoff conditions. Tables 7 and 8 further compare the evaluation indicators of the two models at different forecast periods (T = 1, 3, and 6 days). Taking the Yulongkash River as an example, the tables show that the runoff correction effect of the LSTM error-corrected model is generally better than that of the AR model under different forecast periods. When the foreseeability period is 1 day, the KGE of both the LSTM and AR models is 0.98 during the training period, but the NSE of the LSTM model is 0.99, higher than the 0.97 of the AR model; the MAE is 6.39m. 3 / s, lower than the AR model's 10.70m 3 / s; RMSE is 13.30m 3 / s, lower than the AR model's 23.62m 3 / s. During the testing period, the LSTM model achieved a KGE of 0.91, higher than the AR model's 0.80; and an NSE of 0.92, also higher than the AR model's 0.91, indicating a more significant short-term correction effect. When the forecast period was 3 days, the LSTM model achieved a KGE of 0.97 during the training period, higher than the AR model's 0.94; during the testing period, the LSTM model achieved a KGE of 0.86, higher than the AR model's 0.78, and an NSE of 0.89, higher than the AR model's 0.85, maintaining a clear advantage. When the forecast period was 6 days, the LSTM model achieved a KGE of 0.82 during the testing period, higher than the AR model's 0.66. Overall, across all forecast periods, the LSTM error correction model outperformed the AR model in most metrics, including KGE and NSE. Especially in short-term forecasting, it better maintained the consistency of the statistical characteristics of runoff simulation and reduced errors, highlighting its effectiveness in maintaining robust predictive performance across different time scales.

[0123] Table 7. Performance comparison of LSTM error-corrected VIC-glacier model and AR error-corrected VIC-glacier model during the training and testing periods on the Yulongkash River.

[0124]

[0125]

[0126] Table 8. Performance comparison of LSTM error-corrected VIC-glacier model and AR error-corrected VIC-glacier model in the training and testing periods of the Karakax River

[0127]

[0128] By integrating and comparing the results of the Yulong Karakax River and the Karakax River, it is observed that the LSTM error-corrected model is superior to the AR error-corrected model in multiple prediction periods. Short-term prediction (e.g., T = 1 day or 3 days) reveals the advantage of the model in minimizing errors and capturing flood peaks, which is crucial for timely flood warning and emergency response. For medium and long-term forecasts (T = 6 days), although all models face increased uncertainty, the LSTM error-corrected model maintains relatively high NSE values and low RMSE / MAE, reflecting its strong ability to handle complex and variable conditions in high-altitude cold regions. Figure 12 The monthly average flow comparison between the LSTM error-corrected model and the AR error-corrected model and the actual value shows that the VIC-glacier model without error correction underestimates both low-flow months and high-flow months, while the error-corrected model simulates closer to the actual value in high-flow scenarios. Notably, the LSTM error-corrected model maintains high consistency with the actual value in both high-flow peaks and full-cycle fluctuations, with smaller deviations compared to the LSTM error-corrected model, demonstrating its advantage in flow prediction. In summary, these research results highlight the value of the LSTM error-corrected VIC-glacier model in different hydrological environments in the Tarim River Basin. By effectively capturing the runoff patterns in the high-altitude cold region of the Karakoram Mountains, this correction technique provides reliable decision support for the VIC-galcier model. In addition, the hydrological conditions in high-altitude cold regions are complex and the climate is variable, making it difficult for traditional models to accurately predict flood peak flow and flood risk caused by extreme weather. The VIC-glacier model combined with the LSTM error correction technique has multiple advantages such as capturing long-term correlations, extracting local time features, and optimizing uncertainty estimates, providing more reliable predictions and effectively mitigating the uncertainty of water resources caused by glacier changes. This enables decision-makers to take timely flood prevention measures and more accurately predict future runoff trends.

[0129] The main sources of uncertainty in traditional runoff prediction include data input, model parameters, and model structure. In terms of data input, errors and uncertainties exist in the data measurement and collection process. In terms of model structure, the VIC-glacier runoff prediction model is a simplification and abstraction of the complex runoff process. Different model structures depict the runoff formation and evolution process in different ways, and cannot completely accurately reflect the complex physical process of flood, so there is structural uncertainty. For the LSTM error correction model, the complexity of the model structure, the activation function, and the training algorithm may cause the model to fail to converge or converge to a non-ideal solution, thereby increasing the uncertainty of the prediction results.

[0130] The Bootstrap method, proposed by Efron, a professor of statistics in the United States, is an effective interval prediction method that overcomes the limitations of traditional point prediction, which can only provide deterministic prediction values. This method is simple to calculate and easy to implement, and is widely used to quantify and predict the uncertainty interval related to given estimates.

[0131] Figure 13 The 1-day, 3-day, and 6-day runoff prediction results and 90% confidence intervals based on the Bootstrap model are shown. To quantify the coverage of these intervals, the observed values within each prediction period are analyzed to determine whether they fall within the confidence interval, and the following coverage rates are calculated: 1-day prediction period: more than 90% of the observed values fall within the confidence interval, indicating that the model has high accuracy in short-term prediction and can effectively handle uncertainty to provide stable interval prediction; 3-day prediction period: the coverage rate is about 85%, and the confidence interval is slightly larger, especially during the peak stage of the flood period, reflecting the increase in uncertainty as the prediction period is extended; 6-day prediction period: when the prediction period is extended to 6 days, the coverage rate drops to about 75%, and the confidence interval significantly expands during the extreme event processing of the flood period, indicating that the uncertainty of long-term prediction is greater. As the prediction period extends from 1 day to 6 days, the confidence interval gradually widens. One possible reason is that as the prediction period extends, future runoff changes are influenced by more random factors such as climate variability, precipitation type, and temperature in the high-altitude mountainous region, making it difficult for the model to accurately predict long-term changes, thus widening the prediction interval. Another possible reason is that during the peak stage of the flood period, model errors significantly increase, especially in long-term prediction, where the unpredictability of flood peaks makes it difficult for the model to capture the accurate occurrence time and intensity of these extreme events, leading to the widening of the confidence interval during this period. Future improvements can focus on optimizing the model structure or incorporating more extreme weather data to improve the model's prediction ability for extreme events.

[0132] The runoff components in the alpine mountainous area are diverse, and each runoff source cannot be directly distinguished, so how to quantify the contribution proportion of glacier melting to total runoff and predict the future runoff trend is a hot research problem in the alpine mountainous area at present. Therefore, the present application achieves the purpose of distinguishing the glacier runoff by coupling the VIC model and the glacier degree-day model, and further optimizes the coupled model by using the automatic calibration of SCE-UA and the error correction method of the LSTM model, so as to improve the accuracy of the model, effectively alleviate the uncertainty of water resources caused by glacier change, and more accurately predict the future runoff trend. The results show that:

[0133] (1) The runoff simulation of the two tributaries of the Hetian River upstream describes the seasonal and interannual variation characteristics of the river flow, especially the significant increase of the peak value in summer and the stable dry season flow in winter, and the main driving factors of these trends may be related to climate change;

[0134] (2) The contribution rate of glacier melt water to the annual runoff of Yulong Kashi River and Kalakashi River reaches 78.5% and 62% respectively;

[0135] (3) Glacier ablation starts in May, and the proportion of runoff supply gradually increases, which is consistent with the time when the monthly average temperature reaches the critical temperature of glacier ablation;

[0136] (4) The prediction accuracy of the VIC-glacier model corrected by error is improved in different prediction periods, and the LSTM error correction model is superior to the traditional AR error correction model in multiple prediction periods. In short-term prediction, the NSE of the LSTM correction model in the training period and the test period reaches 0.9, and the MAE and RMSE are significantly reduced compared with the uncorrected VIC-glacier model; although the long-term prediction accuracy is not as good as the short-term prediction, the LSTM error correction model still maintains a high NSE value and a low RMSE and MAE value.

[0137] Correctly analyzing the runoff components is crucial for water resource management in the alpine mountainous area of Xinjiang. With the continuous climate change, predicting the future runoff trend in the alpine mountainous area is particularly important for ensuring the water stability of the local ecosystem, agriculture and local residents in Xinjiang.

[0138] The above only describes the preferred embodiments of the present application and is not intended to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for improving the runoff simulation of alpine mountainous areas by VIC-glacier model combined with LSTM, characterized in that, The steps are as follows: Step one: data collection: the input data includes DEM data, vegetation data, soil data, meteorological data, hydrological data and glacier boundary data; and the GPM precipitation product is used to drive the meteorological data; Step two: establishment of the VIC-glacier coupled model: the degree-day factor model is introduced to quantify the glacier ablation process, and the degree-day factor model is coupled with the VIC model to construct the VIC-glacier model; Step three: parameter calibration of the VIC-glacier model: the SCE-UA automatic optimization algorithm is combined with the artificial trial and error method to calibrate the VIC-glacier model, and the optimal parameter value is obtained; Step four: error correction of the simulation results of the VIC-glacier model by the LSTM model to realize runoff correction.

2. The method of claim 1, wherein the LSTM is used to improve the simulation of alpine mountain runoff in the VIC-glacier model. The DEM data comes from the SRTM plan of the geographic spatial data cloud, and the spatial distribution rate is 90m x 90m; considering the spatial resolution of precipitation data and underlying surface remote sensing data and the running efficiency, 215 0.08333° x 0.08333° calculation grids are established in the study area; the vegetation data is the global 1km land cover data released by the University of Maryland, and the vegetation in the study area is mainly open shrublands and grasslands according to the proportion of different vegetation types in the grid; the soil data comes from the 5' resolution soil database provided by FAO; the meteorological data comes from the station observation of the China Meteorological Administration, including daily minimum temperature, daily maximum temperature and daily average wind speed; the meteorological data in the unit grid is calculated by using the inverse distance weighted interpolation method; the hydrological data comes from the daily runoff data of Tongguzilake hydrological station from 2010 to 2018 and the daily runoff data of Uruwat hydrological station from 2010 to 2018 provided by the Tarim River Basin Management Bureau; the glacier boundary data comes from the second glacier inventory data set.

3. The method of claim 1, wherein the LSTM is combined to improve the VIC-glacier model for simulating runoff in alpine regions. The calculation method of the glacier ablation amount is: M=DDF×PDD (1); Where M is the glacial meltwater equivalent; PDD is the positive accumulated temperature for that period; DDF is the degree-day factor; T t Let H be the average daily temperature on a certain day t. t As a logical variable, when T t When T > 0, the value is 1. t The value is 0 when ≤0.

4. The method of claim 1, wherein the LSTM is combined to improve the VIC-glacier model for simulating runoff in alpine regions. The parameter calibration method of the VIC-glacier model is as follows: firstly, the initial value of the model parameter is adjusted by using the artificial trial and error method; then, the SCE-UA algorithm is used for automatic optimization calibration to minimize the error between the model output and the actual observation data; during the iteration calculation process of the SCE-UA algorithm, the initial value determined by the artificial trial and error method is appropriately adjusted according to the results to improve the parameter precision.

5. The method of claim 4, wherein the LSTM is used to improve the simulation of alpine mountain runoff in the VIC-glacier model. The error correction method of the simulation results of the VIC-glacier model by the LSTM model is as follows: the LSTM error correction method is used to simulate the residual sequence of the measured value and the model simulation value and the precipitation data as the input; specifically, the residual sequence of the simulated runoff of the VIC-glacier model and the observation value is extracted, the LSTM neural network is introduced to train the spatio-temporal nonlinear characteristics of the residual sequence, the Adam optimizer and the early stopping strategy are used to prevent overfitting, and the predicted residual is dynamically superimposed on the output of the VIC-glacier model to realize runoff correction.

6. The method of claim 1-5, wherein the method is characterized in that, A variety of performance indicators are used to evaluate the accuracy and robustness of simulated runoff, including the Nash-Sutcliffe efficiency coefficient (NSE), mean absolute error (MAE), root mean square error (RMSE), Kling-Gupta efficiency coefficient (KGE), median absolute error (MdAE), and maximum absolute error (MaxAE). The calculation formulas are as follows: The calculation formula of the Nash-Sutcliffe efficiency coefficient is as follows: The calculation formula of the mean absolute error is as follows: The calculation formula of the root mean square error is as follows: The calculation formula of the Kling-Gupta efficiency coefficient is as follows: The calculation formula of the median absolute error is as follows: MdAE = median(|Q i,s - Q i,o |) (7); The calculation formula of the maximum absolute error is as follows: MaxAE = max(|Q i,s - Q i,o |) (8); where: Q i,o refers to the observed flow series data; Q i,s is the model simulated flow series data; refers to the observed monthly and daily average surface runoff; r is the Pearson correlation coefficient between the simulated and observed values; β is the bias ratio; is the variability rate, CV represents the coefficient of variation.

7. The method of claim 5, wherein the LSTM is used to improve the simulation of alpine mountain runoff in the VIC-glacier model. The expression of the LSTM error prediction is as follows: ε t+k = f(P, Q, ε t+k-1 , ε t+k-2 ) (12); wherein ε t is the flood forecast error at the current time, ε t-1 denotes the flood forecast error at the previous time; P is the rainfall at the corresponding time; Q is the previous flood forecast flow at the corresponding time; ε t+k denotes the error predicted by the model at the future time.

Images