Variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution
Through Bayesian model and multi-dimensional joint distribution method, a dynamic vegetation model and Bayesian hierarchical model were established, and the Copula function was used to simulate the impact of vegetation parameters on hydrological parameters, which solved the problem of insufficient simulation accuracy of existing hydrological models in vegetation changing areas. High-precision hydrological simulation and future water resource prediction are achieved.
Patent Information
- Application Number
- CN202411312990.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-20
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-09-20
AI Technical Summary
It is difficult for existing hydrological models to accurately simulate the dynamic changes of hydrological parameters in areas where vegetation changes significantly, and the complex relationship between vegetation and hydrological parameters is difficult to describe, resulting in insufficient simulation accuracy.
Using the Bayesian model and multi-dimensional joint distribution variable parameter hydrological simulation method, by collecting and sorting hydrological, climate, topography, and vegetation data, dynamic vegetation model and Bayesian hierarchical model are established, and the Copula function is used to simulate vegetation parameters for hydrological parameters. The influence of hydrological model parameters can be realized.
The long-term hydrological simulation accuracy in changing environments is improved, and the impact of vegetation changes on hydrological parameters can be described more accurately, which enhances the adaptability and prediction ability of the simulation.
Smart Images

Figure CN118839621B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of hydrological models, and in particular relates to a variable parameter hydrological simulation method based on a Bayesian model and multidimensional joint distribution. Background Art
[0002] Vegetation and hydrological elements are very sensitive to changes in environmental factors such as climate and underlying surface. In particular, water conservation and ecological projects such as returning farmland to forest, returning pasture to grassland, terraces, and silt dams have had a dramatic impact on vegetation and hydrology. Most existing hydrological models use static model parameters and are mainly suitable for stable environments. However, for areas where vegetation and hydrological conditions have changed significantly, it is urgent to carry out ecohydrological simulation research based on dynamic parameters.
[0003] Regarding the hydrological simulation problem under non-steady-state environment, the current approach is mainly from two angles: improving the model structure and improving the parameter estimation method. For example, in the literature 1 (Process-based characterization of evapotranspiration sources over the North American monsoon region [J]. Water Resources Research, 2016, 52 (1): 358-384), the introduction of vegetation coverage improved the evapotranspiration module of the hydrological model and improved the description of vegetation in the VIC model. However, in the past hydrological simulation, vegetation change information was rarely included in the model, and static vegetation parameters were mostly used as the default module for calculation. Reference 2 (Study on the influence of LAI and FVC vegetation parameters on soil moisture simulation of VIC model [J]. Journal of China Institute of Water Resources and Hydropower Research, 2018, 16(02): 141-148+155) simulated the soil moisture content of the region using remote sensing leaf area and coverage dynamic data. However, the study area of this document is Nagqu, Tibet, where vegetation changes are not as significant as those in the Loess Plateau. In addition, the data time is short (2008-2011), making it difficult to express the impact of vegetation changes on hydrological parameters. Reference 3 (Analysis of time-varying characteristics of hydrological model parameters considering the influence of vegetation and human activities [J]. Advances in Water Science, 2018, 29(05): 625-635) sets the time-varying parameters of the hydrological model to the polynomial form of several explanatory variables. However, this is only an ideal generalization. The relationship between vegetation and hydrological parameters is often difficult to describe using algebraic relationships. In addition, the current parameter estimation methods mainly include two categories: optimization algorithms and probabilistic algorithms. Among them, the probabilistic algorithm is represented by the Bayesian algorithm that takes uncertainty into account. By combining the prior information of the parameters with the likelihood function, the posterior distribution of the evolutionary parameters can be continuously updated. Reference 4 (Distributed Hydrological Model Parameter Estimation and Uncertainty Research [D]. Nanjing University, 2018) constructed a Bayesian differential evolution algorithm to estimate hydrological parameters. At present, the Bayesian method is mostly used to directly estimate the parameters of the hydrological model, and is less used to estimate the nonlinear relationship between vegetation parameters and hydrological parameters.
[0004] In today's vegetation restoration, it is necessary to develop variable-parameter hydrological simulation methods that fully consider the dynamic changes of vegetation, so that hydrological parameters can change with the state of vegetation restoration, in order to improve the adaptability of hydrological simulation methods under vegetation changes, enhance the level of simulation of eco-hydrological processes under changing environments, and provide new technical support for quantifying the relationship between vegetation and hydrological parameters. Summary of the invention
[0005] In order to address the shortcomings of existing hydrological simulations, such as ignoring information on dynamic changes in vegetation, designing most model parameters as static, and making it difficult to describe the complex relationship between vegetation and hydrological parameters, the present invention uses multi-source remote sensing vegetation information to construct a hydrological model that takes dynamic vegetation into consideration in areas where vegetation changes significantly. This model fully incorporates multi-dimensional vegetation information into hydrological simulations in order to accurately characterize the eco-hydrological process of the watershed, and then establishes prior distribution, Bayesian hierarchical model, high-dimensional joint distribution and other methods to simulate the influence of vegetation parameters on hydrological parameters. The accuracy of long-term hydrological simulation under a changing environment is improved by dynamically adaptive adjustment of hydrological parameters to vegetation parameters.
[0006] The present invention adopts the following technical solution:
[0007] The variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution includes the following steps:
[0008] Step 1: Collect hydrological, climate, topographic, vegetation, and soil data in the study basin, organize site data in time series, and clip spatial data according to the basin vector range;
[0009] Step 2: Based on the vegetation data collected in step 1, prepare annual dynamic vegetation parameter files, establish dynamic vegetation data for the VIC distributed hydrological model, and replace the vegetation parameters in the hydrological model year by year to construct a hydrological model that takes vegetation dynamics into account;
[0010] Step 3: Establish a Bayesian hierarchical model to estimate the dynamic hydrological parameters of the hydrological model, that is, estimate the relationship between vegetation parameters and hydrological model parameters through the Bayesian hierarchical model, and use this relationship to deduce the hydrological model parameters under different vegetation states, thereby driving the hydrological model to perform hydrological simulation;
[0011] Step 4: Calibrate the hydrological model in sections and obtain the hydrological model parameters for each period to obtain the hyper-prior distribution of the relationship between vegetation parameters and hydrological model parameters. The long-term climate data compiled in step 1 and the long-term vegetation data produced in step 2 are divided into multiple sub-periods according to a certain length. The VIC distributed hydrological model is driven and calibrated separately in each sub-period to obtain the hydrological model parameters in each sub-period.
[0012] Step 5: Using the hydrological model parameters of each sub-period obtained in step 4, establish the Copula joint distribution function between climate, vegetation and hydrological model parameters to simulate the impact of climate and vegetation on hydrological model parameters, and use it as the hyper-prior information of the hierarchical model in step 3;
[0013] Step 6: Use the Copula joint distribution function constructed in step 5 as the hyper-prior information in step 3 to improve the Bayesian hierarchical model, build a process model and a hierarchical parameter model; combine the climate data in step 1, the vegetation data in step 2 and the improved Bayesian hierarchical model to establish a conditional distribution to solve the hydrological model parameters in different periods with vegetation changes, and then drive the hydrological model considering vegetation dynamics in step 2 to perform hydrological simulation, set the number of iterations, update the parameters multiple times, and obtain the calibration results.
[0014] Preferably, the vegetation data in step 2 include annual land cover types, monthly leaf area LAI data and vegetation coverage FVC data. The calculation grid units of the hydrological model are divided with a certain spatial resolution, and the proportion of land cover types on each grid is counted in turn. Then, the annual changes in LAI and FVC of various land cover types in each grid are counted, and annual vegetation parameter files and vegetation parameter libraries are generated, thereby establishing dynamic vegetation input for the hydrological model.
[0015] Preferably, in step three, the Bayesian hierarchical model has four layers, the first layer is a data model, the second layer is a process model, and the third and fourth layers are parameter models; the Bayesian hierarchical model includes sample information, likelihood function, hyper-prior distribution, hyper-prior parameters, prior distribution, prior parameters and posterior distribution; sample information refers to measured climate, vegetation and hydrological data; likelihood function refers to the probability of reproducing observed runoff when a certain set of parameters is used to simulate runoff values; hyper-prior distribution and hyper-prior parameters are used to describe the complex relationship between vegetation parameters and hydrological model parameters; prior distribution and prior parameters determine the hydrological model parameters that change with vegetation based on hyper-prior information; posterior distribution is the model parameters updated by prior information and likelihood function.
[0016] Preferably, the Copula joint distribution function established in step 5 is as follows:
[0017] Assume that F is an n-dimensional random variable (X1,...,X n ), and the corresponding marginal distributions are u1,u2,…,u n , then there exists an n-dimensional Copula function such that the subset of n-dimensional random variables X1,...,Xn ( x 1 ,x 2 ,..., x n )∈[-∞,+∞] n ,have:
[0018]
[0019] To solve high-dimensional Copula functions above three dimensions, the Vine method is often used. For an n-dimensional Vine Copula statistical model, there are n-1 trees in total, and the node set is N = {N1, N2, N3, …, N n-1}, edge set E = {E1,E2,E3,...,E n-1}, edge e in E = j(e), k(e) | D(e), where j(e), k(e) are the two nodes connected to edge e, and D(e) is the condition set. Then the Copula density function corresponding to edge e is expressed as c j(e),k(e)|D(e) . Let n-dimensional random variables (X1,...,X n ), use X D(e) represents the subvector determined by the condition set D(e), and the marginal density function of the i-th random variable is f i , then the joint density function can be expressed as:
[0020]
[0021] Specifically, the process model and hierarchical parameter model constructed in step six are as follows:
[0022]
[0023] In the formula, Q sim is the simulated runoff volume, f VIC 、f copula1 、f copula2 They represent the process model and the two constructed Copula function relationships respectively. are all parameters of the VIC distributed hydrological model, B represents the infiltration capacity shape coefficient, d2 represents the thickness of the second soil layer, D smax It indicates the maximum value of base flow generated by the bottom soil in one day. D s Indicates the proportion of nonlinear growth of base flow when it occurs D smax The proportion of W s It indicates the ratio of the bottom soil moisture content to the maximum soil moisture content when nonlinear growth of base flow occurs. d 3 represents the thickness of the third soil layer; P represents rainfall, T represents temperature, and LAI represents leaf area; is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and B, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and B, is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and d2, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and d2, is the likelihood function parameter.
[0024] Furthermore, the parameters of the hydrological model constructed in step 6 are modeled as follows:
[0025]
[0026] in, Set it to uniform distribution U, and the remaining parameters to normal distribution N; For the i The mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean of the likelihood function parameters; For the i The variance of the likelihood function parameter.
[0027] Furthermore, in step six, the Monte Carlo sampling method based on Markov chain and the Metropolis-Hastings method are used to perform high-dimensional sampling and hierarchical model posterior probability distribution inference, establish conditional distribution, and solve the hydrological model parameters in different periods with vegetation changes, thereby driving the dynamic vegetation VIC model in step two to perform variable parameter hydrological simulation, calibrate the parameters on the climate and vegetation data of the year and simulate the runoff of the year, and splice the simulation results in time to obtain long-term hydrological simulation results; set the number of iterations in the hierarchical Bayesian model, use the log-likelihood function of the regression model with correlation, heteroscedasticity and non-Gaussian error to discriminate the likelihood, calibrate and update the parameters multiple times to obtain the calibration results.
[0028] The beneficial effects of the present invention are:
[0029] The present invention overcomes the problem that the existing hydrological simulations rarely consider vegetation change information, and mostly use default unchanged vegetation parameters for hydrological simulations. The dynamic changes of vegetation caused by natural changes and strong human activities in a changing environment are not adequately generalized, and the model parameters are mostly designed to be fixed and unchanged, which is difficult to meet the non-steady-state status of the basin. In the variable parameter research method, the complex relationship between vegetation and hydrological parameters is difficult to describe. It is often necessary to initially set simple parameter change scenarios based on experience to calibrate the hydrological parameters. It is difficult to update and correct the parameter change plan during the calibration process, and there is a certain subjective error problem.
[0030] The present invention uses Copula function to establish a complex multi-dimensional joint relationship among precipitation, temperature, vegetation, and hydrological model parameters, and then updates the Copula function parameters through the Bayesian hierarchical model iteration, updates the relationship between precipitation, temperature, vegetation, and hydrological model parameters, and then updates the hydrological model parameter values under climate and vegetation changes, and finally obtains a hydrological simulation with high long-term simulation accuracy under a changing environment. The dynamic adaptive adjustment of hydrological model parameters to vegetation parameters is realized, which has certain theoretical value and practical significance for improving the accuracy of long-term hydrological simulation under a changing environment and the prediction of future water resources, and provides a new solution for adapting to hydrological simulation under vegetation changes. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 The graphs of the leaf area index LAI and vegetation coverage FVC over time in Example 1, wherein (a) is the evolution of LAI and (b) is the evolution of FVC;
[0032] Figure 2: This is a comparison between the default vegetation parameter library and the remote sensing dynamic vegetation in Example 1, wherein (a) is the LAI in the default vegetation parameter library, (b) is the FVC in the default vegetation parameter library, (c) is the remote sensing LAI data extracted from the evergreen broad-leaved forest, (d) is the remote sensing FVC data extracted from the evergreen broad-leaved forest, (e) is the remote sensing LAI data extracted from the cultivated land, and (f) is the remote sensing FVC data extracted from the cultivated land;
[0033] Figure 3 The Bayesian hierarchical model structure diagram in Example 1;
[0034] Figure 4 It is a contour map of vegetation parameters and hydrological parameters in the Yanhe River Basin in Example 1, wherein (a) is a relationship diagram of model parameters d2, P and LAI, (b) is a relationship diagram of model parameters d2, T and LAI, (c) is a relationship diagram of model parameter d2, cultivated land area ratio and forest and grassland area ratio, (d) is a relationship diagram of model parameters B, P and LAI, (e) is a relationship diagram of model parameters B, T and LAI, (f) is a relationship diagram of model parameter B, cultivated land area ratio and forest and grassland area ratio;
[0035] Figure 5 It is a parameter structure diagram of the Bayesian hierarchical model of the Yanhe River Basin involved in Example 1;
[0036] Figure 6 is the runoff simulation result of Example 1;
[0037] Figure 7 The verification results of evapotranspiration calculation in Example 1, where (a) is the comparison between the simulation results and the PML-V2 product, and (b) is the comparison between the simulation results and the GLEAM product. DETAILED DESCRIPTION
[0038] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.
[0039] Embodiment 1:
[0040] The Yanhe River Basin is located in the middle reaches of the Yellow River Basin, with a total area of about 5,891 square kilometers. The basin control station is the Ganguyi Hydrological Station. The Yanhe River Basin is located in a semi-arid and semi-humid area, with uneven rainfall distribution throughout the year. The vegetation types in the basin are mainly cultivated land, forest land, and grassland. Since 2000, the vegetation has significantly turned green. It is a typical area for returning farmland to forest (grass), such as Figure 1 From 1980 to 2020, the leaf area index LAI and vegetation coverage FVC in the Yanhe River Basin showed a significant increasing trend. The changes were relatively slow before 2000, and the growth rate was obvious after 2000, with LAI increasing by 130% and FVC increasing by 38%.
[0041] The specific method of this embodiment includes:
[0042] Step 1: Collect time series data such as runoff, rainfall, maximum temperature, minimum temperature, leaf area LAI, vegetation coverage FVC, land cover type, and remote sensing evapotranspiration in the Yanhe River Basin. The sequence length is 1980-2020. The station data are sorted according to the time series, and the spatial data are clipped according to the basin vector range; in addition, collect and clip time-invariant data such as digital elevation DEM and soil texture of the study area.
[0043] Step 2: Based on the vegetation data collected in step 1, including annual land cover types, monthly leaf area LAI data and vegetation coverage FVC data, the calculation grid units of the VIC distributed hydrological model are divided with a spatial resolution of 0.25°, and the proportion of land cover types on each grid is counted in turn. Then, the annual changes in LAI and FVC of various land cover types in each grid are counted, and annual vegetation parameter files and vegetation parameter libraries are generated. In this way, the dynamic vegetation input of the VIC distributed hydrological model is established, and the vegetation parameters in the model are replaced year by year, so as to construct a hydrological model that takes vegetation dynamics into consideration.
[0044] Figure 2 is a comparison between the default vegetation parameter library and the remote sensing dynamic vegetation in Example 1, where Figure 2 (a) in the figure is the LAI in the default vegetation parameter library. Figure 2 (b) is the FVC in the default vegetation parameter library. Figure 2 (c) is the remote sensing LAI data extracted from the evergreen broad-leaved forest. Figure 2 (d) in the figure is the remote sensing FVC data extracted from the evergreen broad-leaved forest. Figure 2 (e) in the figure is the remote sensing LAI data extracted from cultivated land. Figure 2 (f) in the figure is the remote sensing FVC data extracted from cultivated land. It can be seen that the default vegetation parameter library of VIC is very different from the vegetation information extracted based on remote sensing. First, the default vegetation parameter library assumes that the values of LAI for different vegetation types are the same every month and every year, because the default vegetation parameter library counts the LAI for different vegetation types on a global scale, ignoring the spatial heterogeneity of vegetation and its temporal variation. Secondly, the default vegetation parameter library assumes that the FVC of each vegetation is constant at 1, but in reality the vegetation coverage is often between 0-1 and varies over time. Therefore, Figure 2 The evidence shows that within the specific study area, the spatiotemporal information of vegetation described by the default vegetation parameter library cannot reflect the actual basin vegetation, while remote sensing data can take the spatiotemporal heterogeneity of vegetation into account in the model.
[0045] Step 3: Establish a four-layer Bayesian hierarchical model. The first layer is the data model, the second layer is the process model, and the third and fourth layers are parameter models. The Bayesian hierarchical model includes sample information, likelihood function, hyper-prior distribution, hyper-prior parameters, prior distribution, prior parameters, posterior distribution, etc. Figure 3 As shown. Sample information refers to the measured climate, vegetation and hydrological data; likelihood function refers to the probability of reproducing the observed runoff when a certain set of parameters is used to simulate the runoff value; hyper-prior distribution and hyper-prior parameters are used to describe the complex relationship between vegetation parameters and hydrological model parameters; prior distribution and prior parameters can determine the hydrological model parameters that change with vegetation based on hyper-prior information; posterior distribution is the model parameters after the prior information and likelihood function are updated. The structure of the Bayesian hierarchical model is shown in the formula:
[0046]
[0047] In the formula, p represents the conditional probability distribution, Q, η, θ 1 ,θ 2 They represent data, process, prior parameters, and hyper-prior parameters respectively. The Bayesian hierarchical model is used to estimate the dynamic hydrological parameters of the VIC distributed hydrological model, that is, the relationship between vegetation parameters and hydrological model parameters is estimated through the Bayesian hierarchical model, and the hydrological model parameters under different vegetation states are deduced using this relationship, thereby driving the hydrological model to perform hydrological simulation.
[0048] Step 4: Calibrate the hydrological model in sections, obtain the hydrological model parameters for each period, and obtain the hyper-prior distribution of the relationship between vegetation parameters and hydrological model parameters. The application of the Bayesian hierarchical model requires a high degree of confidence in prior information or hyper-prior information, and obtains certain parameter samples as prior information through the section-wise calibration method. The long-term climate data in step 1 and the long-term vegetation data in step 2 are divided into multiple sub-periods according to a certain length ω, and the hydrological model is driven and calibrated separately in each sub-period to obtain the hydrological model parameters in each sub-period.
[0049] In this embodiment, the long time series data from 1980 to 2020 is divided into multiple sub-periods every 5 years. The parameters are calibrated and the runoff of the year is simulated based on the climate and vegetation conditions of each sub-period. The simulation results are spliced in time to obtain the long-term hydrological simulation results.
[0050] Step 5: Using the hydrological model parameters for each time period obtained in step 4, analyze the changing relationship between vegetation parameters and hydrological model parameters, and construct a Copula function.
[0051] Considering that the impact of vegetation change on hydrological model parameters is affected by multiple factors and the relationship is difficult to describe with a simple linear relationship, rainfall P, temperature T, and leaf area LAI were selected as explanatory variables of hydrological model parameters. Two sensitive parameters B and d2 of the VIC distributed hydrological model were selected to analyze the parameter change law. Vegetation coverage FVC is highly correlated with leaf area index LAI, and the consistency of change is very high. Here, only LAI is selected as the explanatory variable.
[0052] In the present invention, two Copula functions are established to simulate the relationship between rainfall P, temperature T, leaf area LAI, two hydrological parameters B and d2. The first is a Copula function constructed by variables P, T, LAI, and B, and the second is a Copula function constructed by variables P, T, LAI, and d2.
[0053] Figure 4 It is a contour map of vegetation parameters and hydrological model parameters in the Yanhe River Basin in Example 1, where Figure 4 (a) is the relationship diagram of model parameters d2, P and LAI. Figure 4 (b) is the relationship diagram of model parameters d2, T and LAI. Figure 4 (c) is the relationship diagram between the model parameter d2, the proportion of cultivated land area and the proportion of forest and grassland area. Figure 4 (d) is the relationship diagram of model parameters B, P and LAI. Figure 4 (e) is the relationship diagram of model parameters B, T and LAI. Figure 4 (f) is the relationship diagram between the model parameter B, the proportion of cultivated land area and the proportion of forest and grassland area. From the contour map of parameter changes, it can be seen that the parameter d2 increases with the increase of LAI and the increase of the proportion of forest and grassland area; the parameter B decreases with the increase of LAI and the increase of the proportion of forest and grassland area. In the VIC distributed hydrological model, the parameter d2 represents the thickness of the second soil layer, B represents the infiltration capacity shape coefficient, d2 is inversely proportional to the runoff, and B is proportional to the runoff. The change of hydrological model parameters with vegetation shows the law that runoff decreases with the increase of vegetation.
[0054] The Copula joint distribution function between climate, vegetation and hydrological model parameters is established as hyper-prior information. Since there is a high degree of nonlinearity and complexity between climate, vegetation and hydrological model parameters, and different elements may not obey the same type of distribution characteristics. Therefore, the Copula joint distribution function that can characterize the correlation and statistical laws between multidimensional variables is used to simulate the impact of climate and vegetation on hydrological model parameters.
[0055] Assume that F is an n-dimensional random variable (X1,...,X n ), the corresponding marginal distributions are u1,u2,…,u n, then there exists an n-dimensional Copula function such that the subset of n-dimensional random variables X1,...,Xn ( x 1 ,x 2 ,...,x n )∈[-∞,+∞] n ,have:
[0056]
[0057] (x1,x2,...,x n ) represents a subset of n-dimensional random variables (X1,...,Xn).
[0058] To solve high-dimensional Copula functions above three dimensions, the Vine method is often used. For an n-dimensional Vinecopula statistical model, there are n-1 trees in total, and the node set is N = {N1, N2, N3, …, N n-1}, edge set E = {E1,E2,E3,...,E n-1}, edge e in E = j(e), h(e)| D(e), where j(e) and h(e) are two nodes connected to edge e, and D(e) is the condition set. Then the Copula density function corresponding to edge e is expressed as c j(e),h(e)|D(e) Each node of the first tree represents a marginal distribution of a random variable, and the other nodes represent binary marginal distributions. n-1 is the node of the n-1th tree, E n-1 It is the edge set of the n-1th tree, and each edge of each tree represents a copula.
[0059] Suppose n-dimensional random variables (X1,...,X n ), use X D(e) represents the subvector determined by the condition set D(e), let k The marginal density function of a random variable is f k , then the joint density function of X can be expressed as:
[0060]
[0061] In the formula, F( x j(e) | x D (e) ), F( x h(e) | x D (e) ) are the conditional distribution functions on the nodes respectively.
[0062] Step 6: Use the Copula function constructed in step 5 as the hyper-prior information in step 3 to improve the Bayesian hierarchical model. The constructed process model and hierarchical parameter model are as follows:
[0063]
[0064] In the formula, Q sim is the simulated runoff volume, f VIC 、f copula1 、f copula2 They represent the process model and the two constructed Copula function relationships respectively. All of them are parameters of the VIC distributed hydrological model. B represents the infiltration capacity shape coefficient, d2 represents the thickness of the second soil layer, and the first soil layer thickness is 0.1 m by default. D smax It indicates the maximum value of base flow generated by the bottom soil in one day. D s Indicates the proportion of nonlinear growth of base flow when it occurs D smax The proportion of W s It indicates the ratio of the bottom soil moisture content to the maximum soil moisture content when nonlinear growth of base flow occurs. d 3 represents the thickness of the third soil layer; P represents rainfall, T represents temperature, and LAI represents leaf area; is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and B, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and B, is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and d2, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and d2, is the likelihood function parameter.
[0065] Figure 5 This is a parameter structure diagram of the Bayesian hierarchical model of the Yanhe River Basin in Example 1. The parameter modeling and settings are as follows:
[0066]
[0067] in, Set it to uniform distribution U, and the remaining parameters to normal distribution N; For the iThe mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean of the likelihood function parameters; For the i The variance of the likelihood function parameter.
[0068] Combined with the climate data in step 1, the vegetation data in step 2 and the constructed parameter model, the Monte Carlo sampling method (MCMC) based on Markov chain and the Metropolis-Hasting algorithm are used for high-dimensional sampling and hierarchical model posterior probability distribution inference, and conditional distribution is established to solve the hydrological model parameters in different periods with vegetation changes, and then drive the dynamic vegetation hydrological model in step 2 to perform variable parameter hydrological simulation. The parameters are calibrated on the climate and vegetation data of the year and the runoff of the year is simulated. The simulation results are spliced in time to obtain the hydrological simulation results of the long time series. The number of iterations is set in the Bayesian hierarchical model, and the log-likelihood function (GL) of the regression model with correlation, heteroscedasticity and non-Gaussian error is used to discriminate the likelihood, and the parameters are continuously updated to obtain the calibration results.
[0069] Reference 5 (Research on Bayesian Probabilistic Hydrological Forecasting Method Based on Copula Function [D], Wuhan University, 2017) establishes a copula function for measured flow and forecast flow, constructs a prior distribution based on the measured flow at the current moment and in the previous period, and then forecasts the flow through the Bayesian method. This method uses the copula function to improve the accuracy of hydrological forecasting. It is essentially a data-driven method that can only be applied to forecasts with a shorter forecast period and can only obtain flow results. It is impossible to obtain the variables of various intermediate states in the entire hydrological calculation process, and it is difficult to explain the reasons for changes in runoff and hydrological processes. The present invention uses Copula to construct the relationship between environmental factors and hydrological model parameters to simulate runoff over a long period of time.
[0070] Reference 6 (Research on the Dynamic Law of Hydrological Model Parameters in a Changing Environment [D], Tsinghua University, 2015) takes the Hanjiang River Basin as the research area, establishes the dynamic relationship between model parameters and environmental factors, and uses it as a dynamic parameter scheme to improve the hydrological model. After screening and selection, a multivariate regression model of hydrological parameters and NDVI is established, and a quadratic function relationship between parameters and NDVI is established. In the method of reference 6, only NDVI is selected to describe vegetation changes by comparing the size of the correlation coefficient. However, the multivariate regression relationship is a very simple relationship. The correlation coefficient only indicates the strength of the linear relationship. The changes in hydrological model parameters are affected by a variety of environmental variables (precipitation, temperature, vegetation) and the relationship between them is very complex. It is difficult to describe the quantitative function. This document uses a segmented calibration method to obtain samples and establish a quadratic function relationship between parameters and NDVI. After the relationship is established, this relationship is used to achieve dynamic simulation. However, once this relationship is determined, it cannot be changed. The accuracy of the simulation depends largely on the reliability of the initial setting of the relationship between NDVI and hydrological parameters, and cannot be adjusted. However, the present invention introduces the Bayesian method on the basis of establishing the complex relationship between the hydrological model parameters and various environmental factors. By continuously updating the likelihood function to obtain a more accurate posterior distribution, the initially set relationship can be corrected to improve the adaptability to the changing environment.
[0071] Embodiment 2:
[0072] Evaluate and verify the simulation results. Evaluate the accuracy of the runoff and evapotranspiration simulation results and compare their performance differences with several previous parameter estimation methods. The accuracy evaluation function of runoff and evapotranspiration simulation is the Nash efficiency coefficient. NSE , water volume error RE and correlation coefficient r , the formula is as follows:
[0073]
[0074] In the formula, t 1 and tn are the start and end times of the runoff or evapotranspiration series, respectively; is the measured runoff or evapotranspiration at time t, Simulated runoff or simulated evapotranspiration at time t; is the average value of measured runoff or evapotranspiration, is the average value of simulated runoff. In this embodiment, there is no large-scale measured observation value of evapotranspiration at present, and the remote sensing product is used instead of the real value as the measured value.
[0075] Figure 6 This is the runoff simulation result of Example 1. Runoff Nash efficiency coefficient under hydrological variable parameter simulation based on Bayesian hierarchical model and multidimensional joint distribution NSE is 0.624, water volume error RE The value is 4.6%, which indicates that the variable parameter method can accurately simulate the change of basin runoff under vegetation change. The runoff simulation mainly depends on NSE and water volume error, and the correlation coefficient is not an important indicator.
[0076] Figure 7 is the verification result of the total evaporation simulation value of Example 1, Figure 7 (a) is the comparison between the simulation results and the PML-V2 product. Figure 7 (b) shows the comparison between the simulation results and the GLEAM product. The simulated soil water has good consistency with the two commonly used products, with a correlation coefficient of 0.94 with the PML-V2 product and a correlation coefficient of 0.97 with the GLEAM product. In terms of quantity, the scatter plot of the simulation results and the PML-V2 product is closer to the 1:1 line, so the simulation results are consistent with both products and are closer to the results of the PML-V2 product, which also shows that the method proposed in this application can achieve good results in simulating runoff and evapotranspiration.
[0077] It should be noted that the above specific implementation methods and examples are preferred implementation methods proposed for the present invention, and cannot be used to limit the protection scope of the present invention. Any equivalent changes or equivalent modifications made on the basis of this technical solution in accordance with the technical ideas proposed by the present invention still fall within the scope of protection of the technical solution of the present invention.
Claims
1. A variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution, characterized in that: The following steps are involved: Step 1: Collect hydrological, climate, topographic, vegetation and soil data in the study basin, organize site data in time series, and clip spatial data according to the basin vector range; Step 2: Based on the vegetation data collected in step 1, prepare annual dynamic vegetation parameter files, establish dynamic vegetation data for the VIC distributed hydrological model, and replace vegetation parameters in the VIC distributed hydrological model year by year to construct a hydrological model that takes vegetation dynamics into account; Step 3: Establish a Bayesian hierarchical model to estimate the dynamic hydrological parameters of the hydrological model, that is, estimate the relationship between vegetation parameters and hydrological model parameters through the Bayesian hierarchical model, and use this relationship to deduce the hydrological model parameters under different vegetation states, thereby driving the hydrological model to perform hydrological simulation; The Bayesian hierarchical model consists of four layers. The first layer is the data model, the second layer is the process model, and the third and fourth layers are parameter models. The Bayesian hierarchical model includes sample information, likelihood function, hyper-prior distribution, hyper-prior parameters, prior distribution, prior parameters and posterior distribution. Sample information refers to the measured climate, vegetation and hydrological data. The likelihood function refers to the probability that the simulated runoff value reproduces the observed runoff when a certain set of parameters is used. Hyper-prior distribution and hyper-prior parameters are used to describe the complex relationship between vegetation parameters and hydrological model parameters. Prior distribution and prior parameters determine the hydrological model parameters that change with vegetation based on hyper-prior information. The posterior distribution is the model parameter updated by the prior information and likelihood function. Step 4: Calibrate the hydrological model in sections, obtain the hydrological model parameters for each period, and obtain the hyper-prior distribution of the relationship between vegetation parameters and hydrological model parameters; divide the long-term climate data compiled in step 1 and the long-term vegetation data produced in step 2 into multiple sub-periods according to a certain length, drive and calibrate the VIC distributed hydrological model separately in each sub-period, and obtain the hydrological model parameters in each sub-period; Step 5: Using the hydrological model parameters of each sub-period obtained in step 4, establish the Copula joint distribution function between climate, vegetation and hydrological model parameters to simulate the impact of climate and vegetation on hydrological model parameters, and use it as the hyper-prior information of the hierarchical model in step 3; Step 6: Use the Copula joint distribution function constructed in step 5 as the hyper-prior information in step 3 to improve the Bayesian hierarchical model, build a process model and a hierarchical parameter model; combine the climate data in step 1, the vegetation data in step 2 and the improved Bayesian hierarchical model to establish a conditional distribution to solve the hydrological model parameters at different periods with vegetation changes, and then drive the hydrological model considering vegetation dynamics in step 2 to perform hydrological simulation, set the number of iterations, update the parameters multiple times, and obtain the calibration results; The constructed process model and hierarchical parameter model are as follows: In the formula, Q sim is the simulated runoff volume, f VIC 、f copula1 、f copula2 They represent the process model and the two constructed Copula function relationships respectively. are all parameters of the VIC distributed hydrological model, B represents the infiltration capacity shape coefficient, d2 represents the thickness of the second soil layer, D smax It indicates the maximum value of base flow generated by the bottom soil in one day. D s Indicates the proportion of nonlinear growth of base flow when it occurs D smax The proportion of W s It indicates the ratio of the bottom soil moisture content to the maximum soil moisture content when nonlinear growth of base flow occurs. d 3 represents the thickness of the third soil layer; P represents rainfall, T represents temperature, and LAI represents leaf area; is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and B, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and B; is the marginal distribution parameter of the Copula function relationship established based on P, T, LAI, and d2, is the Copula function parameter of the Copula function relationship established based on P, T, LAI, and d2, is the likelihood function parameter; The parameters of the constructed hydrological model are as follows: in, Set it to uniform distribution U, and the remaining parameters to normal distribution N; For the i The mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the marginal distribution parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and B; For the i The mean value of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The variance of the Copula function parameters of the Copula function relationship established based on P, T, LAI, and d2; For the i The mean of the likelihood function parameters; For the i The variance of the likelihood function parameter.
2. The variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution according to claim 1 is characterized in that: The vegetation data described in step 2 include annual land cover types, monthly leaf area LAI data and vegetation coverage FVC data. The calculation grid units of the hydrological model are divided with a certain spatial resolution, and the proportion of land cover types on each grid is counted in turn. Then, the annual changes in LAI and FVC of various land cover types in each grid are counted, and annual vegetation parameter files and vegetation parameter libraries are generated, thereby establishing dynamic vegetation input for the hydrological model.
3. The variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution according to claim 1 is characterized in that: The Copula joint distribution function established in step 5 is as follows, assuming that F is an n-dimensional random variable X1,...,X n The joint distribution function of n , then there exists an n-dimensional Copula function such that the subset of n-dimensional random variables X1,...,Xn ( x 1 ,x 2 ,...,x n )∈[-∞,+∞] n ,have: The Vine method is used to solve the high-dimensional scale Copula function above three dimensions. For an n-dimensional Vine Copula statistical model, there are n-1 trees in total, and the node set is N = {N1, N2, N3, ..., N n-1 }, edge set E = {E1,E2,E3,...,E n-1 }, edge e in E = j(e), k(e) | D(e), where j(e), k(e) are the two nodes connected to edge e, and D(e) is the condition set. Then the Copula density function corresponding to edge e is expressed as c j(e),k(e)|D(e) ; Let n-dimensional random variables be X1,...,X n , use X D(e) represents the subvector determined by the condition set D(e), and the marginal density function of the i-th random variable is f i , then the joint density function can be expressed as: .
4. The variable parameter hydrological simulation method based on Bayesian model and multidimensional joint distribution according to claim 1 is characterized in that: In step 6, the Monte Carlo sampling method based on Markov chain and the Metropolis-Hastings method are used to perform high-dimensional sampling and hierarchical model posterior probability distribution inference, establish conditional distribution, solve the hydrological model parameters in different periods with vegetation changes, and then drive the dynamic vegetation hydrological model in step 2 to perform variable parameter hydrological simulation. The parameters are calibrated on the climate and vegetation data of the year and the runoff of the year is simulated. The simulation results are spliced in time to obtain long-term hydrological simulation results. The number of iterations is set in the hierarchical Bayesian model, the log-likelihood function of the regression model with correlation, heteroskedasticity and non-Gaussian error is used to discriminate the likelihood, and the parameters are updated by multiple calibration to obtain the calibration results.
Citation Information
Patent Citations
Method for designing distributed-type hydrographical model based on penetration-storage integrated dynamic runoff yield mechanism
CN102419788A
Method for constructing urban flood loss function lacking disaster data based on dynamic proportion substitution and hierarchical Bayesian
CN112116200A
Precipitation space-time prediction method and electronic equipment
CN114118538A
Regional evapotranspiration change attribution analysis method considering multi-element spatial dependency
CN114462518A
Hydrological ecological response relation and ecological flow determination method and system based on Bayesian hierarchical model
CN115130736A
