Method for Risk Regulation and Comprehensive Regulation and Control of Water Resources in Complex Water Network System

Abstract

The present invention provides a method for water resource risk scheduling and comprehensive regulation of a complex water network system, and the steps are as follows: identifying the topological relationship and risk factor set of the complex water network system; screening the important risk factor set, establishing a joint distribution by using an improved variational Bayesian network, performing random sampling to generate samples of the important risk factor set; obtaining the runoff time lag relationship by using the cross-correlation function in combination with the VMD method; dividing the complex water network system into upper, middle, and lower layers according to the runoff time lag relationship and the topological relationship of the complex water network system; constructing a three-layer nested water resource risk scheduling model for the complex water network system, using the samples of the important risk factor set as input, and obtaining and screening the scheduling plan set. Compared with the traditional risk scheduling, the present invention takes into account the overall situation, clarifies the focus of regional scheduling, effectively balances the economy and safety of the water network system, improves the calculation efficiency of the scheduling model, and provides strong support for the water resource risk scheduling of the complex water network system.

Description

Technical Field

[0001] The present invention relates to the field of water resources management and risk scheduling, and specifically refers to a method for water resources risk scheduling and comprehensive regulation of complex water network systems. Background Art

[0002] With the acceleration of urbanization and the impact of climate change, water resources management and risk scheduling are facing unprecedented challenges. To address this issue, China has been vigorously promoting the construction of the national water network in recent years.

[0003] Water resources risk scheduling refers to the scheduling that takes into account the uncertainties of hydrological, meteorological and other forecast results and potential risks during the operation of the water resources system, with the aim of obtaining risk benefits. Initially, methods such as the recurrence period method and the first-order second-moment method were used to quantitatively estimate the magnitude of risks in water resources risk scheduling; subsequently, methods such as stochastic processes, Markov chains, and Monte Carlo simulations were used to solve the problem due to the randomness of risks; up to now, traditional optimization scheduling methods (linear / nonlinear programming, dynamic programming, etc.), intelligent optimization algorithms, and stochastic dynamic programming have also been applied to water resources risk scheduling. Therefore, the existing water resources risk scheduling methods mainly conduct overall simulation and optimization of the water resources system based on optimization algorithms and stochastic simulations to cope with hydrological uncertainties, supply-demand contradictions, and engineering safety risks.

[0004] With the expansion of the scale of the water network system and the increase in engineering complexity, traditional water resources risk scheduling methods cannot take into account the different requirements of the multi-level projects of the national water network's "outline, branches, and nodes", and cannot effectively respond to emergencies in actual work. The problems that need to be solved urgently currently include unscientific stratification of the scheduling model, low calculation efficiency of the scheduling model, unclear focus of regional scheduling, inaccurate identification of runoff time lag relationships, insufficient consideration of runoff time lag relationships, poor combined distribution effect of risk factors, poor sampling efficiency of the risk factor set, and insufficient identification of the topological relationship of the complex water network system. Therefore, there is an urgent need for a water resources risk scheduling and comprehensive regulation method that can hierarchically optimize the complex water network system. Summary of the Invention

[0005] Object of the Invention: To provide a method for water resources risk scheduling and comprehensive regulation of complex water network systems to solve the above problems existing in the prior art.

[0006] Technical Solution: The method for water resources risk scheduling and comprehensive regulation of complex water network systems includes the following steps:

[0007] Step S1: Collect basic data of the study area; identify the topological relationship of the complex water network system and the risk factor set from it;

[0008] Step S2: Screen out the important risk factor set from the risk factor set, establish the marginal distribution of risk factors one by one; establish the joint distribution using the improved variational Bayesian network; conduct random sampling to generate samples of the important risk factor set.

[0009] Step S3: Read the runoff data of reservoirs and pump stations from the basic data, and use the cross-correlation function combined with the VMD method to obtain the runoff time lag relationship; according to the runoff time lag relationship and the topological relationship of the complex water network system, divide the complex water network system into upper, middle, and lower layers, corresponding to short-term, medium-term, and long-term responses respectively.

[0010] Step S4: Call the three-layer nested water resources risk scheduling model of the complex water network system, use the samples of the important risk factor set as input, obtain the scheduling plan set and screen it; among them, the water resources risk scheduling model includes constructing the objective functions of the middle, branch, and node levels of the water network system for the upper, middle, and lower layers respectively.

[0011] The present invention proposes a method for water resources risk scheduling and comprehensive regulation of a complex water network system. Through topological relationship recognition, improved variational Bayesian network, improved Latin hypercube sampling, runoff time lag relationship analysis, and constructing a three-layer nested water resources risk scheduling model of the complex water network system, a practical method for water resources risk scheduling and comprehensive regulation of the complex water network system is provided, solving the existing problems, ensuring the sustainable development of the complex water network system, and providing a solid water resources guarantee for social and economic development.

[0012] According to one aspect of the present application, in the step S4, the construction steps of the water resources risk scheduling model are specifically as follows:

[0013] Determine the objective functions and weights of the middle, branch, and node levels of each of the upper, middle, and lower layers one by one, and obtain the upper, middle, and lower layer objective functions by weighting each layer respectively.

[0014] Based on the scheduling characteristics of each layer, use the corresponding algorithm to calculate, and calculate the optimal scheduling plans of each layer in sequence from the lower layer to the upper layer, and use the optimal scheduling plan of the lower layer as the boundary condition of the middle layer, and the optimal scheduling plan of the middle layer as the boundary condition of the upper layer.

[0015] According to one aspect of the present application, in the step S3, according to the runoff time lag relationship and the topological relationship of the complex water network system, dividing the complex water network system into upper, middle, and lower layers is specifically as follows:

[0016] Based on the runoff time lag relationship between adjacent pump stations and reservoirs, divide the time lag step into three types of time steps: short-term, medium-term, and long-term responses, and combine the three types of time steps to divide the complex water network system into three time response subsystems: short-term, medium-term, and long-term.

[0017] Based on the topological relationship of the complex water network system, calculate the network complexity of adjacent pumping stations and reservoirs; calculate the risk complexity of each node based on the risk factors of the pumping stations and reservoirs; combine the network complexity and the risk complexity to divide the complex water network system into three spatial risk subsystems: high, medium, and low.

[0018] Establish the mapping relationship between each time response subsystem and the spatial risk subsystem, sort them, and determine the importance degree for relative water resource risk scheduling; divide them into upper, middle, and lower layers based on system security and economy.

[0019] According to one aspect of the present application, in step S3, the cross-correlation function is combined with the VMD method to obtain the runoff time lag relationship, specifically as follows:

[0020] Read the runoff data of each node including pumping stations and reservoirs, perform preprocessing, and decompose it using the VMD method. Combine the central frequency method to determine the number of IMFs and obtain the optimal IMFs.

[0021] For the optimal IMFs of the runoff data of adjacent nodes, calculate the cross-correlation function in sequence to obtain the cross-correlation between the IMFs of the runoff data of adjacent nodes, and take the peak value of the cross-correlation function of each pair of IMFs as the time lag.

[0022] Calculate the cross-correlation coefficient between each IMF and the original signal, and calculate the weight of each pair of IMFs accordingly. Determine the time lag between adjacent nodes through the weighted average method to obtain the runoff time lag relationship between each node of the system.

[0023] According to one aspect of the present application, step S1 is further as follows:

[0024] Step S11: Determine the research area range, and collect the basic composition of the complex water network system and the basic data of water conservancy projects.

[0025] Step S12: Construct structured data, and use the knowledge graph technology to identify the topological relationship of the complex water network system to obtain the generalized topological relationship of the complex water network system in the research area. The topological relationship includes source-sink relationship, supply-demand relationship, regression relationship, and subordination relationship.

[0026] Step S13: Use expert consultation to identify the set of risk factors existing in the research area.

[0027] According to one aspect of the present application, step S2 is further as follows:

[0028] Step S21: Screen the important risk factor set; determine the most suitable prior distribution of each risk factor in the important risk factor set one by one; determine the cross-correlated and independent risk factors in the important risk factor set.

[0029] Step S22: Based on the optimal prior distribution, an improved variational Bayesian network is used to establish a joint distribution for the cross-correlated risk factors;

[0030] Step S23: For the joint and independent distributions of the risk factors, improved Latin hypercube sampling is used to extract samples respectively, and the initial risk factor set samples are integrated; the genetic algorithm is used to optimize the samples to obtain the final risk factor set.

[0031] According to one aspect of the present application, the step S22 is further as follows:

[0032] Step S22a: Comparative learning is used to perform feature compression on the cross-correlated risk factors in the important risk factor set, and low-dimensional dense representations are extracted;

[0033] Step S22b: An improved variational Bayesian network is constructed based on the optimal prior distribution, and with the low-dimensional dense representation as the input, in the latent space of the variational autoencoder, the contribution degree of the latent variables to the risk factors is quantified through gradient weight analysis; a progressive L1 sparse penalty is imposed on the latent variables based on the contribution degree ranking, the final latent variables are screened, and their posterior distributions are read to establish a joint distribution, where the Wasserstein distance is used as the difference metric for the joint distribution.

[0034] According to one aspect of the present application, the step S23 is further as follows:

[0035] Step S23a: Based on the curvature k(x) of the risk factor PDF, dynamic stratification is performed on the joint and independent distributions respectively, and the interval width is adaptively adjusted according to the curvature. The high-curvature region is compressed to 1 / (1 + k(x)) of the original width, and the low-curvature region is widened by n times;

[0036] Step S23b: 100% of the samples are generated by Latin hypercube for each stratification, and x% of the samples are extracted for the non-tail region of the PDF; for the tail region, importance sampling is used to obtain (100 - x)% of the samples in the tail region, and the samplings of the joint and independent distributions are integrated to form the initial risk factor samples;

[0037] Step S23c: The initial risk factor samples are used as the input of the genetic algorithm for optimization to obtain the final risk factor set.

[0038] According to one aspect of the present application, the step S3 is further as follows:

[0039] Step S31: Read the runoff data of each node including pumping stations and reservoirs and perform preprocessing; use the cross-correlation function combined with the VMD method to calculate the runoff time lag relationship;

[0040] Step S32: According to the runoff time lag relationship between adjacent pumping stations and reservoirs and the topological relationship of the complex water network system, the entire complex water network system is divided into upper, middle, and lower layers; the upper layer has a short response time and a high risk, and real-time scheduling is adopted; the middle layer has a medium response time and a moderate risk, and short-term scheduling is adopted; the lower layer has a long response time and a low risk, and medium- and long-term scheduling is adopted.

[0041] According to one aspect of the present application, in step S4, objective functions for each layer of the outline, items, and nodes in the water network system are constructed for the upper, middle, and lower layers respectively, specifically as follows:

[0042] For the upper layer, considering safety, the objective function of the outline layer is to minimize the maximum water flow rate at the key cross-section of the water conveyance line, the objective function of the item layer is to minimize the maximum water flow rate at the water intake, and the objective function of the node layer is to maximize the total flood control safety degree of the reservoir group;

[0043] For the middle layer, considering economy and safety, the objective function of the outline layer is to minimize the maximum water flow rate at the key cross-section of the water conveyance line and maximize the regional economic benefit, the objective function of the item layer is to minimize the maximum water flow rate at the water intake and minimize the total water shortage of the regional social economy, and the objective function of the node layer is to maximize the total flood control safety degree of the reservoir group and maximize the comprehensive power generation benefit of the reservoir group;

[0044] For the lower layer, considering economy, the objective function of the outline layer is to maximize the regional economic benefit, the objective function of the item layer is to minimize the total water shortage of the regional social economy, and the objective function of the node layer is to maximize the comprehensive power generation benefit of the reservoir group.

[0045] Beneficial effects: The proposed method of integrating knowledge graph technology through structured data can efficiently analyze the physical connections and dynamic interaction relationships of complex water network systems, significantly improve the accuracy of topological relationship recognition, and consider more comprehensive topological relationships such as source-sink relationships and supply-demand relationships, providing a reliable basis for subsequent hierarchical optimization. Constructing a joint distribution based on the improved variational Bayesian network can significantly reduce the dimensional redundancy of risk factors, solve the problems of being easily trapped in local optima and redundant latent variables, and establish a more suitable joint distribution for mutually correlated risk factors. Based on the improved Latin hypercube sampling, through dynamic stratification and combining importance sampling, the problem of insufficient tail coverage of Latin hypercube sampling can be effectively solved, ensuring more uniform sampling. Additionally, optimizing the samples through the genetic algorithm to obtain a more accurate set of important risk factor samples, ensuring that the risk factors are closer to the actual complex working conditions. Using the cross-correlation function combined with the VMD method to analyze unsteady runoff data to obtain accurate runoff time lag relationships. According to the runoff time lag relationships and the topological relationships of complex water network systems, the complex water network system is divided into upper, middle, and lower layers, providing a scientific basis for the hierarchical division of the subsequent water resource risk scheduling model. This hierarchical strategy effectively solves the problem of insufficient consideration of runoff time lag relationships in traditional methods and comprehensively considers network complexity and risk complexity. Applying this stratification to carry out risk scheduling can effectively enhance the rapid response ability of complex water network systems to emergencies and optimize water resource allocation at the same time. The constructed three-layer nested water resource risk scheduling model for complex water network systems is divided into upper, middle, and lower layers according to different runoff time lags and risk levels. The upper, middle, and lower layers correspond to real-time scheduling, short-term scheduling, and medium- and long-term scheduling respectively. Each layer considers the engineering characteristics and differentiated requirements of the outline, items, and knots, conducts hierarchical scheduling, takes into account the overall situation, clarifies the focus of regional scheduling, effectively balances the economy and safety of the water network system, and improves the calculation efficiency of the scheduling model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is the flowchart of the present invention.

[0047] Figure 2 is the flowchart of step S1 of the present invention.

[0048] Figure 3 is the flowchart of step S2 of the present invention.

[0049] Figure 4 is the flowchart of step S3 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] As Figures 1 to 4 shown, the following technical solutions are proposed. According to one aspect of the present application, a method for water resource risk scheduling and comprehensive regulation of a complex water network system is provided, which is characterized by including the following steps:

[0051] Step S1: Collect the basic data of the study area; identify the topological relationship of the complex water network system and the risk factor set from it;

[0052] Step S2: Screen out the important risk factor set from the risk factor set, establish the marginal distribution of each risk factor one by one; establish the joint distribution using the improved variational Bayesian network; perform random sampling to generate samples of the important risk factor set;

[0053] Step S3: Read the runoff data of reservoirs and pumping stations from the basic data, use the cross-correlation function combined with the VMD method to obtain the runoff time lag relationship; according to the runoff time lag relationship and the topological relationship of the complex water network system, divide the complex water network system into upper, middle, and lower layers, corresponding to short-term, medium-term, and long-term responses respectively;

[0054] Step S4: Call the three-layer nested water resource risk scheduling model of the complex water network system, use the samples of the important risk factor set as input, obtain and screen the scheduling plan set; among them, the water resource risk scheduling model includes constructing the objective function for each layer of the outline, items, and nodes in the water network system for the upper, middle, and lower layers respectively.

[0055] In this embodiment, to obtain a more accurate topological relationship, the present invention constructs structured data and uses knowledge graph technology to identify the topological relationship of the complex water network system. And the present invention considers topological relationships such as source-sink relationship, supply-demand relationship, regression relationship, and subordination relationship. Therefore, the identification of the topological relationship in the study area is more comprehensive, considering the internal structure of the water network system and its interactions. The accurate and comprehensive topological relationship provides a reliable basis for the subsequent layering of the complex water network system.

[0056] In the water resource risk scheduling of the complex water network system, establishing the joint distribution of risk factors is the basis for risk factor sampling. However, the traditional Bayesian network is prone to problems such as falling into local optima and redundant latent variables, resulting in insufficient accuracy in joint distribution modeling and low computational efficiency. To solve this problem, the present invention proposes an improved variational Bayesian network. By screening latent variables and using the Wasserstein distance as the difference measure of the joint distribution, this problem can be effectively solved, a more suitable joint distribution can be established for mutually correlated risk factors, and the computational efficiency can be improved.

[0057] Traditional Latin hypercube sampling has the disadvantages of insufficient coverage of the sample tails and uneven distribution. The present invention improves Latin hypercube sampling, adaptively adjusts the layering based on the curvature of the probability density function (PDF), and combines importance sampling to extract tail risk factor samples, significantly improving the coverage rate of extreme event samples and ensuring that the samples are more representative, providing high-quality input data for the subsequent risk scheduling model.

[0058] The non - steady characteristics of runoff data make it difficult to obtain an accurate runoff time - lag relationship directly using the cross - correlation function. To solve this problem, the present invention proposes a method combining the cross - correlation function with VMD. The runoff data is decomposed into multiple relatively stable IMFs by the VMD method, which is convenient for the subsequent cross - correlation function to obtain more characteristic information from them, thereby improving the accuracy of identifying the runoff time - lag relationship.

[0059] Traditional risk scheduling lacks sufficient consideration of the runoff time - lag relationship, and the model stratification is not scientific. According to the runoff time - lag relationship and the topological relationship of the complex water network system, the present invention divides the complex water network system into upper, middle, and lower layers, fully considering the runoff time - lag and risks, which is convenient for subsequent hierarchical scheduling, optimizing water resource allocation, and improving water resource management efficiency.

[0060] To clarify the focus of regional scheduling, effectively balance the economy and security of the water network system, and improve the calculation efficiency of the scheduling model, the present invention constructs a three - layer nested water resource risk scheduling model for the complex water network system. Real - time scheduling, short - term scheduling, and medium - and long - term scheduling are carried out in the upper, middle, and lower layers respectively. And each layer uses a corresponding algorithm for calculation based on the scheduling characteristics, and the optimal scheduling scheme of the lower layer is used as the boundary condition of the upper layer. Hierarchical scheduling reduces the calculation amount, improves the efficiency, and more scientifically and effectively faces different situations in each region, providing solid technical support for the safe operation of the complex water network system and regional economic development.

[0061] Generally speaking, the present invention proposes a series of innovative solutions for the problems urgently to be solved in the water resource risk scheduling of the complex water network system, providing reliable technical support for the scientific management and optimal allocation of the complex water network system. Based on knowledge graph technology, a comprehensive and accurate complex water network topology relationship is constructed; based on the improved variational Bayesian network, the calculation efficiency is improved, and a more suitable joint distribution is established; based on the improved Latin hypercube sampling, a more representative sample set of important risk factors is extracted; based on the method combining the cross - correlation function with VMD, the accuracy of identifying the runoff time - lag relationship is improved; based on the runoff time - lag relationship and the topological relationship of the complex water network system, the complex water network system is divided into upper, middle, and lower layers, fully considering the runoff time - lag and risks in the water resource risk scheduling, providing a scientific basis for the subsequent hierarchical scheduling of the scheduling model; a three - layer nested water resource risk scheduling model for the complex water network system is constructed, realizing the organic coordination of real - time scheduling, short - term scheduling, and medium - and long - term scheduling, greatly improving the calculation efficiency of the risk scheduling of the complex water network system, and effectively balancing the economy and security.

[0062] According to one aspect of the present application, in step S4, the construction steps of the water resource risk scheduling model are specifically as follows:

[0063] Determine the objective functions and weights of the outline, items, and knots in each of the upper, middle, and lower layers one by one. By weighting each layer, the upper, middle, and lower layer objective functions are obtained respectively.

[0064] In this embodiment, to effectively balance subjective and objective factors, ensure the rationality and scientificity of weight allocation, and provide reliable objective function support for the water resource risk scheduling of complex water network systems, the subjective and objective combined weighting method is used to determine the weights.

[0065] Standardize the objective functions of the general, item, and conclusion in each of the upper, middle, and lower layers. Use the Delphi Method to identify the internal dependencies within the objective functions of the general, item, and conclusion in each of the upper, middle, and lower layers. Based on the results of the Delphi Method, determine the control layer and network layer in the ANP network structure, determine the judgment matrix, and conduct consistency tests and corrections. According to the judgment matrix, construct the ANP supermatrix and solve for the weights. Again, use the Delphi Method to qualitatively evaluate the rationality of the weights, quantify the expert consensus degree through the Kendall coordination coefficient, evaluate the consistency of the experts on the weights, and after reaching an expert consensus, output the final subjective weights. Calculate the objective weights respectively through CRITIC (weight method) and EWM (entropy weight method), and then obtain the comprehensive objective weight through the arithmetic mean method.

[0066] Calculate the Jensen-Shannon divergence between the subjective and objective weights to quantify the difference between the two. The specific formula is as follows:

[0067] JSD(ω 主观 ||ω 客观 )=1 / 2·D KL (ω 主观 ||M)+1 / 2·D KL (ω 客观 ||M);

[0068] In the formula, JSD is the Jensen-Shannon divergence, D KL is the KL divergence, ω 主观 and ω 客观 are the subjective weight and the comprehensive objective weight respectively, and M is the arithmetic mean of the subjective weight and the comprehensive objective weight.

[0069] Take it as a correction factor and perform proportional allocation on the different parts between the subjective and objective weights. The specific formula is as follows: ω j综合 =ω j主观 +α•JSD•ω j客观 ;

[0070] In the formula, ω j综合 is the comprehensive weight of the jth objective function, ω j主观 is the subjective weight of the jth objective function, ω j客观 is the objective weight of the jth objective function, JSD is the Jensen-Shannon divergence, and α is the correction coefficient, usually taken as 0.5.

[0071] Through the integration of proportional allocation, the weights of the objective functions of the general outline, items, and knots in each of the upper, middle, and lower layers are obtained to determine the objective functions of the upper, middle, and lower layers.

[0072] Based on the scheduling characteristics of each layer, corresponding algorithms are used for calculation. Starting from the lower layer to the upper layer, the optimal scheduling plans for each layer are calculated in sequence, and the optimal scheduling plan of the lower layer is used as the boundary condition of the middle layer, and the optimal scheduling plan of the middle layer is used as the boundary condition of the upper layer.

[0073] In this embodiment, the specific constraint conditions considered in the objective functions of the general outline, items, and knots in each of the upper, middle, and lower layers are as follows:

[0074] (1) Outflow discharge constraint: Q j (t) ≥ Q min,j ; Q j (t) ≤ Q max,j ; In the formula, Q j (t) is the outflow discharge of the j-th reservoir at the t-th time period, Q min,j is the minimum downstream discharge limit value of the j-th reservoir, Q max,j is the flood control safety discharge value of the downstream river of the j-th reservoir, and j is the reservoir index.

[0075] (2) Initial / final water level constraint: Z j (1) = Z j,s ; Z j (T) = Z j,e ; In the formula, Z j (1) is the water level of the j-th reservoir at the 1st time period, Z j,s is the starting water level of the j-th reservoir at the beginning of the scheduling, Z j (T) is the water level of the j-th reservoir at the T-th time period, Z j,e is the water level at the end of the scheduling of the j-th reservoir, and j is the reservoir index.

[0076] (3) Reservoir characteristic water level constraint: Z min,j ≤ Z j (t) ≤ Z max,j ; In the formula, Z j (t), Z min,j , Z max,j are respectively the water level, dead water level, and normal storage water level of the j-th reservoir at the t-th time period, and j is the reservoir index.

[0077] (4) Discharge variable amplitude constraint: |Q j '(t) - Q j '(t - 1)| < ΔQ j,con ; In the formula, Q j '(t) is the outflow discharge of the j-th reservoir at the t-th time period, and ΔQ j,conis the allowable discharge variation range for adjacent periods of the j-th reservoir, where j is the reservoir index.

[0078] (5) Output constraint: N min,j ≤ N j (t) ≤ N max,j ; In the formula, N j (t), N min,j , N max,j are respectively the output, guaranteed output and installed capacity of the j-th reservoir at the t-th period, and j is the reservoir index.

[0079] (6) Storage capacity constraint: V j (t) ≥ V 死,j ; V j (t) ≤ V j obj (t);

[0080] In the formula, V j (t) is the storage volume of the j-th reservoir at the t-th period, V 死,j is the dead storage capacity of the j-th reservoir, V j obj (t) is the target storage volume of the j-th reservoir at period t. During the flood season, V j obj (t) is equal to the flood limit storage capacity, and during the non-flood season, V j obj (t) is equal to the beneficial storage capacity.

[0081] (7) Reservoir water balance constraint: V j (t) = V j (t - 1) + intQ j (t)Δt - Q j '(t)Δt - L j (t);

[0082] In the formula, intQ j (t) is the inflow of the j-th reservoir at period t, L j (t) is the water loss of the j-th reservoir at period t, Δt is the time interval, and other variables are the same as above. Note the dimension conversion.

[0083] (8) Non-negativity constraint: All variables are not less than 0, that is, X > 0.

[0084] Considering that real-time scheduling is adopted in the upper layer, the ant colony hunting coupled with particle swarm optimization algorithm is used to solve the upper layer of the model; considering that short-term scheduling is adopted in the middle layer, the multi-objective game theory optimization is used to solve the middle layer of the model; considering that medium and long-term scheduling is adopted in the lower layer, dynamic programming is used to solve the lower layer of the model. The model calculates the optimal scheduling schemes of each layer in sequence from the lower layer to the upper layer, and uses the optimal scheduling scheme of the lower layer as the boundary condition of the middle layer and the optimal scheduling scheme of the middle layer as the boundary condition of the upper layer.

[0085] After each layer of the model obtains the scheduling scheme set, a comprehensive screening method based on NSGA-II combined with fuzzy TOPSIS is adopted. Through the NSGA-II method for preliminary screening, non-dominated sorting and crowding degree are calculated to screen out a high-quality solution set. The fuzzy TOPSIS is used for secondary screening to accurately rank the scheduling schemes, ensuring that the screening results not only conform to the objective characteristics of the objective function but also can reflect the subjective preferences of the decision-makers. Finally, the optimal scheduling schemes of the upper, middle, and lower layers are obtained. The calculation formulas for fuzzy distance and fuzzy similarity are as follows:

[0086] d i + = sqrt[∑ j=1 m {d(v* ij , v* j + )} 2 , d i - = sqrt[∑ j=1 m {d(v* ij , v* j - )} 2 ; where: v* j + , v* j - are fuzzy decision matrices; d i + , d i - are the Euclidean distances between the i-th scheduling scheme and the fuzzy positive ideal solution and the fuzzy negative ideal solution respectively; m is the number of attributes; d(v* ij , v* j + ) is the Euclidean distance between the j-th attribute of the i-th scheduling scheme and the positive ideal solution; d(v* ij , v* j - ) is the Euclidean distance between the j-th attribute of the i-th scheduling scheme and the negative ideal solution; v* ij is the fuzzy evaluation value of the j-th attribute of the i-th scheme; i is the scheme index; j is the attribute index; C i=(d i - ) / (d i + +d i - ); where: C i is the fuzzy similarity, and other variables are the same as above.

[0087] According to one aspect of the present application, in the step S3, according to the runoff time lag relationship and the topological relationship of the complex water network system, the complex water network system is divided into upper, middle, and lower layers, specifically as follows:

[0088] Based on the runoff time lag relationship between adjacent pump stations and reservoirs, the time lag step length is divided into three types of time step lengths: short-time, medium-time, and long-time responses. Combining the three types of time step lengths, the complex water network system is divided into three time response subsystems: short-time, medium-time, and long-time.

[0089] Based on the topological relationship of the complex water network system, calculate the network complexity of adjacent pump stations and reservoirs; calculate the risk complexity of each node based on the risk factors of the pump stations and reservoirs; combine the network complexity and the risk complexity to divide the complex water network system into three spatial risk subsystems: high, medium, and low.

[0090] Establish the mapping relationship between each time response subsystem and the spatial risk subsystem, perform sorting, and determine the importance degree of the relative water resource risk scheduling; divide it into upper, middle, and lower layers based on system security and economy.

[0091] In this embodiment, for the scientific stratification of the subsequent water resource risk scheduling model, according to the runoff time lag relationship and the topological relationship of the complex water network system, fully considering the runoff time lag and risk, the complex water network system is scientifically divided into upper, middle, and lower layers.

[0092] In the current risk scheduling model, the runoff time lag relationship is not considered, but its essence has an important impact on risk scheduling. The shorter the time lag step length, the more urgent the incoming water, and the greater the possible risk. For more scientific scheduling, the present invention divides the time lag step length into three types of time step lengths: short-time, medium-time, and long-time responses. The time lag step length of the short-time response is less than or equal to 1 day, the time lag step length of the medium-time response is greater than 1 day and less than or equal to 1 dekad, and the time lag step length of the long-time response is greater than 1 dekad. And combining the three types of time step lengths, the complex water network system is divided into three time response subsystems: short-time, medium-time, and long-time.

[0093] There are still deficiencies in the model layering when only considering the runoff time lag relationship, which may lead to resource waste in areas with short time lag steps but low risks. Since the network complexity of the water network system is denser, there may be more risk factors and higher risk levels in the relevant areas. Therefore, the present invention calculates the network complexity of adjacent pump stations and reservoirs based on the topological relationship of the complex water network system, calculates the risk complexity of each node based on the risk factors of the pump stations and reservoirs, quantifies the water network risk by combining the network complexity and the risk complexity through the arithmetic mean method, and uses the K-means algorithm to divide the complex water network system into three spatial risk subsystems: high, medium, and low.

[0094] The formula for calculating the network complexity is as follows: C net j = n / N; where C net j is the network complexity of the j-th node, n is the number of edges directly related to the node, and N is the total number of edges in the complex water network system.

[0095] The formula for calculating the risk complexity is as follows: C risk j = 1 / K · ∑ k=1 K R k j ; where C risk j is the risk complexity of the j-th node, K is the number of types of risk factors of the j-th node, and R k j is the value of the k-th type of risk factor of the j-th node after normalization.

[0096] The water network risk is quantified by combining the network complexity and the risk complexity through the arithmetic mean method. The specific formula for the comprehensive risk is as follows: C j = 1 / 2 (C net j + C risk j ).

[0097] A mapping relationship between each time response subsystem and the spatial risk subsystem is established. Each time response subsystem is respectively mapped to three types of spatial risk subsystems, forming 9 types of mapping relationships, and a spatio-temporal risk mapping matrix is constructed:

[0098] M = [m 11 , m 12 , m 13 ; m 21 , m 22 , m 23 ; m 31 , m 32 , m 33 ; where m ijIt is the proportion of the j-th spatial risk subsystem node in each type of mapping relationship to the total nodes in the i-th time response subsystem. Here, i = 1, 2, 3, corresponding to the short-term, medium-term, and long-term time response subsystems respectively, and j = 1, 2, 3, corresponding to the high, medium, and low spatial risk subsystems respectively.

[0099] After establishing the mapping relationship, extract the key mapping features of each type of mapping relationship. The specific formula is as follows: T = m ij ·(1 / k - n )·C - n ; where: T is the key mapping feature; m ij is the mapping coefficient; k - n is the average runoff delay of the n-th type of mapping relationship; C - n is the average comprehensive risk of the n-th type of mapping relationship; n is the mapping relationship category index.

[0100] Based on the key mapping features of each type of mapping relationship, calculate the information entropy and assign weights through the entropy weight method, then calculate the closeness degree through TOPSIS, and sort the mapping relationships from high to low to determine the importance degree of each type of mapping relationship relative to the water resource risk scheduling.

[0101] In the above sorting, the earlier the ranking, the greater the risk of faster incoming water. If risk scheduling is carried out uniformly, it may lead to waste of some resources and huge computational workload. Therefore, for scientific stratification of subsequent risk scheduling, effectively balancing the economy and security of the water network system, and improving the computational efficiency of the scheduling model, based on the security of the complex water network system, the top a% before sorting is classified as the upper layer, based on security and economy, the top a% - b% before sorting is classified as the middle layer, and based on economy, the latter (100 - a - b)% after sorting is classified as the lower layer.

[0102] According to one aspect of the present application, in the step S3, the cross-correlation function is combined with the VMD method to obtain the runoff delay relationship, specifically as follows:

[0103] Read the runoff data of each node including pumping stations and reservoirs, perform preprocessing, and decompose using the VMD method. Combine the central frequency method to determine the number of IMFs to obtain the optimal IMFs;

[0104] Calculate the cross-correlation function for the optimal IMFs of the runoff data of adjacent nodes in sequence to obtain the cross-correlation between the IMFs of the runoff data of adjacent nodes, and take the peak value of the cross-correlation function of each pair of IMFs as the delay;

[0105] Calculate the cross - correlation coefficients between each IMF and the original signal, and calculate the weights of each pair of IMFs based on this. Determine the time delay between adjacent nodes through the weighted average method to obtain the runoff time delay relationship between each node of the system.

[0106] In this example, it is necessary to pre - process the runoff data of each node including pump stations and reservoirs in the complex water network system, check the data integrity. If there are missing values, fill the missing values using the interpolation method. Use low - pass filtering to remove data noise and improve data quality. And perform standardization processing on the processed data to eliminate the influence of dimensions to obtain the final data.

[0107] In the analysis of the runoff data of each node including pump stations and reservoirs in the complex water network system, the VMD technology has significant advantages. Runoff data belongs to non - steady - state information. Using the VMD technology, the runoff data can be decomposed into multiple relatively stable IMF series components, which can obtain characteristic information from the original data more comprehensively and provide richer information support for subsequent identification of the runoff time delay relationship. It mainly uses the alternating direction method of multipliers, Fourier transform, etc. to complete the sequence decomposition. The specific principle is as follows: min {uk},{wk} {∑ k=1 K ||dt[(d(t)+j / πt)*u k (t)]e -jw k t )||2 2}; s.t∑ k=1 K u k =f;

[0108] In the formula, {u k} is the k - th IMF after variational mode decomposition, {w k} is the k - th center frequency after variational mode decomposition, f is the original non - steady - state sequence, K is the number of mode decompositions, k is the index of the IMF order, d(t) is the impulse function, d is the partial derivative symbol, j is the imaginary unit, * is the convolution operation, e -jw k t is the frequency - domain modulation term, t is the independent variable of the signal in the time domain, || ||2 2 is the square of the L2 norm.

[0109] In practical applications, a key issue of VMD is how to determine the appropriate number K of IMFs decompositions. The number of IMFs will significantly affect whether more useful information can be extracted from the original data subsequently. If the number is too large or too small, it may lead to serious overlap of IMFs. To avoid this situation, the present invention determines the optimal number of IMFs based on the central frequency method. The central frequency of VMD refers to the average frequency of each mode function, which reflects the main vibration characteristics of the mode. The central frequency can be obtained by calculating the mean of each mode function or through peak detection. By calculating the residual sum of squares (RSS), the K value that minimizes these error metrics is selected. The formula for the residual sum of squares (RSS) is as follows: RSS = ∑ i=1 N (y i – y* i ) 2 ; In the formula, y i is the actual value of the i-th original water-lifting data, y* i is the i-th predicted value of the reconstructed signal, N is the length of the original water-lifting data, and i is the number of bits of the data sequence.

[0110] The cross-correlation function can describe the similarity of two random discrete signals at different time delays. Its specific definition is as follows: R xy [k]=∑ -∞ ∞ x[n]•y[n+k]; In the formula, R xy [k] is the cross-correlation function value at k time delays, x[n] and y[n] are both random discrete signals, k is the time delay, n is the number of bits of the time sequence, and ∞ is the infinity symbol.

[0111] The present invention combines the cross-correlation function with the VMD method, calculates the cross-correlation function for the optimal IMFs of the runoff data of adjacent nodes in turn, obtains the cross-correlation between the IMFs of the runoff data of adjacent nodes, and then determines the time delay of each pair of IMFs. Through the cross-correlation coefficient, the correlation strength between each IMF and the original signal can be quantified. The stronger the correlation of an IMF, the greater its weight, and the weaker the correlation of an IMF, the smaller its weight.

[0112] The specific definition of the cross-correlation coefficient is as follows: ρ i = [∑ n=1 N (C i [n] – C - i )·(x[n] - x - )] / sqrt{[∑ n=1 N (C i [n] – C - i )2 ·(x[n] – x - ) 2}; where: ρ i is the cross-correlation coefficient between the i-th IMF and the original signal; C i [n] is the i-th IMF, i.e., the n-th value of the i-th sequence; C - i is the mean of the i-th sequence; x[n] is the n-th value of the original sequence; x - is the mean of the original sequence x[n]; N is the length of the original water-lifting data; n is the number of bits of the time series.

[0113] The calculation formula for weight allocation is as follows: ω i =|ρ i | / (∑ j=1 K |ρ j |);

[0114] In the formula, ω i is the weight of the i-th IMF, |ρ i | is the cross-correlation coefficient between the i-th IMF and the original signal, K is the total number of IMFs, j is the index of the IMF sequence number, and all weights need to satisfy ∑ i=1 K ω i = 1.

[0115] After obtaining the weights of each IMF relative to the original signal, the IMFs perform an arithmetic average on their respective weights relative to the original signal, and the weight ω j of the IMFs can be obtained. After determining the time delay k j and the weight ω j for each pair of IMFs, the overall time delay k 整体 can be calculated by the weighted average method. The specific formula is as follows: k 整体 =∑ j J k j •ω j ;

[0116] In the formula, k j is the time delay of the j-th pair of IMFs, ω j is the weight of the j-th pair of IMFs, and J is the total number of IMF pairs.

[0117] As Figure 2 shown, according to one aspect of the present application, the step S1 is further:

[0118] Step S11, determine the research area range, and collect the basic composition of the complex water network system and the basic data of water conservancy projects;

[0119] In this embodiment, the basic components of the complex water network system include rivers (main streams, tributaries), lakes, reservoirs, pumping stations, artificial waterways, water conveyance pipelines, etc. The basic data of the water conservancy project includes hydrometeorological data, water resources data, and design parameter data of water conservancy infrastructure. Among them, the hydrometeorological data includes runoff data of pumping stations and reservoirs, etc. The water resources data includes water resources quantity, water quality, water resources utilization, water resources supply and demand, etc. The design parameter data of water conservancy infrastructure includes design parameters of projects such as reservoirs, irrigation, hydropower stations, and flood control.

[0120] The basic components of the complex water network system in the above data can be obtained from public data. However, due to confidentiality reasons, the basic data of the water conservancy project needs to be obtained from relevant water conservancy departments. The purpose of obtaining the basic components of the complex water network system is to construct structured data subsequently and use knowledge graph technology to identify the topological relationship of the complex water network system. The purpose of obtaining the basic data of the water conservancy project is to further understand the general situation of the study area and collect data for subsequent identification of runoff time lag relationships and construction of a three-layer nested water resources risk scheduling model for the complex water network system.

[0121] Step S12: Construct structured data, use knowledge graph technology to identify the topological relationship of the complex water network system, and obtain the generalized topological relationship of the complex water network system in the study area. The topological relationships include source-sink relationship, supply-demand relationship, regression relationship, and subordination relationship.

[0122] Define the nodes in the complex water network system as water source nodes, user nodes, water discharge nodes, and regulation nodes. The water source nodes include lakes, reservoirs, rivers, etc. The user nodes include domestic water users, agricultural irrigation water users, industrial water users, etc. The water discharge nodes include sewage treatment plants, natural regression water bodies, etc. The regulation nodes include pumping stations, gates, water treatment plants, etc. Define the edge as the relationship between two nodes, that is, the source-sink relationship, supply-demand relationship, regression relationship, and subordination relationship. Define the attributes according to the node classification. The attributes of the water source nodes are defined as flow rate, storage capacity, etc. The attributes of the user nodes are defined as water demand. The attributes of the water discharge nodes are defined as water discharge volume. The attributes of the regulation nodes are defined as regulation volume. And store the nodes, edges, and attributes in the database to obtain structured data.

[0123] Convert the node, edge, and attribute data in the structured data into the RDF triple format (subject - predicate - object) required by the knowledge graph through knowledge extraction and entity alignment, construct an ontology, and complete the semantic modeling of the complex water network system using the RDF ontology language. Then, form new knowledge forms through knowledge reasoning, and jointly undergo quality assessment with the original knowledge to complete knowledge fusion. Import the final RDF triples into the Neo4j graph database, construct the knowledge graph of the complex water network system, identify the topological relationship of the complex water network system, and obtain the generalized topological relationship of the complex water network system in the study area.

[0124] In this embodiment, the RDF triple can be expressed as (h, r, t), where h is the head entity, r is the relationship, and t is the tail entity. For example, the triple (water source node A, supply-demand relationship, user node A) means that water source node A supplies water to user node A.

[0125] The knowledge graph can be represented as a directed graph, defined as follows: G = (V, b); where V is the set of nodes {V1, V2, …, V M}, M is the number of node types, and each type of node V i represents an N-dimensional spectral feature vector, and N is the number of nodes included in each type of node; b is the set of edges {b1, b2, …, b n}, and n is the number of edge types.

[0126] Identifying topological relationships based on the knowledge graph requires constructing an ontology, defined as follows: H x =(V x , P x , R x ); where H x is the xth ontology, V x is the set of nodes, P x is the set of attributes, R x is the set of relationships, and x is the index of the ontology order.

[0127] Knowledge graph technology can integrate the node, edge, and attribute data in a complex water network system into a unified and inferable knowledge network. Based on the RDF triple format, the knowledge graph can accurately express the topological relationships in the complex water network system, such as source-sink relationships, supply-demand relationships, regression relationships, and subordination relationships, and clarify the semantics of nodes, edges, and attributes through ontology definitions to support dynamic updates. In addition, the knowledge graph combines with the graph database Neo4j to achieve efficient storage and query, and at the same time discovers implicit relationships through knowledge reasoning, providing comprehensive and dynamic support for water resource risk scheduling, risk analysis, and topological relationship identification. The node, edge, and attribute data in the structured data can be updated regularly to update the knowledge graph and the topological relationships in the complex water network system, realizing the integration of dynamic data and static topological relationships.

[0128] Step S13: Identify the set of risk factors existing in the study area by means of expert consultation.

[0129] In this embodiment, various possible risk factors are comprehensively considered to ensure that all risk factors are taken into account. According to the risk environment and risk characteristics, the risk factors existing in the complex water network system are classified into three categories: hydrological risk factors, hydraulic risk factors, and technical risk factors, so as to more clearly identify the risk factors existing in the complex water network system of the study area when using expert consultation later. The hydrological risk factors include rainfall, water quality, flood, and drought. The hydraulic risk factors include dam break, collapse, siltation of reservoirs, pipeline rupture, poor drainage, errors in reservoir water level - storage capacity curves, errors in reservoir discharge capacity curves, etc. The technical risk factors include runoff prediction model structure, parameters, flood forecasting errors, and errors in river flood routing parameters.

[0130] As Figure 3 shown, according to one aspect of the present application, the step S2 is further as follows:

[0131] Step S21: Screen the set of important risk factors; determine the most suitable prior distribution for each risk factor in the set of important risk factors one by one; determine the mutually correlated and independent risk factors in the set of important risk factors;

[0132] In this embodiment, a cooperative game model is constructed, and the identified risk factors are used as the game participants of the model; the profit function of the overall risk level of the system is defined, and the Shapley value of each risk factor is calculated; the risk factors are sorted according to the magnitude of the Shapley value, and the risk factors with higher Shapley values are selected as the set of important risk factors.

[0133] The calculation formula of the Shapley value is as follows: φ i =∑ S⊆N\{i} [|S|!(|N| - |S| - 1)! / |N|!][v(S ∪ {i}) - v(S)];

[0134] In the formula, φ i is the Shapley value of the i - th risk factor, N is the set of all risk factors, S is the subset that does not include the i - th risk factor, v(S) is the profit function value of the subset S, and! is the factorial symbol.

[0135] Fit the most suitable prior distribution for each risk factor in the set of important risk factors one by one, use maximum likelihood estimation for parameter estimation, use the K - S test to conduct goodness - of - fit testing, and determine the most suitable prior distribution for each risk factor based on the R2 and RMSE parameters, where the distribution types include P - III distribution, normal distribution, Poisson distribution, generalized extreme value distribution, Weibull distribution, log - normal distribution, gamma distribution, etc.

[0136] The constructed likelihood function formula is as follows: L(θ|x)=p(x|θ)=∏ i=1 n  p(xi ; θ); where L(θ|x) is the likelihood function, θ is the parameter, x is the observed data set, n is the number of the observed data set, i is the number of bits of the observed data set sequence, ∏ is the product operator, and p(x|θ) is the probability of randomly drawing x1, …, x n as follows: p(x|θ)=∏ i=1 n p(x i ; θ).

[0137] Conduct a goodness-of-fit test based on the K-S test to determine whether the fitted distribution line type is appropriate. First, it is necessary to construct the K-S statistic D, and the specific formula is as follows: D = max|F n (x) - F0(x)|; where D is the K-S statistic, F0(x) is the original distribution function, and F n (x) is the distribution function to be tested, and max| | is the maximum value of taking the absolute value.

[0138] Compare the K-S statistic D with the critical value D n,α of the K-S test statistic. When D > D n,α , reject the null hypothesis, that is, the fitted distribution line type is inappropriate.

[0139] After determining the appropriate distribution line type, the optimal prior distribution of each risk factor can be determined based on the R2 and RMSE parameters.

[0140] R 2 The calculation formula is as follows: R 2 = 1 - [∑ i N (y* i - y i ) 2 / [∑ i N (y - i - y i ) 2 ; where: R 2 is the coefficient of determination; y i is the value of the i-th original risk factor; y* i is the value of the risk factor of the i-th fitted distribution line type; y - i is the average value of the original risk factor; N is the length of the original risk factor; and i is the sample index.

[0141] RMSE = sqrt[(1 / N)·∑ i=1 N (y i – y* i )2 ; where: RMSE is the root mean square error; y i is the value of the i-th original risk factor; y* i is the value of the risk factor of the i-th fitted distribution line type; N is the length of the original risk factor; n is the sample index.

[0142] Principal component analysis is used to determine the cross-correlated and independent risk factors in the set of important risk factors. First, standardize the set of important risk factors, calculate the correlation coefficient matrix between risk factors, then calculate the eigenvalues of the correlation coefficient matrix, and extract n as the principal components based on the Kaiser criterion. Finally, calculate the correlation coefficients between the principal components and the original factors to obtain the loading matrix. If the absolute value of the risk factor loading is greater than 0.5, it is a cross-correlated risk factor; if it is less than 0.5, it is an independent risk factor. Integrate the results of n principal component analyses to obtain cross-correlated and independent risk factors.

[0143] Step S22: Based on the optimal prior distribution, use an improved variational Bayesian network to establish a joint distribution for the cross-correlated risk factors;

[0144] Step S23: For the joint and independent distributions of risk factors, use improved Latin hypercube sampling to extract samples respectively, and integrate them to obtain the initial risk factor set samples; use a genetic algorithm to optimize the samples to obtain the final risk factor set.

[0145] According to one aspect of the present application, step S22 is further:

[0146] Step S22a: Use contrastive learning to perform feature compression on the cross-correlated risk factors in the set of important risk factors, and extract low-dimensional dense representations;

[0147] In order to reduce the dimension of risk factors while retaining the global features of risk factors, contrastive learning is used to extract low-dimensional dense representations. First, standardize the cross-correlated risk factor X, and construct positive sample pairs X pos1 =X+ξ1, ξ1~N(0,σ pos1 2 I), X pos2 =X+ξ2, ξ2~N(0,σ pos2 2 I) by adding Gaussian noise, and construct negative sample X neg =X[T:1] by reversing the time sequence. Use the InfoNCE loss function as the contrastive loss function,

[0148] For sample x i calculate positive sample x j and negative sample x k {k≠i} The specific formula is as follows: L InfoNCE = -log[exp(sin(xi , x j ) / τ) / (exp(sin(x i , x j ) / τ) + ∑ k=1 K exp(sin(x i , x k ) / τ))]; where: L InfoNCE is the contrast loss function; τ is the temperature parameter; sin(x i , x j ) is the similarity between the sample x i and the positive sample x j ; sin(x i , x k ) is the similarity between the sample x i and the negative sample x k ; K is the number of negative samples, and k is the index of the negative sample order;

[0149] Select a deep residual network as the encoder, and use the constructed positive and negative samples and the selected contrast loss function to train the encoder. With the goal of minimizing the contrast loss, map the cross-correlation risk factor to a low-dimensional dense space to obtain the corresponding low-dimensional dense representation X'.

[0150] Step S22b, construct an improved variational Bayesian network based on the optimal prior distribution, and use the low-dimensional dense representation as the input. In the latent space of the variational autoencoder, quantify the contribution of the latent variable to the risk factor through gradient weight analysis; impose a progressive L1 sparse penalty on the latent variable based on the contribution ranking, screen the final latent variable, and read its posterior distribution to establish a joint distribution, where the Wasserstein distance is used as the difference metric of the joint distribution.

[0151] In this embodiment, to solve the problems that the traditional Bayesian network is prone to falling into local optima and having redundant latent variables, resulting in insufficient accuracy of joint distribution modeling and low computational efficiency, the present invention solves this problem by improving the variational Bayesian network. First, input the low-dimensional dense representation X' into the variational autoencoder to obtain the latent variable x based on the optimal prior distribution. To avoid falling into local optima, the Wasserstein distance is used as the difference metric of the joint distribution in the improved variational Bayesian network. Map the latent variable to the reconstructed output x recon (k) = f θ (k) (x), where x recon (k) is the reconstructed output of the latent variable, and f θ (k)(x) represents the k-th output node of the decoder, corresponding to the k-th risk factor. Calculate the gradient g of the latent variable for each risk factor j (k) =dx recon (k) / (dx j ), where d is the partial derivative, and x j is the j-th latent variable. The contribution degree G of the latent variable to the risk factor is synthesized by the arithmetic mean method j =∑ k=1 K g j (k) / K, where G j is the contribution degree of the j-th latent variable to the risk factor, K is the total number of risk factors, k is the index number of the risk factor, and other variables are the same as above.

[0152] Apply a progressive L1 sparse penalty according to the ranking of the contribution degree of the latent variable, and dynamically adjust the penalty intensity to gradually sparsify the low-contribution latent variables while retaining the high-contribution latent variables. The specific principle formula is as follows: λ j (t)=λ0•exp(-α•rank(G j )) • (1+γt); In the formula, λ j (t) is the penalty intensity of the j-th latent variable at time t, λ0 is the initial penalty intensity, α is the contribution degree decay rate, γ is the time decay coefficient, and rank(G j ) is the rank in descending order of G j .

[0153] Finally, establish a joint distribution according to the posterior distribution read from the final latent variable.

[0154] According to one aspect of the present application, the step S23 is further as follows:

[0155] Step S23a: Based on the curvature k(x) of the risk factor PDF, perform dynamic stratification on the joint and independent distributions respectively. The interval width is adaptively adjusted according to the curvature. The high-curvature region is compressed to 1 / (1 + k(x)) of the original width, and the low-curvature region is widened by n times;

[0156] Considering that the change at the high-curvature part of the probability density function curve is drastic and finer stratification is needed to capture the change, while the change in the low-curvature region is gentle and the stratification width can be widened to reduce the stratification. Therefore, the present invention performs dynamic stratification on the joint and independent distributions respectively based on the curvature k(x) of the risk factor PDF. The specific calculation formula for the curvature is as follows: k(x)=(|f''(x)|) / (1+(f'(x)) 2 ) (3 / 2); where k(x) is the curvature of the PDF function, f''(x) is the first derivative of the PDF function, and f''(x) is the second derivative of the PDF function.

[0157] In this embodiment, the gradient change rate of the curvature function is incorporated into the determination, and the curvature values are divided into two categories: high and low, by the K-means clustering algorithm. The high-curvature region is compressed to 1 / (1 + k(x)) of the original width, and the low-curvature region is widened by n times.

[0158] Step S23b: Each layer generates 100% of the samples through Latin hypercube sampling. For the non-tail region of the PDF, x% of the samples are extracted; for the tail region, importance sampling is used to obtain (100 - x)% of the samples in the tail region. The sampling results of the combined joint and independent distributions are integrated to form the initial sample of the risk factor.

[0159] Based on the dynamic stratification in step S23a, the joint and independent distribution risk factors generate 100% of the samples through Latin hypercube sampling. Based on the curvature of the probability density function, the region where the curvature k(x) < α and is not a maximum point is defined as the tail region, and other regions are defined as non-tail regions. And x% of the samples are extracted for the non-tail region of the PDF.

[0160] To ensure sufficient sampling of the sample tail, the present invention combines importance sampling to sample the data in the tail region. For the tail region, importance sampling is used to obtain (100 - x)% of the samples, and the sampling results are integrated to obtain the initial sample of the risk factor.

[0161] Step S23c: The initial sample of the risk factor is used as the input of the genetic algorithm for optimization to obtain the final risk factor set.

[0162] In this embodiment, the Wasserstein distance is used to measure the overall difference between the sample distribution and the theoretical distribution, and the genetic algorithm is used to minimize the error. The threshold is set by calculating the Wasserstein distance of the initial sample of the risk factor. When the Wasserstein distance between the subsequent optimized sample distribution and the theoretical distribution exceeds the threshold, a penalty term for the sample quantity is introduced into the fitness function based on the penalty function to reduce the sample quantity. Optimization stops until the value of the fitness function converges or reaches the maximum number of iterations, and the final risk factor set is obtained.

[0163] The calculation formula of the Wasserstein distance is as follows: W(p,q) = inf γ~∏(p,q) E x,y~γ [||x - y||]; where: W(p,q) is the Wasserstein distance; ∏(p,q) is the set of all possible joint distributions of the distributions p and q; E x,y~γ[||x-y||] is the expected value of the sample pair distance; p and q are the sample and theoretical probability distributions respectively; inf represents the infimum; γ is the joint distribution.

[0164] As Figure 4 shown, according to one aspect of the present application, the step S3 is further as follows:

[0165] Step S31: Read the runoff data of each node including pumping stations and reservoirs, and perform preprocessing; use the cross-correlation function combined with the VMD method to calculate the runoff time lag relationship.

[0166] Step S32: According to the runoff time lag relationship between adjacent pumping stations and reservoirs and the topological relationship of the complex water network system, divide the entire complex water network system into upper, middle, and lower layers; the upper layer has a shorter response time and higher risk, and real-time scheduling is adopted; the middle layer has a medium response time and moderate risk, and short-term scheduling is adopted; the lower layer has a longer response time and lower risk, and medium- and long-term scheduling is adopted.

[0167] Traditional water resource risk scheduling methods usually adopt a unified scheduling strategy, which is difficult to adapt to water network subsystems with different response times and different risks, resulting in low scheduling efficiency and insufficient utilization of water resources. To clarify whether the focus of regional scheduling is on safety or economy, in this embodiment, by identifying the runoff time lag relationship between adjacent pumping stations and reservoirs, and combining the topological relationship of the complex water network system to quantify risks, the complex water network system is divided into upper, middle, and lower layers, and corresponding scheduling schemes are adopted for different layers to achieve the optimal allocation and efficient scheduling of water resources, and significantly improve the operation efficiency and resource utilization rate of the water network system.

[0168] According to one aspect of the present application, in the step S4, the objective functions of each layer of the outline, items, and nodes in the water network system are constructed for the upper, middle, and lower layers respectively, specifically as follows:

[0169] For the upper layer considering safety, the objective function of the outline layer is to minimize the maximum water passing flow rate of the key section of the water conveyance line, the objective function of the item layer is to minimize the maximum water passing flow rate of the water intake, and the objective function of the node layer is to maximize the total flood control safety degree of the reservoir group;

[0170] For the middle layer considering economy and safety, the objective function of the outline layer is to minimize the maximum water passing flow rate of the key section of the water conveyance line and maximize the regional economic benefit, the objective function of the item layer is to minimize the maximum water passing flow rate of the water intake and minimize the total water shortage of the regional social economy, and the objective function of the node layer is to maximize the total flood control safety degree of the reservoir group and maximize the comprehensive power generation benefit of the reservoir group;

[0171] For the lower layer considering economy, the objective function of the outline layer is to maximize the regional economic benefit, the objective function of the item layer is to minimize the total water shortage of the regional social economy, and the objective function of the node layer is to maximize the comprehensive power generation benefit of the reservoir group.

[0172] In this embodiment, the present invention constructs a three-layer nested water resource risk scheduling model for a complex water network system. Real-time scheduling, short-term scheduling, and medium- and long-term scheduling are carried out in the upper, middle, and lower layers respectively. Each layer considers the engineering characteristics and differentiated requirements of the main trunks, branches, and nodes, and conducts hierarchical scheduling, taking into account the overall situation, effectively balancing the economy and security of the water network system, and improving the calculation efficiency of the scheduling model.

[0173] Since the complex water network system in the upper layer has a short response time and high risk, real-time scheduling is adopted, so this layer considers the security of the complex water network system scheduling; since the complex water network system in the middle layer has a medium response time and moderate risk, short-term scheduling is adopted, so this layer considers the economy and security of the complex water network system scheduling; since the complex water network system in the lower layer has a long response time and low risk, medium- and long-term scheduling is adopted, so this layer considers the economy of the complex water network system scheduling.

[0174] The objective functions of the upper, middle, and lower layers are obtained by synthesizing the weights of the objective functions of each level of the main trunks, branches, and nodes. The specific formula is as follows: F = ω 纲层级 F 纲层级 + ω 目层级 F 目层级 + ω 结层级 F 结层级 ; where ω 纲层级 、ω 目层级 、ω 结层级 are the weight coefficients of the objective functions of each level of the main trunks, branches, and nodes respectively, and F 纲层级 、F 目层级 、F 结层级 are the objective functions of each level of the main trunks, branches, and nodes respectively.

[0175] The objective functions of each level of the main trunks, branches, and nodes in the upper layer are as follows:

[0176] (1)Main trunk level: F 纲层级 = min[max t∈[1,T] {Q(t)}] = min[max t∈[1,T] {∑ j=1 m Q j '(t) + Q j,区 (t)}];

[0177] Where F 纲层级 is the objective function of the main trunk level, Q(t) is the flow rate of the key section of the water conveyance line, t ∈ [1, T] is the range of the risk scheduling period, T is the total number of risk scheduling periods, m is the number of reservoirs hydraulically connected to the key section of the water conveyance line at the main trunk level of the upper layer, j is the index of the reservoir, Q j '(t) is the response flow rate of the j-th reservoir at the t-th period after flood routing to the key section of the water conveyance line, Qj,区 (t) is the sectional flow between the t-th period of the j-th reservoir and the key section of the water conveyance line.

[0178] (2) Objective level: F 目层级 = min[max t∈[1,T] {Q(t)'}] = min[max t∈[1,T] {∑ j=1 n Q j ''(t) + Q j,区 '(t)}];

[0179] In the formula, F 目层级 is the objective function of the objective level, Q(t)' is the flow rate of the water intake, Q j ''(t) is the response flow rate of the j-th reservoir at the t-th period after flood routing to the water intake, Q j,区 '(t) is the sectional flow between the t-th period of the j-th reservoir and the water intake, n is the number of reservoirs hydraulically connected to the water intake, and other variables are the same as above.

[0180] (3) Result level: F 结层级 = min{max t∈[1,T] ∑ j=1 n Z j (t)}; In the formula, F 结层级 is the objective function of the result level, Z j (t) is the water level of the j-th reservoir at the t-th period, n is the number of reservoirs hydraulically connected to the result level, and other variables are the same as above.

[0181] The objective functions of the class, objective, and result levels in the middle layer are as follows:

[0182] Class level: F 纲层级,1 = min[max t∈[1,T] {Q(t)}] = min[max t∈[1,T] {∑ j=1 m Q j '(t) + Q j,区 (t)}];

[0183] In the formula, F 纲层级,1 is the first objective function of the class level, m is the number of reservoirs hydraulically connected to the key section of the water conveyance line at the class level in the middle layer, and other variables are the same as above.

[0184] F 纲层级,2 = max{∑ j=1 J ∑ k=1 K ∑ t=1T (∑ i I b jk Q ijkt -∑ i I c ijk Q ijkt )}; where F 纲层级,2 is the second objective function at the outline level, b jk is the water supply benefit of the k-th water use department in the j-th sub-region, Q ijkt is the water supply volume from the i-th water supply source to the k-th water use department in the j-th sub-region at the t-th time period in the planned water year, c ijk is the water supply cost from the i-th water supply source to the k-th water use department in the j-th sub-region, J is the total number of sub-regions, j is the sub-region index, K is the total number of water use departments, k is the water use department index, I is the total number of water sources, i is the water source index, and other variables are the same as above.

[0185] (2) Target level:

[0186] F 目层级,1 = min[max t∈[1,T] {Q(t)'}] = min[max t∈[1,T] {∑ j=1 n Q j ''(t) + Q j,区 '(t)}]; where F 目层级,1 is the first objective function at the target level, and other variables are the same as above.

[0187] F 目层级,2 = min{∑ t=1 T ∑ k=1 K (D kt - S kt )}; S kt = ∑ i=1 N s ikt + ∑ j=1 M s jkt ;

[0188] Where F 目层级,2 is the second objective function at the target level, K is the number of socio-economic water demand areas, k is the socio-economic water demand area index, D kt is the total socio-economic water demand of the k-th water demand area at the t-th time period, S kt is the total socio-economic water supply of the k-th water demand area at the t-th time period, S kt ≤ D kt , s iktis the socioeconomic water supply supplied by the i-th pumping station to the k-th water demand area in the t-th period, N is the total number of pumping stations hydraulically connected to the target level, i is the pumping station index, sjkt is the socioeconomic water supply supplied by the j-th reservoir to the k-th water demand area in the t-th period, M is the total number of reservoirs hydraulically connected to the target level, and j is the reservoir index.

[0189] (3) Knot level:

[0190] F 结层级,1 =min{max t∈[1,T] ∑ j=1 n Z j (t)}; where F 结层级,1 is the first objective function of the knot level, and the other variables are the same as above.

[0191] F 结层级,2 =max t∈[1,T] (∑ j N ∑ t=1 T E j,t )=max t∈[1,T] (∑ j N ∑ t=1 T 9.8•λ•η•Q j,发电 (t)•Δh j,t •Δt j,t ) ;

[0192] In the formula, F 结层级,2 is the second objective function at the knot level, E j,t is the power generation of the jth reservoir in the tth period, λ is the dimension conversion coefficient, η is the turbine output coefficient of the hydropower station, Q j,发电 (t) is the flow rate through the turbine in the jth reservoir in the tth period, Δh j,t is the power generation head of the jth reservoir in the tth period, Δt j,t is the power generation time of the jth reservoir in the tth period, N is the number of reservoirs hydraulically connected to the node level, and the other variables are the same as above.

[0193] The objective functions of the lower levels of class, order, and knot are as follows:

[0194] (1) Level: F 纲层级 =max{∑ j=1 J ∑ k=1 K ∑ t=1 T (∑ i I b jk Qijkt -∑ i I c ijk Q ijkt )}; Wherein, the same as above.

[0195] (2) Item level: F 目层级 = min{∑ t=1 T ∑ k=1 K (D kt -S kt )}; Wherein, the same as above.

[0196] (3) Result level: F 结层级 = max t∈[1,T] (∑ j N ∑ t=1 T E j,t ) = max t∈[1,T] (∑ j N ∑ t=1 T 9.8 • λ • η • Q j,发电 (t) • Δh j,t • Δt j,t )}; Wherein, the same as above.

[0197] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for water resource risk dispatching and comprehensive regulation in a complex water network system, characterized in that: The steps include: Step S1: Collect basic data of the study area and identify the topological relationship and risk factor set of the complex water network system; Step S2: select an important risk factor set from the risk factor set, and establish the marginal distribution of the risk factors one by one; use the improved variational Bayesian network to establish the joint distribution; perform random sampling to generate a sample of the important risk factor set; Step S3, read the reservoir and pump station runoff data from the basic data, and use the cross-correlation function combined with the VMD method to obtain the runoff time-lag relationship; according to the runoff time-lag relationship and the topological relationship of the complex water network system, divide the complex water network system into three layers: upper, middle and lower, corresponding to short-term, medium-term and long-term responses respectively; Step S4, calling the three-layer nested water resource risk scheduling model of the complex water network system, taking the samples of the important risk factor set as input, obtaining and screening the scheduling plan set; wherein the water resource risk scheduling model includes constructing the objective functions of each layer of the water network system, the order, and the structure for the upper, middle, and lower layers respectively; Step S3 is specifically as follows: Based on the time-lag relationship between the runoff of adjacent pumping stations and reservoirs, the time-lag step is divided into three types of time steps: short-term, medium-term and long-term response. Combining the three types of time steps, the complex water network system is divided into three time response subsystems: short-term, medium-term and long-term. Calculate the network complexity of adjacent pumping stations and reservoirs based on the topological relationship of complex water network systems; Calculate the risk complexity of each node based on the risk factors of pumping stations and reservoirs; Combining network complexity and risk complexity, the complex water network system is divided into three spatial risk subsystems: high, medium and low; Establish the mapping relationship between each time response subsystem and the spatial risk subsystem, sort them, and determine the relative importance of water resource risk scheduling; Based on system security and economy, it is divided into three layers: upper, middle and lower; In step S3, the method for obtaining the runoff time-lag relationship is specifically as follows: The runoff data of each node including pumping stations and reservoirs are read, preprocessed, and decomposed using the VMD method, combined with the center frequency method to determine the number of IMFs and obtain the optimal IMFs; The cross-correlation function is calculated for the best IMFs of runoff data of adjacent nodes respectively, and the cross-correlation between the IMFs of runoff data of adjacent nodes is obtained. The peak value of the cross-correlation function of each pair of IMFs is taken as the time lag. The mutual correlation coefficient between each IMFs and the original signal is calculated, and the weight of each pair of IMFs is calculated accordingly. The time lag between adjacent nodes is determined by the weighted average method, and the runoff time lag relationship between each node in the system is obtained.

2. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 1 is characterized in that: In step S4, the steps of constructing the water resources risk scheduling model are specifically as follows: Determine the objective functions and weights of the upper, middle and lower layers one by one, and obtain the upper, middle and lower objective functions by weighting each layer; Based on the scheduling characteristics of each layer, the corresponding algorithm is used to calculate the optimal scheduling plan of each layer from the lower layer to the upper layer, and the optimal scheduling plan of the lower layer is used as the boundary condition of the middle layer, and the optimal scheduling plan of the middle layer is used as the boundary condition of the upper layer.

3. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 1 is characterized in that: The step S1 is further as follows: Step S11, determine the scope of the study area, and collect basic data on the complex water network system and water conservancy projects; Step S12, constructing structured data, using knowledge graph technology to identify the topological relationship of the complex water network system, and obtaining the generalized topological relationship of the complex water network system in the study area, wherein the topological relationship includes source-sink relationship, supply-demand relationship, regression relationship, and affiliation relationship; Step S13: Use expert consultation to identify the risk factor set existing in the study area.

4. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 1 is characterized in that: The step S2 is further as follows: Step S21, screening the important risk factor set; determining the most suitable prior distribution of each risk factor in the important risk factor set one by one; determining the inter-correlated and independent risk factors in the important risk factor set; Step S22: Based on the optimal prior distribution, a joint distribution is established for the inter-correlated risk factors using an improved variational Bayesian network; Step S23: for the joint and independent distribution of risk factors, samples are extracted using improved Latin hypercube sampling, and the samples are integrated to obtain the initial risk factor set samples; the samples are optimized using a genetic algorithm to obtain the final risk factor set.

5. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 4 is characterized in that: The step S22 is further as follows: Step S22a, using contrastive learning to perform feature compression on the risk factors that are mutually correlated in the important risk factor set, and extracting a low-dimensional dense representation; Step S22b, constructing an improved variational Bayesian network based on the optimal prior distribution, and using the low-dimensional dense representation as input, and quantifying the contribution of latent variables to risk factors through gradient weight analysis in the latent space of the variational autoencoder; Based on the contribution ranking, a progressive L1 sparse penalty is applied to the latent variables, the final latent variables are screened, and their posterior distribution is read to establish the joint distribution, where the Wasserstein distance is used as the difference measure of the joint distribution.

6. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 4 is characterized in that: The step S23 is further as follows: Step S23a, based on the curvature k(x) of the risk factor PDF, the joint and independent distributions are dynamically stratified, the interval width is adaptively adjusted according to the curvature, the high curvature area is compressed to the original width 1 / (1+k(x)), and the low curvature area is widened to n times; Step S23b: Generate 100% samples for each layer through Latin hypercube, extract x% samples from the non-tail area of ​​PDF; for the tail area, use importance sampling to obtain (100-x)% samples in the tail area, integrate joint and independent distribution sampling, and combine them into the initial sample of risk factors; Step S23c: using the initial sample of risk factors as input of the genetic algorithm for optimization to obtain a final risk factor set.

7. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 1 is characterized in that: The step S3 is further as follows: Step S31, read the runoff data of each node including the pump station and the reservoir, and perform preprocessing; use the cross-correlation function combined with the VMD method to calculate the runoff time lag relationship; Step S32: According to the runoff time-lag relationship between adjacent pumping stations and reservoirs and the topological relationship of the complex water network system, the entire complex water network system is divided into three layers: upper, middle and lower; the upper layer has a shorter response time and a higher risk, and adopts real-time scheduling; the middle layer has a medium response time and a moderate risk, and adopts short-term scheduling; the lower layer has a longer response time and a lower risk, and adopts medium- and long-term scheduling.

8. The method for water resource risk dispatching and comprehensive control of a complex water network system according to claim 1 is characterized in that: In step S4, the objective functions of each layer of the water network system, namely, the order, and the knot, are constructed for the upper, middle, and lower layers respectively, specifically: The upper layer considers safety. The objective function of the outline layer is to minimize the maximum water flow rate at the key section of the water transmission line, the objective function of the mesh layer is to minimize the maximum water flow rate at the water intake, and the objective function of the structure layer is to maximize the total flood control safety of the reservoir group. The middle layer considers economy and safety. The outline layer takes the minimum maximum water flow rate at the key section of the water transmission line and the maximum regional economic benefit as the objective function. The target layer takes the minimum maximum water flow rate at the water intake and the minimum total water shortage in the regional social economy as the objective function. The structure layer takes the maximum total flood control safety of the reservoir group and the maximum comprehensive power generation benefit of the reservoir group as the objective function. The lower layer considers economic efficiency. The overall layer takes the maximization of regional economic benefits as its objective function, the target layer takes the minimization of regional socio-economic total water shortage as its objective function, and the final layer takes the maximization of the comprehensive power generation benefits of the reservoir group as its objective function.

Images