A runoff scene reduction method and system based on improved Kernel Herding algorithm

By improving the Kernel Herding algorithm for runoff scene reduction, and utilizing the Weighted Generalized Kernel Herding algorithm (WGKH) and an improved step size, the problems of low computational efficiency and unstable results in the existing technology are solved, and efficient and accurate runoff scene reduction is achieved.

CN118350289BActive Publication Date: 2024-11-08HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410694231.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-31
Publication Date
2024-11-08
Estimated Expiration
2044-05-31

AI Technical Summary

Technical Problem

Existing methods for reducing runoff scenarios are computationally inefficient when the initial number of scenarios is large. Furthermore, traditional cluster analysis methods are highly sensitive to the initial cluster centers, and the iterative process is prone to getting trapped in local optima, leading to instability of the scenario set and loss of temporal features.

Method used

An improved Kernel Herding algorithm is used for runoff scene reduction. The distribution difference is measured by the maximum mean difference, and the weighted generalized Kernel Herding algorithm (WGKH) is used for minimization iteration during the iteration process. The control parameter α is adjusted to adjust representativeness and diversity, and the step size is changed to 2/(p+1) to improve the efficiency and accuracy of scene reduction.

Benefits of technology

It improves the efficiency and accuracy of runoff scenario reduction, avoids the cost of manual data annotation, enhances the versatility of the algorithm and the stability of the reduction results, and outperforms traditional methods in terms of time complexity and convergence speed, effectively handling large-scale runoff scenario reduction problems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118350289B_ABST
    Figure CN118350289B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of runoff scene reduction, and specifically discloses a runoff scene reduction method and system based on an improved Kernel Herding algorithm, which comprises: using the maximum mean difference to measure the distribution difference between the retained runoff scene set and the initial runoff scene set; using the Kernel Herding algorithm to iteratively minimize the distribution difference between the retained runoff scene set and the initial runoff scene set, and taking the first p The step size of the iteration is 2 / ( p +1); solve and obtain the corresponding retained runoff scene set when the distribution difference between the retained runoff scene set and the initial runoff scene set is the smallest, and complete the scene reduction of the initial runoff scene set. The present invention can realize accurate, reliable and universal runoff scene reduction, and further improve the efficiency and accuracy of runoff scene reduction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of runoff scene reduction, and more specifically, relates to a runoff scene reduction method and system based on an improved Kernel Herding algorithm. Background Art

[0002] Stochastic simulation of runoff can provide various possible water inflow scenarios for the formulation of reservoir operation and dispatching plans under uncertain conditions. These scenarios, which include but are different from the measured runoff sequences, can implicitly reflect the random characteristics of the runoff process and are the basic basis for water resources management work such as formulating reservoir dispatching rules and evaluating the risk-benefit of reservoir operation. In order to fully approximate the actual runoff situation, stochastic simulation usually needs to generate a large-scale scenario set to ensure the accurate characterization of runoff uncertainty. However, a large number of scenario sets will inevitably cause a huge computational burden, prolong the calculation time and reduce the calculation efficiency, and even cause the dimensionality curse problem for the optimal operation of reservoirs. Therefore, it is very necessary to carry out runoff scenario reduction work on the large-scale scenario set generated by random simulation to improve the calculation efficiency of reservoir operation.

[0003] At present, the existing scene reduction methods mainly include selection reduction method (synchronous back-substitution reduction method, fast forward selection method, etc.) and cluster analysis method (K-means, etc.). The selection reduction method achieves the purpose of scene reduction by selecting scenes one by one from the initial runoff scene set to retain or discard. This type of method has a clear concept, is simple and easy to use, and is widely used in the field of scene reduction. However, due to its nature of needing to traverse the entire scene, the computational efficiency of this type of method will drop significantly when the number of initial scenes is large. Cluster analysis method has good computational efficiency when solving large-scale scene sets and is widely used in scene reduction research. However, this type of method is highly sensitive to the initial cluster center, and the iterative process is prone to fall into local optimality; and it usually uses the mean to generate the cluster center and continuously iterate to obtain the retained runoff scene set, resulting in the retained scene set not necessarily being a subset of the initial runoff scene set, resulting in the loss of the temporal characteristics of the runoff scene.

[0004] Therefore, it is necessary to propose a more accurate, reliable and universal new method for runoff scene reduction to further improve the efficiency and accuracy of runoff scene reduction. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a runoff scene reduction method and system based on an improved KernelHerding algorithm, the purpose of which is to improve the efficiency and accuracy of runoff scene reduction.

[0006] To achieve the above object, according to a first aspect of the present invention, a runoff scene reduction method based on an improved Kernel Herding algorithm is proposed, comprising the following steps:

[0007] The maximum mean difference is used to measure the distribution difference between the retained runoff scenario set and the initial runoff scenario set. The kernel herding algorithm is used to iteratively minimize the distribution difference between the retained runoff scenario set and the initial runoff scenario set, and the first p The step size of the iteration is 2 / ( p +1); solve and obtain the corresponding retained runoff scenario set when the distribution difference between the retained runoff scenario set and the initial runoff scenario set is the smallest, and complete the scenario reduction of the initial runoff scenario set.

[0008] As a further preferred embodiment, a Kernel Herding algorithm is used to iteratively minimize the distribution difference between the retained runoff scene set and the initial runoff scene set, and the iterative formula is:

[0009] ;

[0010] in, m To retain the total number of scenarios in the runoff scenario set, , n is the total number of scenarios in the initial runoff scenario set, represents the initial runoff scenario set excluding the retained scenario set; x i , x l is the first i , l A runoff scenario, x j The first j runoff scenarios; is an iterative function; For the front p times, before p +The set of runoff scenarios retained after 1 iteration; i p+1 Based on The reserved p +1 index of runoff scenario; Based on The reserved p +1 runoff scene; k is a positive definite radial kernel function; α is a control parameter.

[0011] As a further preference, the positive definite radial kernel function adopts a Gaussian kernel function.

[0012] As a further preferred embodiment, the bandwidth parameter in the Gaussian kernel function for:

[0013] ;

[0014] Among them, Median represents the median value, X n represents the initial runoff scenario set, x , y Represents the runoff scenario in the initial runoff scenario set.

[0015] As a further preferred embodiment, the value range of the control parameter α is: .

[0016] As a further preference, the initial runoff scenario set is generated by performing random runoff simulation on historical runoff data of the watershed.

[0017] According to a second aspect of the present invention, a runoff scene reduction system based on an improved Kernel Herding algorithm is provided, comprising a processor, wherein the processor is used to execute the above-mentioned runoff scene reduction method based on the improved Kernel Herding algorithm.

[0018] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the runoff scene reduction method based on the improved Kernel Herding algorithm is implemented.

[0019] In general, the above technical solution conceived by the present invention has the following technical advantages compared with the prior art:

[0020] 1. The present invention innovatively introduces the unsupervised learning method Kernel Herding algorithm into the field of runoff scene reduction. The algorithm does not require label information and only performs scene reduction based on the distribution characteristics of the data itself, avoiding the cost and complexity of manually annotating data. It is improved by modifying the iteration step size during iterative solution, which can solve the equal probability problem faced by the traditional Kernel Herding algorithm in the process of scene reduction while ensuring that the time complexity of the algorithm remains unchanged. At the same time, the convergence speed is improved, thereby improving the efficiency and accuracy of runoff scene reduction.

[0021] 2. The traditional Kernel Herding algorithm directly uses the maximum mean difference (MMD) for distribution measurement, while the present invention proposes an improved weighted generalized Kernel Herding algorithm, which specifically uses α-MMD to measure the distribution of the retained runoff scene set and the initial runoff scene set. By controlling the numerical value of the built-in parameter α, it can effectively achieve a balance between representativeness and diversity of the retained scenes relative to the initial runoff scene set.

[0022] 3. The weighted generalized Kernel Herding algorithm of the present invention is superior to traditional methods such as selection reduction method (SBR, FFR), cluster analysis method, etc. in terms of time complexity and convergence speed, and its runoff scene reduction result has better performance in energy score index compared with traditional methods, and can efficiently handle the runoff scene reduction problem.

[0023] 4. The weighted generalized Kernel Herding algorithm of the present invention can be regarded as a greedy algorithm in essence. The retention strategy of the algorithm is similar to the greedy selection step in the greedy algorithm, that is, each step selects the local optimal solution, hoping to eventually obtain the global optimal solution. The algorithm gradually selects the retention scene in the direction of minimizing MMD, and its reduction result is unique, which can effectively avoid the defects of cluster analysis method such as sensitivity to initial values ​​and unstable results, and enhances the versatility of the runoff scene reduction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a runoff scenario reduction method based on an improved Kernel Herding algorithm according to an embodiment of the present invention.

[0025] Figure 2 This is an initial runoff scene set generated by GMM in an embodiment of the present invention.

[0026] Figure 3 This is a two-dimensional effect diagram of the runoff scene reduction of the WGKH algorithm in an embodiment of the present invention.

[0027] Figure 4 This is a three-dimensional rendering of the runoff scene reduction using the WGKH algorithm in an embodiment of the present invention.

[0028] Figure 5 It is a corresponding scene probability map of the runoff scene reduction of the WGKH algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] The conventional Kernel Herding algorithm is an unsupervised learning method based on kernel selection. It uses the convexity of MMD to construct a continuous convex optimization problem and iterates the process of minimizing the distribution difference between data. The present invention applies the Kernel Herding algorithm to runoff scenario reduction, improves it, and proposes a weighted generalized Kernel Herding algorithm (WGKH).

[0031] Specifically, an embodiment of the present invention provides a runoff scene reduction method based on an improved Kernel Herding algorithm, which performs runoff scene reduction through WGKH, such as Figure 1 As shown, the following steps are included:

[0032] S1: Obtain the initial runoff scenario set and determine the number of reduced scenarios.

[0033] Specifically, the runoff random simulation is performed on the historical runoff data of the basin to generate n A large-scale initial runoff scenario set Reducing the number of scenarios means selecting representative and diverse typical scenarios from the initial large-scale runoff scenario set to form a set of retained runoff scenarios. The scale is preset as m .

[0034] S2: The maximum mean difference is used to measure the distribution difference between the retained runoff scenario set and the initial runoff scenario set.

[0035] Furthermore, before performing the maximum mean difference metric, the positive definite radial kernel function to be used needs to be determined in advance. Positive definite characteristic radial kernel function For mapping the distribution into the Reproducing Kernel Hilbert Space (RKHS), the present invention preferably uses a Gaussian kernel function as a positive definite radial kernel function:

[0036] (1)

[0037] The bandwidth parameter is a built-in parameter of the Gaussian kernel function. It is preferably set to the median value of the squared Euclidean distance between the initial runoff scene sets. The median value can be regarded as a compromise choice for extreme cases and is robust. Specifically:

[0038] (2)

[0039] Specifically, the maximum mean difference (MMD) is a method to measure the difference between two distributions in the RKHS function space. The definition of MMD is as follows:

[0040] set up P , Q yes Two Borel probability measures on . Assume f From its regenerative nucleus k The associated reproducing kernel Hilbert space The unit ball sampling in , that is, if , then the MMD between distributions P and Q is defined as:

[0041] (3)

[0042] In the formula, and All are independent samples.

[0043] The meaning of this formula is to find a mapping function f , this mapping function f The variables can be mapped to a high-dimensional space, and then the expected difference between the two distributed random variables after mapping is found. This difference is the Mean Discrepancy, and then the supremum of the Mean Discrepancy is found. This maximum value is the MMD.

[0044] An empirical version of the MMD can be derived that is able to measure the set of retained runoff scenarios To the initial runoff scenario set The distance between and The constructed empirical distribution replaces the distribution in formula (3) P and Q :

[0045] (4)

[0046] Furthermore, the present invention designs a generalized maximum mean difference (Alpha - Maximum Mean Discrepancy, α-MMD) to measure distribution difference, which is an improvement on MMD. It measures the representativeness and diversity of the selected data by re-weighting each item in equations (3) and (4), and the preference for the diversity of the selected scenes is determined by a single parameter α.

[0047] According to the definition of formula (3), the distribution P and Q The α-MMD between is defined as:

[0048] (5)

[0049] For the retained runoff scenario set and the initial runoff scenario set:

[0050] (6)

[0051] Specifically, the larger the built-in control parameter α is, the more the runoff scenario set is retained. The weaker the diversity, the stronger the representativeness; conversely, the stronger the diversity, the weaker the representativeness; the following is a specific explanation.

[0052] definition :

[0053] (7)

[0054] According to the kernel function k The similarity criterion is used Evaluate The self-similarity of Evaluate and The former can be interpreted as a quantification of the diversity of the selected scenes. Formula (6) can be rewritten as:

[0055] (8)

[0056] The ability of α-MMD to quantify representativeness and diversity is explained from the perspective of regularization. By definition, Considered as a constant term, when minimizing α-MMD, it is equivalent to minimizing :

[0057] (9)

[0058] Among them, (*) is the objective function of minimizing MMD, which encourages sampling in high-quality areas under the target distribution; (**) is a diversity regularizer that discourages sampling at points close to existing samples, and the built-in parameter α controls the trade-off between representativeness and diversity in the regularization. More specifically, a larger value of α means a greater penalty on diversity. Therefore, minimizing the α-MMD between the retained runoff scene set and the initial runoff scene set is the correct strategy to maintain both representativeness and diversity.

[0059] S3: The improved Kernel Herding algorithm is used to iteratively minimize the distribution difference, select the required number of scene sets, obtain the retained runoff scene set, and complete the scene reduction.

[0060] Specifically, the steps include:

[0061] S31: Using the convexity of MMD, minimize Considered as a continuous convex optimization problem, according to formula (6), formula (10) can be derived, which is the iterative formula of the improved Kernel Herding algorithm:

[0062] (10)

[0063] in, , m To retain the total number of scenarios in the runoff scenario set, n is the total number of scenarios in the initial runoff scenario set, represents the initial runoff scenario set excluding the retained scenario set; x i , x l is the first i , l A runoff scenario, x j The first j runoff scenarios; is an iterative function; For the front p times, before p +The set of runoff scenarios retained after 1 iteration; i p+1 Based on The reserved p +1 index of runoff scenario; Based on The reserved p +1 runoff scene.

[0064] S32: Based on the above iterative formula, the minimization process is iterated to obtain the initial runoff scenario set. As input, the following steps are involved:

[0065] S321: Initial Season p =1, step length ;

[0066] S322: To facilitate calculation, transform equation (10) to ,calculate ;

[0067] S323: Calculate the retention scenario :

[0068] (11)

[0069] S324: For any , update the process kernel variables involved in equation (10) :

[0070] (12)

[0071] S325: , step length ;Each time a reserved scenario is selected, it is deleted from the initial runoff scenario set;

[0072] S326: Repeat steps S323 to S325 until the number of retained scenes reaches the reduced scene size m , output the retained runoff scenario set .

[0073] In the above iterative process, the step size in the conventional Kernel Herding algorithm is 1 / p , which will result in the retained scene set being uniformly probable; the present invention sets the step length to 2 / ( p +1) to improve the equal probability problem faced during scenario reduction.

[0074] The following are specific embodiments:

[0075] Taking the monthly historical runoff data of Geheyan Reservoir in Qingjiang River Basin from 1951 to 2022 as the implementation object, the Gaussian mixture model was used to perform random runoff simulation on the historical data to generate initial runoff scenario sets with scales of 1000 and 10000 years.

[0076] The WGKH proposed in this invention is used to reduce the initial runoff scene set, and the reduction results are compared and analyzed with those of the K-means clustering method, the simultaneous back-reduction method (SBR), and the fast forward selection method (FFS). The reduced runoff scene is required to accurately reflect the probability distribution characteristics of historical runoff and be related to the actual scene; considering the probability distribution characteristics of the scene, the energy score index is used to measure the scene quality.

[0077] The Energy Score (ES) indicator is used to measure the probability distribution characteristics of the runoff scenario. The ES indicator is a measure of the difference between the cumulative distribution function of the reduced runoff scenario and the actual runoff scenario. It is a negative score. The smaller the value, the closer the two are, that is, the reduced scenario is more realistic. See the following formula for details:

[0078] (13)

[0079] Where: Represents historical runoff t Moment mean; Representing scenes i , j probability; Representing scenes i , j of t Time flow, N is the total number of scenes.

[0080] like Figure 2 As shown in Figure 1, an initial runoff scenario set with a scale of 1000 years is generated using GMM.

[0081] Table 1 shows the iteration time required for runoff scenario reduction by the WGKH method, K-means clustering method, synchronous back-reduction method (SBR) and fast forward selection method (FFS) of the present invention when different numbers of retained scenarios are set (the initial runoff scenario set size is 1000 years, and the average level of 10 iterations is taken); Table 2 shows the energy scores of the method, K-means clustering method, synchronous back-reduction method (SBR) and fast forward selection method (FFS) of the present invention when the number of retained scenarios is 10, 20 and 50 respectively. It can be seen from Tables 1 and 2 that the present invention is superior to other traditional methods in terms of convergence speed and reduced scene quality, and can efficiently handle the runoff scenario reduction problem;

[0082] ;

[0083] ;

[0084] Figures 3 to 5The two-dimensional and three-dimensional effect diagrams of runoff scene reduction and the probability diagrams of the corresponding scenes of the method of the present invention when the number of retained scenes is 10 are given. It can be seen that the present invention modifies the iteration step size of the traditional Kernel Herding algorithm so that the proposed WGKH algorithm can obtain the corresponding probability of the retained scenes in the scene reduction field, solving the equal probability problem faced by the traditional algorithm in the scene reduction process. Table 3 shows the iteration time required for runoff scene reduction of the present invention and the traditional Kernel Herding algorithm when setting different numbers of retained scenes (the initial runoff scene set size is 10,000 years, and the average level of 25 iterations is taken). It can be seen from Table 3 that the convergence speed of the present invention is improved compared with the traditional algorithm;

[0085] .

[0086] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A runoff scene reduction method based on an improved Kernel Herding algorithm, characterized in that: The steps include: The generalized maximum mean difference α-MMD is used to measure the set of retained runoff scenarios. With the initial runoff scenario set The distribution difference between The Kernel Herding algorithm is used to iteratively minimize the distribution difference between the retained runoff scenario set and the initial runoff scenario set, and the step length of the pth iteration is 2 / (p+1) during the iterative process. The specific steps are as follows: The iterative formula is derived based on the distribution difference: Where m is the total number of scenes in the retained runoff scene set, p∈{1,2,…,m}, n is the total number of scenes in the initial runoff scene set, represents the initial runoff scenario set excluding the retained scenario set; x i 、x l is the i-th and l-th runoff scenario in the initial runoff scenario set, x j is the jth runoff scenario in the retained runoff scenario set; is an iterative function; is the set of runoff scenarios retained after the first p and first p+1 iterations; i p+1 Based on The index of the p+1th runoff scenario retained; Based on The p+1th runoff scenario is retained; k is a positive definite radial kernel function; α is a control parameter; Based on the above iterative formula, the minimization process is iterated and the initial runoff scenario set X n As input, the following steps are involved: (1) Initially, p = 1, step size β1 = 1, S0 = 0, (2) For any X i ∈X n ,calculate (3) Calculate the reserved scene i p : (4) For any Update process kernel variable S p (x i ): (5) Step size β p =2 / (p+1); each time a reserved scenario is selected, it is deleted from the initial runoff scenario set; (6) Repeat steps (3) to (5) until the number of retained scenarios reaches the reduced scenario size m, and output the retained runoff scenario set In this way, the retained runoff scenario set corresponding to the minimum distribution difference between the retained runoff scenario set and the initial runoff scenario set is solved, thereby completing the scenario reduction of the initial runoff scenario set.

2. The runoff scene reduction method based on the improved Kernel Herding algorithm according to claim 1, characterized in that: The positive definite radial kernel function adopts a Gaussian kernel function.

3. The runoff scene reduction method based on the improved Kernel Herding algorithm according to claim 2, characterized in that: The bandwidth parameter σ in the Gaussian kernel function is: σ=Mean({||xy||2∣x,y∈X n }) Among them, Median represents the median value, X n represents the initial runoff scenario set, and x and y represent the runoff scenarios in the initial runoff scenario set.

4. The runoff scene reduction method based on the improved Kernel Herding algorithm according to claim 1, characterized in that: The value range of the control parameter α is:

5. The runoff scene reduction method based on the improved Kernel Herding algorithm according to any one of claims 1 to 4, characterized in that: The initial runoff scenario set is generated by performing random runoff simulation on the historical runoff data of the basin.

6. A runoff scene reduction system based on an improved Kernel Herding algorithm, characterized in that: It includes a processor, which is used to execute the runoff scene reduction method based on the improved Kernel Herding algorithm as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the runoff scene reduction method based on the improved Kernel Herding algorithm as described in any one of claims 1 to 5 is implemented.

8. A computer program product, characterized in that It comprises a computer program, which, when executed by a processor, implements the runoff scene reduction method based on the improved Kernel Herding algorithm as described in any one of claims 1 to 5.