A distribution network partition collaborative optimization control method and system

Abstract

The present invention discloses a distribution network partition collaborative optimization control method and system, the method comprising: obtaining electrical distance and operating characteristics, and partitioning the distribution network according to the electrical distance and operating characteristics by using a cluster analysis algorithm; obtaining local data of each partition, and for each partition, building a local model by using the local data; using a horizontal federated learning architecture, updating local model parameters based on the local data of each partition, introducing differential privacy technology and aggregating the local model parameters of each partition to update the global model; optimizing the global model by a collaborative optimization control algorithm to obtain a first global model, generating a local control strategy according to the first global model, and performing coordinated execution, and improving the scalability and robustness of the system through distributed execution and coordination mechanisms, and being able to better cope with dynamic changes in the distribution network.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network, and in particular to a distribution network partition collaborative optimization control method and system. Background Art

[0002] With the large-scale access of renewable energy and distributed resources, the structure and operation characteristics of distribution networks are becoming increasingly complex. Traditional centralized control methods face problems such as high computational complexity, insufficient data privacy protection, and poor scalability. Although distributed control methods have solved these problems to a certain extent, it is difficult to achieve global optimal control due to limited information sharing. In addition, existing distributed control methods often ignore the partitioning characteristics of distribution networks and find it difficult to fully utilize the differentiated resources and operation characteristics of different regions.

[0003] In response to the problems existing in the prior art, this patent realizes coordinated optimization control of distribution network partitions while protecting data privacy; fully utilizes the characteristics of distribution network partitions to improve the efficiency and accuracy of optimization control; and constructs a global optimization model in a distributed environment to improve the overall operating efficiency of the distribution network and its ability to carry distributed resources. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a distribution network partition collaborative optimization control method and system to solve the problem that the existing distributed control methods often ignore the partition characteristics of the distribution network and find it difficult to fully utilize the differentiated resources and operating characteristics of different regions.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a distribution network partition collaborative optimization control method, comprising:

[0008] Acquire electrical distance and operation characteristics, and partition the distribution network using a cluster analysis algorithm according to the electrical distance and operation characteristics;

[0009] Get the local data of each partition, and build a local model for each partition using the local data.

[0010] Using a horizontal federated learning architecture, local model parameters are updated based on the local data of each partition, differential privacy technology is introduced, and local model parameters of each partition are aggregated to update the global model;

[0011] The global model is optimized by a collaborative optimization control algorithm to obtain a first global model, and a local control strategy is generated according to the first global model and coordinatedly executed.

[0012] As a preferred solution of the distribution network partition collaborative optimization control method of the present invention, the distribution network partitioning by using the cluster analysis algorithm includes:

[0013] The electrical distance between nodes is defined as the inverse of the absolute value of the admittance between nodes;

[0014] The square of the electrical distance between nodes is divided by twice the square of the bandwidth parameter of the Gaussian kernel function, the result is negative, and the similarity matrix is ​​obtained using the exponential function;

[0015] Subtract the diagonal matrix from the similarity matrix to get the Laplacian matrix;

[0016] The Laplacian matrix and the corresponding eigenvector are equal to the product of the eigenvalue of the Laplacian matrix and the diagonal matrix and the corresponding eigenvector to obtain the eigenvalue;

[0017] The eigenvector corresponding to the smallest non-zero eigenvalue is selected to form a first matrix, and the row vectors of the first matrix are clustered to obtain the partition of the distribution network.

[0018] As a preferred solution of the distribution network partition collaborative optimization control method of the present invention, wherein: updating the local model parameters based on the local data of each partition includes:

[0019] Define global model parameters and local model parameters for each partition;

[0020] The local model parameters of each partition after iteration are the difference between the local model parameters of the partition in the previous iteration and the product of the learning rate and the gradient of the loss function of the partition to the local model parameters in the previous iteration;

[0021] The global model parameters are obtained by summing the products of the local model parameters of each partition and the corresponding number of samples and dividing by the total number of samples.

[0022] As a preferred solution of the distribution network partition collaborative optimization control method described in the present invention, the differential privacy technology includes:

[0023] During the aggregation process, the noise added to the local model parameters of each partition is the sum of the local model parameters and Gaussian noise after iteration for each partition;

[0024] Gaussian noise includes mean and variance, and the variance is expressed as:

[0025]

[0026] Among them, δ is the privacy failure probability, ε is the preset privacy budget, and Δf is the sensitivity.

[0027] As a preferred solution of the distribution network partition collaborative optimization control method described in the present invention, the global model includes:

[0028] Through iterative training and model aggregation, a global model is constructed and the global loss function is defined as:

[0029]

[0030] The optimization goal is to minimize the global loss function, which is expressed as:

[0031] w * =argmin w F(w)

[0032] Among them, n k is the number of samples in the kth partition, n is the total number of samples, w is the global model parameter, F k is the loss function of the kth partition, and arg min is the parameter of the minimum value.

[0033] As a preferred solution of the distribution network partition collaborative optimization control method of the present invention, wherein: optimizing the global model by means of a collaborative optimization control algorithm comprises:

[0034] Initialize global model parameters;

[0035] For each round of communication t=1,2,…T, the central server broadcasts the current global model to all partitions;

[0036] Each partition performs E rounds of local updates;

[0037] Each partition sends the updated model parameters to the central server;

[0038] The central server aggregates model parameters and outputs a first global model.

[0039] As a preferred solution of the distribution network partition collaborative optimization control method described in the present invention, the local control strategy includes:

[0040] The central server sends the global model to each partition;

[0041] Each partition generates a control strategy based on the global model and local data, which is expressed as:

[0042] u k =g k (x k ,w * )

[0043] Among them, x k is the state variable of partition k, g k is the control strategy generation function, w* is the global model;

[0044] Each partition executes local control strategies, and adjacent partitions exchange boundary information for coordination.

[0045] In a second aspect, the present invention provides a distribution network partition collaborative optimization control system, comprising:

[0046] A partitioning module, used for obtaining electrical distance and operation characteristics, and partitioning the distribution network using a clustering analysis algorithm according to the electrical distance and operation characteristics;

[0047] A model building module is used to obtain local data of each partition and build a local model for each partition using the local data;

[0048] An aggregation module, used to update local model parameters based on the local data of each partition by using a horizontal federated learning architecture, introduce differential privacy technology and aggregate the local model parameters of each partition to update the global model;

[0049] The optimization module is used to optimize the global model through a collaborative optimization control algorithm to obtain a first global model, generate a local control strategy according to the first global model, and perform coordinated execution.

[0050] In a third aspect, the present invention provides a computing device, comprising:

[0051] Memory and processor;

[0052] The memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the distribution network partition collaborative optimization control method are implemented.

[0053] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the distribution network partition collaborative optimization control method.

[0054] Compared with the prior art, the present invention has the following beneficial effects: the present invention realizes global optimization in a distributed environment through distribution network partitioning and a federated learning framework, effectively balances the needs of local autonomy and global coordination, adopts differential privacy technology, realizes effective aggregation of model parameters while protecting sensitive data of each partition, solves the problem of data privacy protection, makes full use of the distribution network partitioning characteristics, improves the efficiency and accuracy of optimization control, is suitable for large-scale distribution network systems, improves the scalability and robustness of the system through distributed execution and coordination mechanisms, and can better cope with the dynamic changes of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0056] Figure 1 The figure is a schematic diagram of the overall process logic of the distribution network partition collaborative optimization control method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0058] Example 1

[0059] Reference Figure 1 , is an embodiment of the present invention, and provides a distribution network partition collaborative optimization control method, comprising:

[0060] S100: obtaining electrical distance and operation characteristics, and partitioning the distribution network using a cluster analysis algorithm according to the electrical distance and the operation characteristics;

[0061] S200: Acquire local data of each partition, and build a local model for each partition using the local data;

[0062] Specifically, with the goal of minimizing operating costs and network losses, the local model is expressed as:

[0063]

[0064] st

[0065]

[0066] V min ≤V i,t ≤V max

[0067] P i,min ≤P i,t ≤P i,max

[0068] O i,min ≤Q i,t ≤Q i,max

[0069] |I l,t |≤I l,max

[0070] Among them, P i,t , Q i,t are the active and reactive outputs of the i-th generating unit at time t, P L,j,t is the active power demand of the jth load at time t, V i,t is the voltage amplitude of node i at time t, I l,t is the current of line l at time t, C i is the power generation cost coefficient of the ith power generation unit, λ is the weight coefficient of network loss cost and power generation cost, N is the total number of power generation units, L is the total number of lines, T is the total number of time periods, R l is the resistance of the lth circuit, I l,t is the current of the lth line at time t, M is the total number of loads, P loss,t is the network loss at time t, V min 、V max is the minimum and maximum voltage of node i at time t, P i,min , P i,max is the minimum and maximum active output of the ith generating unit, O i,min , Q i,max is the minimum and maximum reactive power output of the ith generating unit, I l,max is the maximum current of the lth line.

[0071] S300: Using the horizontal federated learning architecture, the local model parameters are updated based on the local data of each partition, and the differential privacy technology is introduced to aggregate the local model parameters of each partition to update the global model;

[0072] S400: Optimizing the global model through a collaborative optimization control algorithm to obtain a first global model, generating a local control strategy according to the first global model, and performing coordinated execution.

[0073] It should be noted that compared with the traditional centralized control method, the present application reduces the computational complexity, makes the control of the distribution network more efficient, improves the overall operating efficiency of the distribution network and its ability to carry distributed resources, and enhances the scalability and robustness of the system.

[0074] In the embodiment of the present application, the above step S100 includes the following sub-steps A1-A5;

[0075] In A1: the electrical distance between nodes is defined as the reciprocal of the absolute value of the admittance between the nodes;

[0076] In A2: the square of the electrical distance between nodes is divided by twice the square of the bandwidth parameter of the Gaussian kernel function, the result is negative, and the similarity matrix is ​​obtained using the exponential function;

[0077] In A3: the Laplace matrix is ​​obtained by subtracting the diagonal matrix from the similarity matrix;

[0078] In A4: The Laplacian matrix and the corresponding eigenvector are equal to the product of the eigenvalue of the Laplacian matrix and the diagonal matrix and the corresponding eigenvector to obtain the eigenvalue;

[0079] In A5: the eigenvector corresponding to the smallest non-zero eigenvalue is selected to form a first matrix, and the row vectors of the first matrix are clustered to obtain the partitions of the distribution network.

[0080] In an optional embodiment, the cluster analysis algorithm for distribution network partitioning may be a spectral clustering algorithm, which represents data points as nodes in a graph, represents the similarity between nodes as edge weights, and performs clustering by analyzing the eigenvalues ​​and eigenvectors of the Laplace matrix of the graph;

[0081] The cluster analysis algorithm for distribution network partitioning can also be a hierarchical clustering algorithm, which builds a hierarchical tree from bottom to top (agglomerative) or top to bottom (divisive), and partitions based on the similarity between nodes defined by electrical distance;

[0082] It is also possible to use the K-medoids algorithm similar to K-means, using actual data points as cluster centers and performing cluster partitioning based on electrical distance;

[0083] It can also be said that the fuzzy clustering algorithm allows nodes to belong to multiple partitions at the same time, which is suitable for handling the situation where the partition boundaries are fuzzy and conforms to the overlapping characteristics of the actual network.

[0084] In a preferred embodiment, the present application uses a spectral clustering algorithm to partition the distribution network;

[0085] Specifically, the electrical distance between nodes i and j is defined as:

[0086]

[0087] Among them, Y ij is the admittance between nodes i and j;

[0088] Construct similarity S, whose elements are:

[0089]

[0090] Where, σ is the bandwidth parameter of the Gaussian kernel function;

[0091] Calculate the Laplace matrix L, expressed as: L = DS;

[0092] Where D is a diagonal matrix,

[0093] The generalized eigenvalue is solved as: Lv = λDv, where v is the eigenvector and λ is the eigenvalue;

[0094] The eigenvectors corresponding to the k smallest non-zero eigenvalues ​​are selected to form the first matrix, namely, the matrix V. The row vectors of the matrix V are clustered using k-means to obtain k partitions of the distribution network.

[0095] It should be noted that the spectral clustering algorithm can directly reflect the electrical correlation between nodes through electrical distance, takes into account the topological structure of the entire network, and tends to produce partitions of relatively balanced size; the hierarchical clustering algorithm can intuitively display the hierarchical relationship of the partitions, but the computational complexity is relatively high, and it is suitable for small and medium-sized networks; the K-medoids algorithm is insensitive to outliers and the results are more stable, but the computational overhead is large and the convergence speed is slow; the partitioning results of the fuzzy clustering algorithm are less interpretable, and the partitioning method using the spectral clustering algorithm can better reflect the actual electrical correlation than the traditional geographical location or topological structure partitioning.

[0096] In the embodiment of the present application, the above step S300 includes the following sub-steps B1-B3:

[0097] In B1: define global model parameters and local model parameters for each partition;

[0098] In B2: the local model parameters of each partition after iteration are the difference between the local model parameters of the partition in the previous iteration and the product of the learning rate and the gradient of the loss function of the partition to the local model parameters in the previous iteration;

[0099] In B3: The product of the local model parameters of each partition and the corresponding number of samples is summed and divided by the total number of samples to obtain the global model parameters.

[0100] Specifically, define the global model parameter as w and the local model parameter of the kth partition as w k ;

[0101] In each iteration, each partition updates the model parameters based on local data as follows:

[0102]

[0103] Among them, η is the learning rate, F k is the loss function of the kth partition, is the local model parameter of the kth partition at the tth iteration;

[0104] The model parameters of the central server aggregating each partition are expressed as:

[0105]

[0106] Among them, n k is the number of samples in the kth partition, and n is the total number of samples.

[0107] It should be noted that horizontal federated learning is a federated learning architecture that is suitable for scenarios where the business types of each data holder are similar, the user characteristics obtained are similar, but the user space has little overlap or basically no overlap. In horizontal federated learning, the feature dimension of the data is the guide, and the parts with the same characteristics of the participants but different users are taken out for joint training. In this way, the sample space for training can be expanded, thereby improving the accuracy and generalization ability of the model. In horizontal federated learning, each participant uses local data to train the model and uploads the model parameters or gradients to the central server in encrypted form. The central server aggregates all model gradients, and then encrypts the aggregated global model parameters and transmits them back to each participant. In this process, all interactions of model parameters are encrypted to ensure data security and privacy.

[0108] The horizontal federated learning architecture is adopted, which is suitable for the characteristics of distribution network partitions with similar characteristics but less overlap in user space. The parameter update and aggregation mechanism take into account the weight of the number of samples. This design not only protects data privacy, but also makes full use of distributed data for model training.

[0109] In the embodiment of the present application, after completing steps B1-B3 in the above step S300, the differential privacy technology further includes the following sub-steps B4-B5;

[0110] In B4: during the aggregation process, adding noise to the local model parameters of each partition is the sum of the local model parameters and Gaussian noise after iteration for each partition;

[0111] In B5: Gaussian noise includes mean and variance, and the variance is expressed as:

[0112]

[0113] Among them, δ is the privacy failure probability, ε is the preset privacy budget, and Δf is the sensitivity.

[0114] Specifically, the Laplace mechanism is used to add noise to the model parameters of each partition during the parameter aggregation process as follows:

[0115]

[0116] Among them, N(0,σ 2 ) has a mean of 0 and a variance of σ 2 Gaussian noise;

[0117] The noise variance is determined according to the preset privacy budget and sensitivity.

[0118] It should be noted that differential privacy is a powerful privacy protection technology that aims to protect personal data from being traced back to individual users. This technology adds a certain amount of random noise during data processing to ensure that even if the data set contains sensitive information, attackers cannot identify specific individuals through analysis results. The core idea of ​​differential privacy is to ensure that changes in a single record in the database will not have a significant impact on the query results, thereby protecting personal privacy. The implementation of differential privacy usually relies on two main mechanisms: the Laplace mechanism and the exponential mechanism.

[0119] The Laplace mechanism is applicable to the protection of numerical results and achieves differential privacy protection by adding random noise that follows the Laplace distribution to the query results.

[0120] The exponential mechanism is suitable for the protection of discrete results. It selects an output value so that the quality of the output value is proportional to the evaluation result of an availability function, while ensuring that the selection process meets the requirements of differential privacy.

[0121] This application balances privacy protection and model performance by presetting the privacy budget and failure probability. This mechanism ensures the availability of the model while protecting sensitive data.

[0122] In the embodiment of the present application, the global model is optimized in the above step S400, including the following sub-steps C1-C2;

[0123] In C1: Through iterative training and model aggregation, a global model is constructed and the global loss function is defined as:

[0124]

[0125] In C2: The optimization goal is to minimize the global loss function expressed as:

[0126] w * =argmin w F(w)

[0127] Among them, n k is the number of samples in the kth partition, n is the total number of samples, w is the global model parameter, F k is the loss function of the kth partition, and arg min is the parameter of the minimum value.

[0128] In the embodiment of the present application, after completing steps C1-C2 in the above step S400, the following sub-steps C3-C7 are also included:

[0129] In C3: Initialize global model parameters;

[0130] In C4: For each round of communication t=1,2,…T, the central server broadcasts the current global model to all partitions;

[0131] In C5: Each partition performs E rounds of local updates;

[0132] In C6: each partition sends the updated model parameters to the central server;

[0133] In C7: the central server aggregates model parameters and outputs a first global model.

[0134] In an optional embodiment, the distribution network partition collaborative optimization control algorithm may be a federated average algorithm, which performs distributed machine learning and allows multiple clients to collaborate in training models without the need to centralize data to a central location;

[0135] The distribution network partition collaborative optimization control algorithm can also be a distributed gradient descent algorithm, where each node calculates the gradient based on local data and exchanges gradient information with adjacent nodes to update model parameters;

[0136] It can also be a decentralized alternating direction multiplier method, by introducing an augmented Lagrangian function to coordinate the solutions of each sub-problem;

[0137] It can also be a dual decomposition algorithm, which converts the original problem into a dual problem through Lagrangian dualization, and then solves the dual problem;

[0138] In a preferred embodiment, the present application uses a federated algorithm for optimization;

[0139] Specifically, initialize the global model parameter w0;

[0140] For each round of communication t=1,2,…T:

[0141] The central server broadcasts the current global model w to all partitions t ;

[0142] Each partition k performs E rounds of local updates as follows:

[0143]

[0144] Each partition k will have the updated model parameters Send to the central server;

[0145] The central server aggregation model parameters are:

[0146]

[0147] Output the final global model, i.e. the first global model w T;

[0148] It should be noted that the optimization using the federated average algorithm aims to achieve distributed machine learning, allowing multiple clients (such as mobile devices or distributed servers) to collaborate on training models without centralizing the data to a central location. It uses data distributed across different clients while protecting user privacy; the distributed gradient descent algorithm is simple to implement, has low computational overhead, and converges quickly. However, due to the need to frequently exchange gradient information, it has high requirements for communication quality and is insufficient in data privacy protection; the distributed alternating direction multiplier method has good convergence and can handle non-convex optimization problems. However, it has high computational complexity, requires a central coordination node, and has high communication overhead; the dual decomposition algorithm is suitable for handling large-scale problems, but has a slow convergence speed, is more sensitive to the selection of initial values, and has insufficient privacy protection.

[0149] This application uses a federated learning algorithm for model optimization. Federated learning is a distributed machine learning framework that allows multiple parties to collaboratively train models without sharing original data. It achieves data privacy protection through local training and model aggregation, and can be combined with differential privacy technology to provide quantifiable privacy protection. It has high communication efficiency and only requires the exchange of model parameters, and has good model performance and system scalability.

[0150] While protecting data privacy, the federated learning algorithm achieves effective use of global information through model parameter aggregation, and has significant advantages in communication efficiency, model performance, and system scalability. This algorithm not only solves the data privacy problem in traditional distributed algorithms, but also provides a quantifiable level of privacy protection through differential privacy mechanisms, while ensuring that the control performance reaches a level close to the global optimal level.

[0151] In the embodiment of the present application, after completing steps C3-C7 in the above step S400, the following sub-steps C7-C9 are further included:

[0152] In C7: the central server sends the global model to each partition;

[0153] In C8: Each partition generates a control strategy based on the global model and local data, which is expressed as:

[0154] u k =g k (x k ,w * )

[0155] Among them, x k is the state variable of partition k, g k is the control strategy generation function, w * is the global model;

[0156] In C9: Each partition executes a local control strategy, and adjacent partitions exchange boundary information for coordination.

[0157] It should be noted that each partition generates a local control strategy based on the global model, and adjacent partitions exchange boundary information for coordination. This execution method not only ensures the real-time control, but also realizes the coordination between partitions.

[0158] Combining the federated learning framework with differential privacy protection not only solves the data privacy problem in traditional distributed control, but also provides a quantifiable level of privacy protection through the differential privacy mechanism while ensuring control performance.

[0159] The above is a schematic scheme of a distribution network partition collaborative optimization control method of this embodiment. It should be noted that the technical scheme of the distribution network partition collaborative optimization control system and the technical scheme of the distribution network partition collaborative optimization control method mentioned above belong to the same concept. For the details not described in detail in the technical scheme of the distribution network partition collaborative optimization control system in this embodiment, please refer to the description of the technical scheme of the distribution network partition collaborative optimization control method mentioned above.

[0160] The distribution network partition collaborative optimization control system in this embodiment includes:

[0161] A partitioning module, used for obtaining electrical distance and operation characteristics, and partitioning the distribution network using a clustering analysis algorithm according to the electrical distance and operation characteristics;

[0162] A model building module is used to obtain local data of each partition and build a local model for each partition using the local data;

[0163] An aggregation module, used to update local model parameters based on the local data of each partition by using a horizontal federated learning architecture, introduce differential privacy technology and aggregate the local model parameters of each partition to update the global model;

[0164] The optimization module is used to optimize the global model through a collaborative optimization control algorithm to obtain a first global model, generate a local control strategy according to the first global model, and perform coordinated execution.

[0165] This embodiment further provides a computing device, which is applicable to the coordinated optimization control of distribution network partitions, including:

[0166] Memory and processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the distribution network partition collaborative optimization control method proposed in the above embodiment.

[0167] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for realizing the coordinated optimization control method of distribution network partitions as proposed in the above embodiment is implemented.

[0168] The storage medium proposed in this embodiment and the method for realizing collaborative optimization control of distribution network partitions proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0169] Through the above description of the implementation methods, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software and necessary general hardware, and of course can also be implemented by hardware. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ReadOnly, Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform the methods of various embodiments of the present invention.

[0170] Example 2

[0171] Referring to Table 1, this embodiment is different from the first embodiment in that it provides a verification test of a distribution network partition collaborative optimization control method to verify and illustrate the technical effects adopted in this method.

[0172] The simulation verification is carried out using the IEEE 123 node power distribution system as an example;

[0173] The system is divided into three sub-areas, and a local optimization control model is deployed in each sub-area, considering objectives such as voltage regulation, power balance, and network loss minimization.

[0174] The proposed federated learning framework is adopted, and the training is iterated for 30 rounds, with 5 local updates in each round.

[0175] The control strategies for each sub-area are generated based on the global model and executed in a coordinated manner.

[0176] Table 1 Performance comparison of different algorithms (IEEE 123 node system)

[0177] Performance Indicators Federated Learning DGD ADMM Dual Decomposition Centralized Voltage deviation (pu) 0.024 0.035 0.031 0.038 0.025 Network loss rate (%) 3.2 4.1 3.8 4.3 3.3 Communication overhead (KB / round) 45 128 156 98 - Calculation time(s) 0.8 0.6 1.2 1.5 2.3 Privacy protection degree (%) 95 60 65 55 0 Convergence rounds 25 45 38 52 -

[0178] As shown in Table 1, compared with other distributed algorithms, federated learning achieves a smaller voltage deviation (0.024pu) and a lower network loss rate (3.2%), which is close to the effect of centralized control.

[0179] The communication overhead of federated learning (45KB / round) is significantly lower than that of other algorithms, thanks to the fact that it only exchanges model parameters instead of raw data. Through the differential privacy mechanism, federated learning achieves 95% privacy protection, which is much higher than other algorithms. Federated learning converges in just 25 rounds, while other algorithms require 38-52 rounds.

[0180] The superiority of the federated learning algorithm in the coordinated optimization control of distribution network partitions has been fully demonstrated, especially in terms of privacy protection, communication efficiency and control performance balance, which effectively solves the problems of high computational complexity, insufficient data privacy protection and poor scalability faced by large-scale distribution network systems. Through technologies such as distribution network partitioning, federated learning framework design, differential privacy protection and distributed execution, global optimization control in a distributed environment is achieved. While protecting data privacy, this application improves the overall operating efficiency of the distribution network and its ability to carry distributed resources, and has broad application prospects.

[0181] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A distribution network partition collaborative optimization control method, characterized in that: include: Acquire electrical distance and operation characteristics, and partition the distribution network using a cluster analysis algorithm according to the electrical distance and operation characteristics; Get the local data of each partition, and build a local model for each partition using the local data. Using a horizontal federated learning architecture, local model parameters are updated based on the local data of each partition, differential privacy technology is introduced, and local model parameters of each partition are aggregated to update the global model; The global model is optimized by a collaborative optimization control algorithm to obtain a first global model, and a local control strategy is generated according to the first global model and coordinatedly executed.

2. The distribution network partition collaborative optimization control method according to claim 1, characterized in that: The use of cluster analysis algorithms to partition distribution networks includes: The electrical distance between nodes is defined as the inverse of the absolute value of the admittance between nodes; The square of the electrical distance between nodes is divided by twice the square of the bandwidth parameter of the Gaussian kernel function, the result is negative, and the similarity matrix is ​​obtained using the exponential function; Subtract the diagonal matrix from the similarity matrix to get the Laplacian matrix; The Laplacian matrix and the corresponding eigenvector are equal to the product of the eigenvalue of the Laplacian matrix and the diagonal matrix and the corresponding eigenvector to obtain the eigenvalue; The eigenvector corresponding to the smallest non-zero eigenvalue is selected to form a first matrix, and the row vectors of the first matrix are clustered to obtain the partition of the distribution network.

3. The distribution network partition collaborative optimization control method according to claim 1 or 2, characterized in that: Updating local model parameters based on the local data of each partition includes: Define global model parameters and local model parameters for each partition; The local model parameters of each partition after iteration are the difference between the local model parameters of the partition in the previous iteration and the product of the learning rate and the gradient of the loss function of the partition to the local model parameters in the previous iteration; The global model parameters are obtained by summing the products of the local model parameters of each partition and the corresponding number of samples and dividing by the total number of samples.

4. The distribution network partition collaborative optimization control method according to claim 3, characterized in that: Differential privacy techniques include: During the aggregation process, the noise added to the local model parameters of each partition is the sum of the local model parameters and Gaussian noise after iteration for each partition; Gaussian noise includes mean and variance, and the variance is expressed as: Among them, δ is the privacy failure probability, ε is the preset privacy budget, and Δf is the sensitivity.

5. The distribution network partition collaborative optimization control method according to claim 4, characterized in that: The global model includes: Through iterative training and model aggregation, a global model is constructed and the global loss function is defined as: The optimization goal is to minimize the global loss function, which is expressed as: w * =argmin w F(w) Among them, n k is the number of samples in the kth partition, n is the total number of samples, w is the global model parameter, F k is the loss function of the kth partition, and arg min is the parameter of the minimum value.

6. The distribution network partition collaborative optimization control method according to claim 5, characterized in that: Optimizing the global model by means of a collaborative optimization control algorithm includes: Initialize global model parameters; For each round of communication t=1,2,…T, the central server broadcasts the current global model to all partitions; Each partition performs E rounds of local updates; Each partition sends the updated model parameters to the central server; The central server aggregates model parameters and outputs a first global model.

7. The distribution network partition collaborative optimization control method according to claim 6, characterized in that: Local control strategies include: The central server sends the global model to each partition; Each partition generates a control strategy based on the global model and local data, which is expressed as: u k =g k (x k ,w * ) Among them, x k is the state variable of partition k, g k is the control strategy generation function, w * is the global model; Each partition executes local control strategies, and adjacent partitions exchange boundary information for coordination.

8. A system using the distribution network partition collaborative optimization control method according to any one of claims 1 to 7, characterized in that: include: A partitioning module, used for obtaining electrical distance and operation characteristics, and partitioning the distribution network using a clustering analysis algorithm according to the electrical distance and operation characteristics; A model building module is used to obtain local data of each partition and build a local model for each partition using the local data; An aggregation module, used to update local model parameters based on the local data of each partition by using a horizontal federated learning architecture, introduce differential privacy technology and aggregate the local model parameters of each partition to update the global model; The optimization module is used to optimize the global model through a collaborative optimization control algorithm to obtain a first global model, generate a local control strategy according to the first global model, and perform coordinated execution.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the distribution network partition collaborative optimization control method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the distribution network partition collaborative optimization control method as described in any one of claims 1 to 7.

Images