Power distribution network fault recovery reconstruction method based on deep residual network

Through the failure recovery method based on deep residual network, the offline training model is used to learn reconstruction strategies, and the problems of large calculation volume and low efficiency of traditional algorithms are solved, and the rapid and feasible reconstruction of distribution network faults is achieved.

CN119026456AInactive Publication Date: 2024-11-26STATE GRID ZHEJIANG ELECTRIC POWER CO LTD JINHUA POWER SUPPLY CO
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410996440.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Traditional distribution network fault recovery algorithms have scenarios where the calculation volume is large, time-consuming, and it is difficult to face the large number of switches, resulting in low fault reconstruction speed and efficiency.

Method used

Using the fault recovery reconstruction method based on deep residual network, the deep residual network model is trained offline, and the reconstruction control strategy is learned from historical data, to avoid repeated trend calculations and switch combination operations, reduce the calculation amount and reduce the calculation time. Ensure that there is no combination explosion when there are many switches.

Benefits of technology

It realizes the reduction of calculation amount and calculation time, improves the speed and efficiency of fault reconstruction, ensures the feasibility of fault reconstruction solutions, and avoids the occurrence of combined explosions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026456A_ABST
    Figure CN119026456A_ABST
Patent Text Reader

Abstract

The invention provides a power distribution network fault recovery reconstruction method based on a deep residual network, and the method specifically comprises the steps: obtaining the historical operation data of a power distribution network, and constructing a sample data set in combination with the topological structure of the power distribution network; training a deep residual network model offline based on the sample data set; collecting operation parameters of the power distribution network in real time, inputting the operation parameters of the power distribution network into the offline trained deep residual network model, obtaining a reconstructed switch combination solution based on the deep residual network model, and verifying the reconstructed switch combination solution; determining whether to update the power distribution network topology based on the verification result, and obtaining a reconstruction optimal solution; and updating the topological structure of the power distribution network based on the reconstructed optimal solution. According to the method, a reconstruction control strategy can be learned in a mode of offline training the deep residual network model, repeated load flow calculation optimization iteration operation is avoided, the calculation amount is reduced, the calculation time is shortened, a combined explosion phenomenon does not occur under the condition that the number of switches is large, and the feasibility of a fault reconstruction solution is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault reconstruction, and in particular to a distribution network fault recovery and reconstruction method based on a deep residual network. Background Art

[0002] Smart grid has gradually become the main development direction of power grid technology, and smart distribution network is an indispensable part of smart grid. The main feature of smart distribution network is that it has complete self-healing function, which can minimize the impact of power grid failure on users. The distribution network is directly facing users, and the strength of its self-healing ability directly affects the power supply quality of the power grid to users. As a key link in realizing the self-healing function of smart distribution network, distribution network fault recovery plays an important role in improving the power supply quality of the power grid.

[0003] Distribution network fault recovery is a multi-objective, multi-dimensional, multi-constrained, multi-period nonlinear configuration optimization problem, and is a type of NP (Non-deterministic Polynomial) hard problem. Traditional distribution network fault recovery strategies can be roughly divided into mathematical optimization methods and heuristic search methods. When these methods are used to deal with NP-hard problems, the reconstruction scheme based on mathematical optimization has the problem of dimensionality disaster, as well as large amount of calculation, long calculation time, and poor real-time performance. Most of the fault recovery methods based on heuristic search are based on switch operations. It is necessary to add some heuristic information according to the characteristics of the problem itself during the search process, so as to determine the direction of the heuristic information and optimize it towards the direction of the optimal solution. Although this method has the characteristics of obtaining an effective solution in real time within the allowed time, and its rules can be used for the universality of distribution networks of various structures, the initial state of the system has a great influence on the search results. When faced with a large number of switches, combinatorial explosion is very likely to occur.

[0004] In summary, the traditional fault recovery reconstruction algorithm has the problems of large amount of calculation, long time consumption, difficulty in dealing with scenarios with a large number of switches, etc., and the speed and efficiency of fault reconstruction are low. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the fault recovery method used in the prior art when performing fault recovery and reconstruction, such as large amount of calculation, long time consumption, and difficulty in coping with scenarios with a large number of switches. A distribution network fault recovery and reconstruction method based on a deep residual network is provided, which learns the reconstruction control strategy from a historical data set by offline training of a deep residual network model. Therefore, when the deep residual network model is applied, repeated iterative operations of power flow calculation and optimization are avoided, the amount of calculation and the calculation time are reduced, and when the optimal reconstruction solution is subsequently determined, there is no need to perform a linear combination of the switch combination. When a large number of switches are encountered, a combination explosion phenomenon will not occur, thereby ensuring the feasibility of the fault reconstruction solution.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A distribution network fault recovery and reconstruction method based on deep residual network includes:

[0008] Obtain the historical operation data of the distribution network and build a sample data set based on the topological structure of the distribution network;

[0009] Offline training of deep residual network model based on sample dataset;

[0010] Collect distribution network operating parameters in real time, input the distribution network operating parameters into the offline trained deep residual network model, obtain the reconstructed switch combination solution based on the deep residual network model, and verify the reconstructed switch combination solution;

[0011] Determine whether to update the distribution network topology based on the verification results to obtain the optimal reconstruction solution;

[0012] The distribution network topology is updated based on the reconstructed optimal solution.

[0013] Furthermore, the method of obtaining historical operation data of the distribution network and constructing a sample data set in combination with the topological structure of the distribution network includes constructing a data set based on the historical operation data of the distribution network and the topological structure of the distribution network, taking network loss as the objective function, and obtaining the switch combination sequence for restoring power supply to all power-off loads in non-fault areas based on the data set based on a genetic algorithm to construct a sample data set.

[0014] Furthermore, the objective function is expressed as:

[0015]

[0016] Among them, f ploss is the network loss after reconstruction, v′ i is the voltage at node i, ∠δ′ i is the voltage angle at node i, R ij is the branch resistance of the branch ij formed by nodes i and j, X ijis the branch reactance of the branch ij formed by nodes i and j, k ij is the switch closing state of the branch ij formed by nodes i and j, and n is the total number of branches in the distribution network.

[0017] Furthermore, the method of inputting the distribution network operation parameters into the offline trained deep residual network model, obtaining the reconstructed switch combination solution based on the deep residual network model, and verifying the reconstructed switch combination solution includes dividing the distribution network parameters according to the collection time, inputting the corresponding distribution network parameters into the deep residual network model according to the collection time, the deep residual network model outputs the corresponding reconstructed switch combination solution, obtaining the objective function value when the deep residual network model outputs the reconstructed switch combination, and calling the objective function value at the previous moment, comparing the obtained objective function value with the objective function value at the previous moment, and verifying the reconstructed switch combination solution according to the comparison result.

[0018] Furthermore, the obtained objective function value is compared with the objective function value at the previous moment, and the reconstructed switch combination solution is verified according to the comparison result, including: when the obtained objective function value is less than the objective function value at the previous moment, the reconstructed switch combination solution passes the verification; when the obtained objective function value is greater than or equal to the objective function value at the previous moment, the reconstructed switch combination solution fails the verification.

[0019] Furthermore, determining whether to update the distribution network topology based on the verification result to obtain the optimal reconstruction solution includes:

[0020] When the reconstructed switch combination solution is verified, it is determined whether the conditions for ending the distribution network topology update iteration are met;

[0021] If the distribution network topology update iteration end condition is not met, then continue to obtain the reconstructed switch combination solution at the next moment and verify it until the distribution network topology update iteration end condition is met;

[0022] If the distribution network topology update iteration end condition is met, the distribution network topology is updated with the current reconstructed switch combination solution, and the updated distribution network topology is taken as the optimal reconstruction solution;

[0023] When the reconstructed switch combination solution verification fails, the distribution network topology result of the previous moment is maintained, and the reconstructed switch combination solution of the next moment is continuously obtained for verification.

[0024] Furthermore, the distribution network topology update iteration termination conditions include a maximum number of iterations, a convergence condition, and a power grid operation restriction condition.

[0025] The beneficial effects of the present invention are:

[0026] By offline training of the deep residual network model, the reconstruction control strategy is learned from the historical data set. Therefore, when the deep residual network model is applied, the repeated iterative operations of power flow calculation optimization are avoided, the amount of calculation and the calculation time are reduced. In the subsequent determination of the optimal reconstruction solution, there is no need to perform linear combination of switch combinations. When encountering a large number of switches, there will be no combination explosion, which ensures the feasibility of the fault reconstruction solution.

[0027] Specifically, when constructing a sample data set for training the deep residual network model, a genetic algorithm is used to analyze the historical operation data of the distribution network to directly obtain the switch combination sequence for restoring power to all power-off loads in non-fault areas. The obtained switch combination sequence is then used as the sample data set. In the subsequent practical application of the deep residual network model, there is no need to perform complex power flow calculations, which optimizes the algorithm calculation amount and improves the speed and efficiency of fault reconstruction.

[0028] When iteratively searching for the optimal solution for reconstruction, the corresponding objective function value is set as the iterative optimization condition for reconstructing the switch combination solution, avoiding the combination explosion and the difficulty in obtaining the optimal solution when encountering a large number of switches, and ensuring the feasibility of the fault reconstruction solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of a process of the present invention;

[0030] Figure 2 It is a complete flow chart of obtaining the optimal reconstruction solution according to an embodiment of the present invention;

[0031] Figure 3 This is a test result diagram of a deep residual network model according to an embodiment of the present invention. Figure 1 ;

[0032] Figure 4 This is a test result diagram of a deep residual network model according to an embodiment of the present invention. Figure 2 . DETAILED DESCRIPTION

[0033] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0034] Example:

[0035] A distribution network fault recovery and reconstruction method based on deep residual network, such as Figure 1 As shown, including,

[0036] Obtain the historical operation data of the distribution network and build a sample data set based on the topological structure of the distribution network;

[0037] Offline training of deep residual network model based on sample dataset;

[0038] Collect distribution network operating parameters in real time, input the distribution network operating parameters into the offline trained deep residual network model, obtain the reconstructed switch combination solution based on the deep residual network model, and verify the reconstructed switch combination solution;

[0039] Determine whether to update the distribution network topology based on the verification results to obtain the optimal reconstruction solution;

[0040] The distribution network topology is updated based on the reconstructed optimal solution.

[0041] The method of obtaining historical operation data of the distribution network and constructing a sample data set in combination with the topological structure of the distribution network includes constructing a data set based on the historical operation data of the distribution network and the topological structure of the distribution network, taking network loss as the objective function, and obtaining a switch combination sequence for restoring power supply to all power-off loads in non-fault areas based on the data set based on a genetic algorithm to construct a sample data set.

[0042] The data set contains a characteristic matrix that reflects the changes in the power flow and topological structure of the distribution network, including information such as the topological connection relationship of the network, load demand and fault branches. The size of the data set is determined by the number of nodes and fault branches in the distribution network.

[0043] When training the deep residual network model, if each feature of the input data is different, the training results will be affected by the size and dimension differences of the input data. Therefore, before the sample data set is input into the deep residual network model, the feature matrix and other information in the sample data set are normalized.

[0044] Since the distribution network is generally operated in an open loop through a closed loop design, it must be radial in normal operation, and reconstruction is to achieve radial operation by switching the tie switch and the section switch. In the distribution network loop characteristics, there is only one tie switch in a loop, so a radial topology is to be formed, and only one switch in a loop can be opened.

[0045] Specifically, in this embodiment, the idea of ​​one-hot encoding is used to treat the switch state in each circuit as a discrete variable, and these discrete values ​​are used as the characteristic encoding of the state, and the fault reconstruction solution combination is further determined by a genetic algorithm. The obtained reconstruction solution combination is the one-hot encoding value assignment for each circuit that requires state transformation.

[0046] In order to optimize the subsequent deep residual network model training, the feature matrix in the data set is matched with the obtained reconstruction solution combination, which is equivalent to adding the corresponding reconstruction solution combination label to the feature matrix.

[0047] There are many objective functions for distribution network reconstruction, the main one is to quickly restore power supply after a fault. Therefore, the operating cost and other aspects are taken into consideration, and the objective function of reconstruction is to minimize the reconstructed network loss.

[0048] The expression of the objective function is:

[0049]

[0050] Among them, f ploss is the network loss after reconstruction, v′ i is the voltage at node i, ∠δ′ i is the voltage angle at node i, R ij is the branch resistance of the branch ij formed by nodes i and j, X ij is the branch reactance of the branch ij formed by nodes i and j, k ij is the switch closing state of the branch ij formed by nodes i and j, and n is the total number of branches in the distribution network.

[0051] Genetic algorithm is a search and optimization method that imitates biological heredity and evolution. It has high robustness and wide adaptability. It has no requirements for the initial state of the problem. Combined with the various fault recovery strategies recorded in the distribution network, it can find the switch combination sequence for restoring power supply to all power-off loads in non-fault areas, that is, the reconstruction solution combination.

[0052] After obtaining the feature matrix and the corresponding reconstruction solution combination, a sample data set is constructed according to the feature matrix and the corresponding reconstruction solution combination.

[0053] And because deep learning is different from traditional machine learning methods, it has end-to-end characteristics. Users only need to input initial data and do not need to participate in data processing processes such as feature extraction and dimensionality enhancement. When the number of layers of the deep learning network is deeper, the expression ability will be stronger in theory, but after the CNN network reaches a certain depth and then deepens, the classification performance will not improve, but will cause the network to converge more slowly, and the accuracy will also decrease. Even if the data set is increased, the classification performance and accuracy will not improve. Therefore, in this embodiment, a deep residual network is specifically used to solve the degradation problem of the deep convolutional network when the depth is deepened.

[0054] The deep residual network (ResNet-50) is a neural network composed of 50 layers (including convolutional layers, pooling layers, batch normalization layers, fully connected layers, etc.), and its structure includes:

[0055] 1) Input layer: 224×224×3 types of input data types;

[0056] 2) Convolutional layer and pooling layer: First, a 7×7 convolutional layer and a 3×3 maximum pooling layer are used to downsample the input data and extract features;

[0057] 3) Residual block: A large number of residual units are set, and these residual units are encapsulated into 4 residual blocks. There is one residual block in each stage, and the ratio of the number of residual units in each residual block is 3:4:6:3;

[0058] 4) Global pooling layer: After the last stage, a global average pooling layer is added to obtain the statistical characteristics of the entire feature;

[0059] 5) Fully connected layer: Expand the output of the global pooling layer, input it into the Softmax classifier, and output the category probability.

[0060] The constructed sample data set is then input into the deep residual network model for offline training.

[0061] The method of inputting the distribution network operation parameters into the offline trained deep residual network model, obtaining the reconstructed switch combination solution based on the deep residual network model, and verifying the reconstructed switch combination solution includes dividing the distribution network parameters according to the collection time, inputting the corresponding distribution network parameters into the deep residual network model according to the collection time, the deep residual network model outputting the corresponding reconstructed switch combination solution, obtaining the objective function value when the deep residual network model outputs the reconstructed switch combination, and calling the objective function value at the previous moment, comparing the obtained objective function value with the objective function value at the previous moment, and verifying the reconstructed switch combination solution according to the comparison result.

[0062] Firstly, the actual operation parameters of the distribution network are collected through the distribution network smart meters, and then the features of the collected distribution network parameters are extracted to construct the corresponding feature matrix and input the trained deep residual network model.

[0063] Since the time variation factor is not taken into account when training the deep residual network model, the operation of the distribution network has both time complexity and space complexity in the actual operation process. Therefore, in order to consider the time-varying characteristics of the load, while determining the reconstructed switch combination solution through the deep residual network model, the previously constructed distribution network reconstruction objective function is used to verify and analyze the reconstructed switch combination solution obtained by the objective function value corresponding to each moment, thereby realizing dynamic reconstruction of the distribution network fault reconstruction.

[0064] By verifying and analyzing the reconstructed switch combination solutions at different times, it is possible to adapt to the time-varying characteristics of the load in actual operation, and to obtain the optimal reconstruction solution for the entire time period by reconstructing the network structure multiple times at different time points.

[0065] The obtained objective function value is compared with the objective function value at the previous moment, and the reconstructed switch combination solution is verified according to the comparison result, including: when the obtained objective function value is less than the objective function value at the previous moment, the reconstructed switch combination solution passes the verification; when the obtained objective function value is greater than or equal to the objective function value at the previous moment, the reconstructed switch combination solution fails the verification.

[0066] The determining whether to update the distribution network topology based on the verification result to obtain the optimal reconstruction solution includes, when the reconstruction switch combination solution is verified, determining whether the distribution network topology update iteration end condition is met;

[0067] If the distribution network topology update iteration end condition is not met, then continue to obtain the reconstructed switch combination solution at the next moment and verify it until the distribution network topology update iteration end condition is met;

[0068] If the distribution network topology update iteration end condition is met, the distribution network topology is updated with the current reconstructed switch combination solution, and the updated distribution network topology is taken as the optimal reconstruction solution;

[0069] When the reconstructed switch combination solution verification fails, the distribution network topology result of the previous moment is maintained, and the reconstructed switch combination solution of the next moment is continuously obtained for verification.

[0070] And when the distribution network topology update iteration end condition is not met, if it is necessary to continue to obtain the reconstructed switch combination solution at the next moment, the objective function value corresponding to the current reconstructed switch combination solution is used as the comparison object of the objective function value corresponding to the reconstructed switch combination solution at the next moment.

[0071] If the verification of the reconstructed switch combination solution fails, the objective function value corresponding to the reconstructed switch combination solution at the previous moment is kept, and the objective function value corresponding to the reconstructed switch combination solution at the current moment is replaced by the reconstructed switch combination solution at the previous moment as the comparison object for the objective function value corresponding to the reconstructed switch combination solution at the next moment.

[0072] The complete process of obtaining the optimal reconstruction solution is as follows: Figure 2 shown.

[0073] The distribution network topology update iteration termination conditions include a maximum number of iterations, a convergence condition, and a power grid operation restriction condition.

[0074] Specifically, the maximum number of iterations can be set according to needs. When the iteration process reaches this number, the iteration is stopped to ensure that the algorithm does not run indefinitely.

[0075] The convergence condition may be a change error of the objective function value. If the objective function value gradually tends to be stable and the change error is within a set error range, it can be considered that convergence has occurred and iteration can be stopped.

[0076] The grid operation conditions can be some voltage stability, power balance and other conditions that need to be met by the updated topology after the distribution network topology is updated, and can be set according to the specific distribution network operation conditions.

[0077] As long as one of the distribution network topology update iteration end conditions is met, the iteration ends and the optimal reconstruction solution is determined.

[0078] Taking the IEEE33 node test system as an example, a sample data set is constructed according to the corresponding distribution network parameters, and the corresponding data in the sample data set is normalized before the simulation starts, so as to obtain the training set and the test set. Specifically, in this embodiment, 3200 data are used for training and 1000 data are used for testing. The test results are as follows: Figure 3 and Figure 4 shown.

[0079] Depend on Figure 3 and Figure 4 It can be seen that the accuracy of the deep residual network model converges to a relatively high accuracy, with an accuracy of 0.97 on the test set. It is possible to learn the reconstruction control strategy from the historical data set to output an accurate reconstructed switch combination solution. And because the deep residual network model is trained offline, no complex power flow calculation is required in the application, and the reconstruction calculation time of the IEEE33-node test system is only 47.8ms.

[0080] The above-described embodiment is only a preferred solution of the present invention and does not limit the present invention in any form. There are other variations and modifications without exceeding the technical solution described in the claims.

Claims

1. A distribution network fault recovery and reconstruction method based on deep residual network, characterized in that: Including, obtaining historical operation data of the distribution network and building a sample data set based on the topological structure of the distribution network; Offline training of deep residual network model based on sample dataset; Collect distribution network operating parameters in real time, input the distribution network operating parameters into the offline trained deep residual network model, obtain the reconstructed switch combination solution based on the deep residual network model, and verify the reconstructed switch combination solution; Determine whether to update the distribution network topology based on the verification results to obtain the optimal reconstruction solution; The distribution network topology is updated based on the reconstructed optimal solution.

2. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 1 is characterized in that: The method of obtaining historical operation data of the distribution network and constructing a sample data set in combination with the topological structure of the distribution network includes constructing a data set based on the historical operation data of the distribution network and the topological structure of the distribution network, taking network loss as the objective function, and obtaining a switch combination sequence for restoring power supply to all power-off loads in non-fault areas based on the data set based on a genetic algorithm to construct a sample data set.

3. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 2 is characterized in that: The expression of the objective function is: Among them, f ploss is the network loss after reconstruction, v′ i is the voltage at node i, ∠δ′ i is the voltage angle at node i, R ij is the branch resistance of the branch ij formed by nodes i and j, X ij is the branch reactance of the branch ij formed by nodes i and j, k ij is the switch closing state of the branch ij formed by nodes i and j, and n is the total number of branches in the distribution network.

4. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 3 is characterized in that: The method of inputting the distribution network operation parameters into the offline trained deep residual network model, obtaining the reconstructed switch combination solution based on the deep residual network model, and verifying the reconstructed switch combination solution includes dividing the distribution network parameters according to the collection time, inputting the corresponding distribution network parameters into the deep residual network model according to the collection time, the deep residual network model outputting the corresponding reconstructed switch combination solution, obtaining the objective function value when the deep residual network model outputs the reconstructed switch combination, and calling the objective function value at the previous moment, comparing the obtained objective function value with the objective function value at the previous moment, and verifying the reconstructed switch combination solution according to the comparison result.

5. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 4 is characterized in that: The obtained objective function value is compared with the objective function value at the previous moment, and the reconstructed switch combination solution is verified according to the comparison result, including: when the obtained objective function value is less than the objective function value at the previous moment, the reconstructed switch combination solution passes the verification; when the obtained objective function value is greater than or equal to the objective function value at the previous moment, the reconstructed switch combination solution fails the verification.

6. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 5 is characterized in that: The step of determining whether to update the distribution network topology based on the verification result and obtaining the optimal reconstruction solution includes: When the reconstructed switch combination solution is verified, it is determined whether the distribution network topology update iteration end condition is met; if the distribution network topology update iteration end condition is not met, the reconstructed switch combination solution at the next moment is obtained and verified until the distribution network topology update iteration end condition is met; If the distribution network topology update iteration end condition is met, the distribution network topology is updated with the current reconstructed switch combination solution, and the updated distribution network topology is taken as the optimal reconstruction solution; When the reconstructed switch combination solution verification fails, the distribution network topology result of the previous moment is maintained, and the reconstructed switch combination solution of the next moment is continuously obtained for verification.

7. The distribution network fault recovery and reconstruction method based on deep residual network according to claim 6 is characterized in that: The distribution network topology update iteration termination conditions include a maximum number of iterations, a convergence condition, and a power grid operation restriction condition.

Citation Information

Cited By

  • Multi-scene knowledge base fusion-oriented power distribution network reconstruction decision interpretable optimal selection method, system and equipment and medium

    CN121413462A

  • Power distribution network reconstruction decision explainable optimization method, system, device and medium facing multi-scene knowledge base fusion

    CN121413462B