Power grid topology identification method based on deep learning network, medium and terminal

Abstract

The invention is suitable for the technical field of electric power Internet of Things, and relates to a power grid topology identification method based on a deep learning network, a medium and a terminal, and the method comprises the steps: constructing a shallow convolutional neural network CNN for power grid topology signal identification; locking a parameter of a second full connection layer in the CNN as 1, and training network parameters except the second full connection layer by using historical power grid signal data to obtain a pre-trained basic model; acquiring a real-time current signal of the current power grid, forming a training sample set, unlocking parameters of the second full-connection layer, and performing fine adjustment on the second full-connection layer by using the training sample set to obtain a topology identification model adaptive to the current power grid environment; and inputting power grid signals acquired in real time into the topology identification model, outputting probability distribution of each topology category, and selecting the category with the highest probability as a topology identification result. The method solves the problems of high calculation complexity, strong model dependence and poor field adaptability in a traditional method, and is high in identification accuracy and strong in adaptability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power Internet of Things, and particularly relates to a power grid topology identification method based on a deep learning network, a medium and a terminal. BACKGROUND

[0002] Power grid topology identification is the core basis of power system automation operation, and directly affects key links such as state estimation, fault location and network reconstruction. With the development of smart grids, the penetration rate of distributed energy (such as photovoltaic and energy storage) in distribution networks is increasing, and the topology change frequency is significantly increasing. The traditional method faces the following technical bottlenecks:

[0003] Strong dependence on data quality: Traditional methods (such as SCADA systems) rely on switch remote signaling monitoring, but there are "perception blind areas" in device failures or non-covered areas, leading to an increase in topology misjudgment rate; Insufficient dynamic response: Existing rule-driven or shallow machine learning methods (such as SVM) are difficult to capture the spatiotemporal correlation characteristics of the power grid, especially the dynamic changes of multi-time section data; Computational efficiency and resource limitations: Complex optimization algorithms (such as ant colony genetic algorithm) face the problem of combinatorial explosion, and deep learning models require a large amount of computing resources for end-side device deployment, making it difficult to meet real-time requirements; Difficulty in small sample adaptation: Existing transfer learning solutions are prone to catastrophic forgetting under a small amount of field data, leading to degradation of pre-trained model performance. Patent No. CN117411190B provides a distribution network topology identification method, device, system and medium based on multiple source information. First, the first topology information of the target distribution network at the last time and the multiple source information of the target distribution network at the current time are obtained, wherein the multiple source information includes: SCADA measurement information, μPMU measurement information and AMI measurement information; According to the μPMU measurement information, it is determined whether the current time is a topology change time, and when the current time is a topology change time, the second topology information of the target distribution network is determined according to the μPMU measurement information, the first topology information and the first neural network established in advance; According to the second topology information, the SCADA measurement information and the AMI measurement information, the third topology information of the target distribution network at the current time is determined. By using more accurate μPMU measurement information for preliminary topology prediction, and then combining SCADA measurement information and AMI measurement information to correct the predicted topology, the characteristics of multiple source information are considered, and the distribution network topology identification is realized. In this patent, the measurement is also performed by the SCADA system, and the topology misjudgment rate is high, which has the same disadvantages as the prior art.

[0004] Therefore, it is urgent to provide a topology identification method with high recognition accuracy and strong adaptability to solve the problems of high computational complexity, strong model dependence and poor field adaptability in traditional methods. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide a power grid topology identification method based on deep learning, so as to solve the problems of high computational complexity, strong model dependency and poor field adaptability in the traditional method; in addition, the present application also provides a power grid topology identification medium and terminal based on deep learning.

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

[0007] In the first aspect, the present application provides a power grid topology identification method based on deep learning, comprising the following steps:

[0008] S10, a shallow convolutional neural network CNN for power grid topology signal identification is constructed, the CNN comprises a data input layer, a first convolutional layer, an activation function layer, a pooling layer, a second convolutional layer, a flattening layer, a first full connection layer, a second full connection layer and a Softmax output layer connected in sequence;

[0009] S20, the parameters of the second full connection layer are locked to 1, and the network parameters except the second full connection layer are trained using historical power grid signal data to obtain a pre-trained basic model;

[0010] S30, real-time current signals of the current power grid are collected to form a training sample set, the parameters of the second full connection layer are unlocked, and the second full connection layer is fine-tuned using the training sample set to obtain a topology identification model adapted to the current power grid environment;

[0011] S40, the real-time collected power grid signals are input into the topology identification model, and the probability distribution of each topology category is output, and the category with the highest probability is selected as the topology identification result.

[0012] Further, in the S10, the input data periodically received by the data input layer is as follows:

[0013]

[0014] Wherein, the period is A, T is the number of groups of real-time current sampling data received in the period A, and D is the number of current sampling data in each group.

[0015] Further, let the topology identification signal period of 4 bits be A, and the number of current sampling data in each group D = B / T, wherein B is the total number of current sampling data in the period A, and when B cannot be divided by T, D is rounded up, and the end of the data is supplemented with corresponding number of 0.

[0016] Further, the sample annotation adopts binary label representation method, and the length of the label vector is 17, which is represented as follows:

[0017]

[0018] wherein, represents that there is no topological signal, represents that there is a topological signal;

[0019] represents that there is a topological signal 0, represents that there is no topological signal 0;

[0020] represents that there is a topological signal 1, represents that there is no topological signal 1;

[0021] represents that there is a topological signal 2, represents that there is no topological signal 2;

[0022] represents that there is a topological signal 3, represents that there is no topological signal 3;

[0023] represents that there is a topological signal 4, represents that there is no topological signal 4;

[0024] …

[0025] represents that there is a topological signal F, represents that there is no topological signal F.

[0026] Further, the error of each label is calculated using a binary cross-entropy loss function:

[0027] .

[0028] Further, the activation function layer adopts a ReLU activation function.

[0029] In a second aspect, the present application further provides a computer readable storage medium, the storage medium stores a computer program, and the computer program is executed by a processor to implement the method described above.

[0030] In a third aspect, the present application further provides an electronic terminal, comprising: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes the method described above.

[0031] The power grid topology identification method based on a deep learning network provided by the present application, medium and terminal have at least the following beneficial effects compared with the prior art:

[0032] The problems of high calculation complexity, strong model dependency and poor field adaptability exist in the grid topology identification in the traditional method. The process of the method is simple and convenient, and the key technical problems of high calculation complexity, strong model dependency and poor field adaptability in the prior art are effectively solved through the structural design of the 'basic model + reserved model', the basic model adopts a convolution layer to extract general features in the grid signal, such as voltage mutation, current waveform change and other topology related modes, thereby avoiding the dependence on the accurate physical model and the state estimation process in the traditional method, and the identification efficiency and generalization ability are significantly improved; in the deployment stage, the reserved model is quickly adapted to the local environment characteristics by fine-tuning a small amount of field data, while the basic model remains unchanged, thereby greatly reducing the field training cost and time overhead. The architecture not only improves the noise resistance and inference speed of the model, but also realizes the smooth transition from general identification to personalized adaptation, enhances the engineering practicability and deployability of the system, improves the identification accuracy, and has good popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the scheme of the present application, a brief introduction will be given below to the drawings used in the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0034] Figure 1 A flowchart of a voltage topology identification method based on a deep learning network provided by an embodiment of the present application;

[0035] Figure 2 A structural diagram of a layer convolutional neural network CNN in a voltage topology identification method based on a deep learning network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0036] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the related drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present application more thorough and comprehensive.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application herein are only for the purpose of describing the specific embodiments and are not intended to limit the present application.

[0038] The application provides a power grid topology identification method based on a deep learning network, applied to a power grid topology identification process in a power system.

[0039] S10, a shallow convolutional neural network (CNN) for power grid topology signal identification is constructed, the CNN comprising a data input layer, a first convolutional layer, an activation function layer, a pooling layer, a second convolutional layer, a flattening layer, a first full connection layer, a second full connection layer and a Softmax output layer connected in sequence; S20, the parameters of the second full connection layer are locked as 1, and historical power grid signal data is used to train the network parameters except the second full connection layer, to obtain a pre-trained base model; S30, real-time current signals of a current power grid are collected to form a training sample set, the parameters of the second full connection layer are unlocked, and the training sample set is used to fine-tune the second full connection layer, to obtain a topology identification model adapted to the current power grid environment; S40, the real-time collected power grid signals are input into the topology identification model, and the probability distribution of each topology category is output, and the category with the highest probability is selected as the topology identification result.

[0040] The application solves the problems of high calculation complexity, strong model dependency and poor field adaptability in the traditional method, and has high identification accuracy and strong adaptability.

[0041] In order to enable personnel in the technical field to better understand the application scheme, the technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings.

[0042] The application provides a power grid topology identification method based on a deep learning network, applied to a power grid topology identification process in a power system, the method divides the model into a base model (Base Model) and a reserved model (AdaptationModel), the base model is used to extract general topology features in power grid signals, and the reserved model is used for rapid adaptive training according to a field environment, through the structural design, the pre-training before model deployment and the light-weight fine-tuning after deployment are combined, the identification accuracy and adaptability are considered, and the problems of high calculation complexity, strong model dependency and poor field adaptability in the traditional method are solved. Figure 1 With Figure 2 In the embodiment, the power grid topology identification method based on the deep learning network comprises the following steps:

[0043] S10, a shallow convolutional neural network (CNN) for power grid topology signal identification is constructed, the CNN comprising a data input layer, a first convolutional layer, an activation function layer, a pooling layer, a second convolutional layer, a flattening layer, a first full connection layer, a second full connection layer and a Softmax output layer.

[0044] Specifically, the data input layer (Data Input) is used to receive time series data of power grid operation signals; the first convolutional layer (Conv1) is used to extract local features of the input signals; the activation function layer (Relu) is used to introduce a nonlinear factor and enhance the expression ability of the model; the pooling layer (Pooling) is used to compress the feature map size and reduce the computational complexity; the second convolutional layer (Conv2) is used to further combine and extract high-order features; the flattening layer is used to compress the multi-dimensional feature vector into one dimension; the first full connection layer (Full Connection1) is used as the core part of the basic model to perform weighted summation on the flattened features; the second full connection layer (Full Connection2) is used as the core part of the reserved model to support on-site fine tuning; and the Softmax output layer (Softmax) is used to map the final features to a probability distribution.

[0045] Further, in the embodiment, the topology identification signal detection device samples the current signal in real time, records a 4-bit topology identification signal period as A, divides the real-time current sampling data received in the period A into T groups, records the number of current sampling data in each A period as B, and then the number of current sampling data in each group in the T groups of current sampling data is D = B / T (when B cannot be divided by T, D is rounded up, and 0 is added at the end of the data):

[0046] .

[0047] Further, in the embodiment, the sample labeling adopts a binary label representation method (One-Hot Encoding) to represent the current. Assuming that there are 17 identification results, the length of the label vector is 17, and each position represents an identification result:

[0048]

[0049] Among them, represents that there is no topology signal, represents that there is a topology signal;

[0050] represents that there is a topology signal 0, represents that there is no topology signal 0;

[0051] represents that there is a topology signal 1, represents that there is no topology signal 1;

[0052] represents that there is a topology signal 2, represents that there is no topology signal 2;

[0053] indicates that the topology signal 3 exists, indicates that the topology signal 3 does not exist;

[0054] indicates that the topology signal 4 exists, indicates that the topology signal 4 does not exist;

[0055] …

[0056] indicates that the topology signal F exists, indicates that the topology signal F does not exist.

[0057] In other embodiments, it can also be an N bit topology identification signal, and the label type of one N bit topology identification signal period is 2 N+1 .

[0058] Further, in the embodiment, the loss function uses a binary cross-entropy loss function to calculate the error of each label:

[0059] .

[0060] S20, the parameters of the second full connection layer (Full Connection2 layer) are locked to 1, and the network parameters except the second full connection layer are trained using historical power grid signal data to obtain a pre-trained basic model.

[0061] Specifically, in the embodiment, the parameters of the layer are not iteratively updated during the entire training process, the prediction probability of each fault type is calculated by forward propagation, the binary cross-entropy loss of each label is calculated, and the network parameters are back propagated and updated.

[0062] S30, real-time current signals of the current power grid are collected to form a training sample set, the parameters of the second full connection layer are unlocked, the second full connection layer is fine-tuned using the training sample set, and a topology identification model adapted to the current power grid environment is obtained.

[0063] S40, the real-time collected power grid signals are input into the topology identification model, the probability distribution of each topology category is output, and the category with the highest probability is selected as the topology identification result.

[0064] Specifically, in the embodiment, in actual deployment, the current power grid current data is labeled according to the binary label representation method, the parameters of the second full connection layer (Full Connection2 layer) are unlocked, and the collected current power grid current data is used for several rounds of training to achieve the purpose of fine-tuning and optimizing the network according to the actual scene.

[0065] Further, in the embodiment, the data collected by the current sensor in the power grid such as an electric meter is input into the network, and the neural network outputs a vector with the same label format as the binary label representation, and each element represents the probability of the identification result.

[0066] In the embodiment, it is assumed that:

[0067]

[0068] According to the output, it can be judged that there is a 10% probability that there is no topology signal, a 83% probability that there is topology signal 0, a 1% probability that there is topology signal 1, a 3% probability that there is topology signal 2, …, and a 3% probability that there is topology signal F. The system takes the result with the highest probability value as the final determination result.

[0069] The embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement any one of the methods in the embodiment.

[0070] The embodiment of the application further provides an electronic terminal, which comprises a processor and a memory.

[0071] The computer readable storage medium in the embodiment can be understood by those skilled in the art that all or part of the steps of the above-mentioned method embodiments can be completed by the hardware of the computer program. The computer program can be stored in a computer readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the storage medium comprises ROM, RAM, a magnetic disc or an optical disc and various storage program codes.

[0072] The electronic terminal provided in the embodiment comprises a processor, a memory, a transceiver and a communication interface. The memory and the communication interface are connected with the processor and the transceiver and complete communication between each other. The memory is used for storing a computer program, the communication interface is used for communication, and the processor and the transceiver are used for running the computer program, so that the electronic terminal executes the steps of the above method.

[0073] Compared with the prior art, the power grid topology identification method based on the deep learning network, the medium and the terminal described in the above embodiments have the problems of high computational complexity, strong model dependency and poor field adaptability in the power grid topology identification in the traditional method. The process of the present application is simple and convenient to operate. Through the structural design of the "basic model + reserved model", the key technical problems of high computational complexity, strong model dependency and poor field adaptability in the prior art are effectively solved. The basic model uses a convolution layer to extract general features in the power grid signal, such as voltage mutation, current waveform change and other topology-related patterns, avoiding the dependence of the traditional method on the accurate physical model and the state estimation process, and significantly improving the identification efficiency and generalization ability. In the deployment stage, the reserved model is fine-tuned through a small amount of field data to quickly adapt to the local environmental characteristics, while the basic model remains frozen, thereby greatly reducing the field training cost and time overhead. This architecture not only improves the noise immunity and inference speed of the model, but also realizes a smooth transition from general identification to personalized adaptation, enhances the engineering practicability and deployability of the system, and improves the identification accuracy and adaptability, and has good popularization prospect.

[0074] Obviously, the above-described embodiments are only the preferred embodiments of the present application, not all the embodiments, and the drawings show the preferred embodiments of the present application, but do not limit the patent scope of the present application. The present application can be realized in many different forms, and on the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing specific embodiments, or make equivalent replacements to some technical features. Any equivalent structure made by using the content of the present application specification and drawings, directly or indirectly applied to other related technical fields, is also within the patent protection scope of the present application.

Claims

1. A power grid topology identification method based on deep learning networks, characterized in that, Includes the following steps: S10. Construct a shallow convolutional neural network (CNN) for power grid topology signal recognition. The CNN includes a data input layer, a first convolutional layer, an activation function layer, a pooling layer, a second convolutional layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a Softmax output layer connected in sequence. S20. Lock the parameters of the second fully connected layer to 1, and use historical power grid signal data to train the network parameters other than the second fully connected layer to obtain a pre-trained basic model. S30. Collect the real-time current signal of the current power grid to form a training sample set, unlock the parameters of the second fully connected layer, and use the training sample set to fine-tune the second fully connected layer to obtain a topology recognition model adapted to the current power grid environment. S40. Input the real-time collected power grid signal into the topology identification model, output the probability distribution of each topology category, and select the category with the highest probability as the topology identification result.

2. The power grid topology identification method based on deep learning networks according to claim 1, characterized in that, In step S10, the input data periodically received by the data input layer takes the following form: ； Wherein, the period is A, T is the number of sets of real-time current sampling data received within the period A, and D is the number of sets of current sampling data.

3. The power grid topology identification method based on deep learning networks according to claim 2, characterized in that, Let A be the period of the 4-bit topology identification signal, and D = B / T be the number of current sampling data in each group, where B is the total number of current sampling data in period A. If B is not divisible by T, D is rounded up, and zeros corresponding to the number of times are added to the end of the data.

4. The power grid topology identification method based on deep learning networks according to claim 3, characterized in that, The sample annotations use binary label representation, with a label vector length of 17, as shown below: ； in, This indicates that no topological signal exists. This indicates the presence of a topological signal; This indicates the existence of a topological signal of 0. This indicates that the topological signal 0 does not exist; This indicates the existence of a topological signal of 1. This indicates that topological signal 1 does not exist; This indicates the existence of a topological signal of 2. This indicates that topological signal 2 does not exist; This indicates the existence of a topological signal of type 3. This indicates that topological signal 3 does not exist; This indicates the existence of a topological signal of 4. This indicates that topological signal 4 does not exist; … This indicates the existence of a topological signal F. This indicates that there is no topological signal F.

5. The power grid topology identification method based on deep learning networks according to claim 4, characterized in that, The error for each label is calculated using the binary cross-entropy loss function: 。 6. The power grid topology identification method based on deep learning networks according to claim 1, characterized in that, The activation function layer uses the ReLU activation function.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

8. An electronic terminal, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the terminal to perform the method as described in any one of claims 1 to 6.

Images