An artificial intelligence-based power distribution network topology identification method and system
By constructing an RF-CNN-Attention hybrid model, which combines data preprocessing and attention mechanisms, the problem of high efficiency and high accuracy in distribution network topology identification under complex power grid environments is solved, and high-precision topology identification is achieved under noise and disaster scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-02-03
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies struggle to achieve efficient and accurate distribution network topology identification in the face of large-scale distributed power source integration and complex power grid environments, especially lacking multi-mode identification capabilities under noise interference and disaster scenarios.
An AI-based RF-CNN-Attention hybrid model is adopted. The KNN algorithm is used to fill in missing values, K-means clustering analysis and min-max normalization are used to process the distribution network measurement data, and random forest is used to select features. The CNN-Attention model is used to extract feature maps and assign different weights through the attention mechanism, and finally output the topology classification results.
It enhances the feature mining capability and processing efficiency of distribution network topology identification, improves identification accuracy and efficiency in noisy and complex environments, and can complete the topology classification of voltage data within 10 milliseconds.
Smart Images
Figure CN122153623A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system automation technology, and more specifically to a distribution network topology identification method and system based on artificial intelligence. Background Technology
[0002] With the large-scale integration of distributed power sources (such as photovoltaic power generation), the topology of urban power distribution networks changes frequently. Traditional topology identification methods based on physical models are limited by problems such as insufficient coverage of measurement equipment, high computational complexity, and poor real-time performance.
[0003] While graph-based abstraction methods simplify power grid structure analysis, they struggle to address the multi-mode identification requirements under noise interference and disaster scenarios. Furthermore, existing machine learning algorithms have shortcomings in feature importance mining, data noise robustness, and multi-label classification generalization capabilities.
[0004] Therefore, how to propose an artificial intelligence-based method and system for distribution network topology identification to achieve efficient and high-precision distribution network topology identification is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a distribution network topology identification method and system based on artificial intelligence, which realizes high-precision and high-efficiency intelligent identification of topology structure in complex power grid environment.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: On the one hand, this invention proposes an artificial intelligence-based method for topology identification of power distribution networks, comprising the following steps: Acquire distribution network measurement data and preprocess the distribution network measurement data; Construct a hybrid RF-CNN-Attention model, consisting of a sequentially connected RF model and a CNN-Attention model; The RF model is used to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics; The preliminary voltage features are input into the CNN-Attention model to extract feature maps. Different weights are assigned to different feature maps through the attention mechanism, and finally the topology classification result is output.
[0007] Preferably, the distribution network measurement data is preprocessed, including: The KNN algorithm is used to calculate the mean of neighboring samples of missing voltage data in the distribution network measurement data based on Euclidean distance, and the missing values are filled in. The K-means algorithm was used to perform cluster analysis on the photovoltaic power generation in the power distribution network measurement data to generate a typical power output scenario dataset. The voltage amplitude in the power distribution network measurement data is mapped to the [0,1] interval by min-max normalization.
[0008] Preferably, the KNN algorithm is used to calculate the mean of neighboring samples of missing voltage data in the distribution network measurement data based on Euclidean distance, and to complete the missing value imputation, including: Data samples were determined based on Euclidean distance. of K One neighbor; Pick K The average of the neighbors for the data sample The j Each feature is filled in, and the filled value is... for: ; in, yes K The first of the neighbors k The first neighbor's j Each feature value.
[0009] Preferably, the RF model is used to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics, including: Calculate the first using random forest. j The importance of each feature The formula is as follows: ; In the formula, The number of decision trees; Features of all samples of the out-of-bag data at random After adding noise interference, the error of the out-of-bag data is calculated again. The prediction error rate calculated using the raw out-of-bag data; The features are sorted in descending order of importance, and the top p% are selected as the initial voltage features.
[0010] Preferably, the CNN-Attention model includes an input layer, a convolutional layer, an attention layer, a pooling layer, a flattening layer, a first fully connected layer, a Dropout layer, a second fully connected layer, and an output layer connected in sequence. The input layer outputs a single-channel two-dimensional tensor adapted for convolution operations, which serves as the input to the convolutional layer. The convolutional layer outputs a local feature map activated by ReLU, which is then input to the attention layer. The attention layer generates a query matrix, a key matrix, and a value matrix, calculates attention weights, and outputs a weighted and enhanced feature map, which is then fed into the pooling layer. The pooling layer outputs a dimensionality-reduced feature map. The flattening layer flattens the dimensionality-reduced feature map into a vector, which is then input into the first fully connected layer. The first fully connected layer outputs a high-dimensional vector mapped by ReLU, which is then fed into the Dropout layer. The Dropout layer outputs a regularized vector, which is then input into the second fully connected layer. The second fully connected layer outputs a topology category space vector, which is then fed into the output layer. The output layer outputs the topology classification probability result processed by Softmax.
[0011] Preferably, the attention layer generates a query matrix, a key matrix, and a value matrix, calculates attention weights, and outputs a weighted and enhanced feature map, including: Each time step t The query vector is Q t The key matrix is K The value matrix is V Then the attention score in the attention mechanism is: In the formula, d k The dimension of the key vector; After the attention score is normalized using the Softmax function, the resulting probability distribution value is the attention weight, as shown in the following formula: ;in It is a query vector With key vector K The dot product; Weighted fusion and residual connection are performed based on the attention weights to output a weighted enhanced feature map.
[0012] Preferably, an artificial intelligence-based distribution network topology identification method further includes: The RF-CNN-Attention hybrid model is trained using the cross-entropy loss function, which measures the difference between the model's predicted labels and the true labels.
[0013] On the other hand, the present invention also proposes an artificial intelligence-based distribution network topology identification system to implement the aforementioned artificial intelligence-based distribution network topology identification method, comprising: The data processing module is used to acquire power distribution network measurement data and preprocess the power distribution network measurement data. The model building module is used to build a hybrid RF-CNN-Attention model, which consists of a sequentially connected RF model and a CNN-Attention model, including: The feature filtering unit uses the RF model to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage features. The topology identification unit is used to input the preliminary voltage features into the CNN-Attention model, extract feature maps, assign different weights to different feature maps through the attention mechanism, and finally output the topology classification result.
[0014] As can be seen from the above technical solution, compared with the prior art, this invention discloses an artificial intelligence-based method and system for distribution network topology identification. It preprocesses distribution network measurement data by performing KNN missing value imputation, photovoltaic output scenario clustering, and normalization preprocessing. It utilizes random forest to filter voltage features to reduce input dimensionality and constructs an RF-CNN-Attention hybrid model. In this model, CNN extracts local features, and the attention mechanism calculates attention weights by generating query, key, and value matrices and assigns different weights to the feature maps. Residual connections are combined to enhance feature representation, and finally, topology label classification is achieved through fully connected layers. This invention improves the feature mining capability and processing efficiency of distribution network topology identification through multi-model fusion and refined feature processing. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0016] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of topology 0; Figures 3(a), 3(b), and 3(c) are schematic diagrams showing the recognition accuracy of different algorithms at noise intensities of 0%, 2%, and 5%, respectively. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] On the one hand, embodiments of the present invention propose a distribution network topology identification method based on artificial intelligence, such as... Figure 1 As shown.
[0019] Calling the IEEE 33-node system in MATLAB 2023b, refer to... Figure 2 The study input active power, reactive power, and photovoltaic (PV) power at 192 time points over two days. Distributed PV systems were integrated at nodes 6, 11, 17, 25, and 29 of the IEEE 33-node system. A 30% fluctuation was set for the load power data on the second day to simulate load fluctuations in the same distribution network on different days under real-world conditions. Using MatPower 7.1, 5376 voltage data points were generated for 28 typical topologies in the IEEE 33-node system. The selected typical topologies included radial networks, ring networks, and islanded topologies.
[0020] The method proposed in this invention includes the following steps: S1. Acquire distribution network measurement data and preprocess the distribution network measurement data, including: The KNN algorithm is used to calculate the mean of neighboring samples of missing voltage data in the distribution network measurement data based on Euclidean distance, and then the missing values are filled.
[0021] Wherein, it is assumed that the data sample set is It contains n samples, each with m features. Taking Euclidean distance as an example, for two samples... and Its Euclidean distance is defined as: ; in, , Samples and The Each feature value. Select the nearest one. Each sample is taken as its neighbor. The missing values are imputed using the average of the neighbors, assuming... The Each feature is missing, and its imputation value is... It can be represented as: ; in, yes The first of the neighbors The first neighbor's Each feature value is filled in. Once all missing values are filled in, the missing data processing is complete.
[0022] The K-means algorithm is used to perform cluster analysis on photovoltaic power generation in distribution network measurement data to generate a dataset of typical power output scenarios.
[0023] The voltage amplitude in the distribution network measurement data is mapped to the [0,1] interval by min-max normalization, thus eliminating the difference in dimensions.
[0024] S2. Construct a hybrid RF-CNN-Attention model, consisting of a sequentially connected RF model and a CNN-Attention model.
[0025] S21. Use the RF model to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics. This step is used to filter key features and reduce the input dimensionality.
[0026] For each sample, all decision trees make predictions independently. In this embodiment, the topology identification problem is essentially a classification problem. It is assumed that the random forest has... A decision tree, then features Its importance is: .
[0027] In the formula, The number of decision trees; Features of all samples of the out-of-bag data at random Add noise interference and recalculate the error of the out-of-bag data; The prediction error rate is calculated using the original out-of-bag data.
[0028] The features are sorted in descending order of importance, and the top p% are selected as the initial voltage features.
[0029] S22. Input the initial voltage features into the CNN-Attention model, extract the feature maps, assign different weights to different feature maps through the attention mechanism, and finally output the topology classification results.
[0030] The CNN-Attention model consists of an input layer, a convolutional layer, an attention layer, a pooling layer, a flattening layer, a first fully connected layer, a dropout layer, a second fully connected layer, and an output layer, all connected in sequence.
[0031] The input layer outputs a single-channel 2D tensor adapted for convolution operations, which serves as the input to the convolutional layer. The convolutional layer outputs a local feature map activated by ReLU, which is then input to the attention layer. The attention layer generates a query matrix, a key matrix, and a value matrix, calculates the attention weights, and outputs a weighted and enhanced feature map, which is then fed into the pooling layer. The pooling layer outputs a dimensionality-reduced feature map. The flattening layer flattens the dimensionality-reduced feature map into a vector, which is then input to the first fully connected layer. The first fully connected layer outputs a high-dimensional vector mapped by ReLU, which is then fed into the Dropout layer. The Dropout layer outputs a regularized vector, which is then input to the second fully connected layer. The second fully connected layer outputs a topology category space vector, which is then fed into the output layer. The output layer outputs the topology classification probability result processed by Softmax.
[0032] The specific hierarchical structure is shown below: Input Layer: To optimize memory management during model training, a loader is pre-created to extract k random data points for training each time. Assuming each data point has n features, the input dimension of the input layer is (k, n). To meet the requirements of a one-dimensional convolutional neural network (Conv1d), a dimension of size 1 is inserted in the middle of the original input to represent it as a single-channel signal, thus constructing a sequence with one channel and k steps. The output dimension of the input layer is (k, 1, n).
[0033] Convolutional Layers: The function of convolutional layers is to extract local temporal features of a sequence through convolutional kernels. The input dimension of a convolutional layer is (k, 1, n). In this paper, the kernel size is set to 1×3, the stride is set to 1, and one unit of zero is padded on both sides of the input data to keep the output size consistent with the input size. The ReLU activation function is applied to each convolutional vector to introduce nonlinearity and enhance the feature extraction capability of the model. With the number of output channels set to m, the output dimension of the convolutional layer is (k, m, n).
[0034] The operation expression for a convolutional layer is: ; in, Indicates the first The first layer Each feature mapping, Indicates the first The first layer Feature mapping; Indicates the first The first layer One convolutional kernel, Indicates the first The first layer There are 1 bias weight; M is the number of input feature maps; express Activation function; This represents the convolution operation.
[0035] The activation function used in this embodiment is computationally simple and can effectively alleviate the gradient vanishing problem. The activation function is expressed as follows: .
[0036] Attention Layer: The role of the attention layer is to capture temporal dependencies, introduce residual connections, and increase the model's weights on important features. The input dimension of the attention layer is (k, m, n). First, three matrices, Query, Key, and Value, are generated through three 1×1 convolutions. Then, attention is calculated according to the steps in Section 3.2.3. After the attention calculation is completed, weighted fusion and residual connections are performed, and the output dimension is still (k, m, n).
[0037] The attention layer assigns different weights to different feature maps through an attention mechanism, guiding the model to focus on features that are crucial to solving the task.
[0038] For the Attention mechanism, the objects for calculating the attention score can be abstracted as the query vector and the key-value matrix. For each time step... Its query vector is The key matrix is The value matrix is Then the attention score in the attention mechanism is: ; In the formula, is the dimension of the key vector.
[0039] The calculated attention scores are normalized using the Softmax function to generate an attention distribution.
[0040] in It is a query vector With key vector The dot product represents the degree of matching, and the softmax function can... Transform it into a probability distribution such that the sum of all attention weights is 1.
[0041] Finally, the attention distribution is applied to the value vector set. The context vector is obtained by weighted summation, which incorporates information from all positions in the input sequence. ; in, This represents the hidden state of the encoder.
[0042] The SE (Squeeze-and-Excitation) attention mechanism enhances feature representation by explicitly modeling the dependencies between channels. Its core can be divided into Squeeze (compression) and Excitation (excitation).
[0043] In the compression step, global average pooling is used to compress the information of each channel into a single value. This applies to the feature map input to the model. The channel statistics generated after compression are: The formula for calculating the Cth element is: ; This operation compresses global spatial information into compact channel values, facilitating subsequent modeling of relationships between channels.
[0044] In the activation step, channel activation values will be generated using two fully connected layers. : ; in, Represents the ReLU activation function. For the Sigmoid function, and These are the weights of the fully connected layer. Finally, the feature map is obtained by scaling channel by channel: .
[0045] Pooling layers are used to reduce the spatial size of feature maps, decrease the number of parameters and computational cost, prevent model overfitting, and extract important features. This embodiment uses max pooling, with both the pooling kernel size and stride set to 2 (default). The input dimension of the pooling layer is (k, m, n), and the output dimension is (k, m, n). ).
[0046] Taking max pooling as an example: In the formula, This refers to the position within the pooling window.
[0047] Flattening layer: The function of the flattening layer is to flatten multi-dimensional features into vectors, with input dimensions (k, m, ..., ...). The output dimension is (k, m×) ).
[0048] First Fully Connected Layer: In the first fully connected layer, a nonlinear mapping is performed, and ReLU activation is used. The input dimension of the first fully connected layer is (k, m×) If mapped to s dimensions, the output dimension is (k, s).
[0049] The first fully connected layer maps the features extracted by the convolutional and pooling layers to the final output: ; In the formula, For the input matrix, This is the weight matrix. This is the bias matrix. For this project, the final output is the distribution network topology feature label.
[0050] Dropout layer: To prevent overfitting, a Dropout layer is placed between two fully connected layers. ; in, This represents the activation value input to the Dropout layer. This represents the probability that a neuron's output will be set to zero. For example, in this embodiment, the activation value is set to 0.5, meaning that 50% of the neuron outputs will be set to zero. Randomly disabling some neurons in the network prevents the model from depending on any single neuron or group of neurons, thereby improving the model's generalization ability.
[0051] The input dimension of the Dropout layer is (k, s), while the output dimension remains unchanged.
[0052] The second fully connected layer outputs the prediction results, achieving the final classification. The input dimension of the second fully connected layer is (k, s), and the output dimension is (k, p), where p is the total number of labels.
[0053] Preferably, the attention layer generates a query matrix, a key matrix, and a value matrix, calculates attention weights, and outputs a weighted and enhanced feature map, including: Each time step t The query vector is Q t The key matrix is K The value matrix is V Then the attention score in the attention mechanism is: In the formula, d k The dimension of the key vector; After the attention score is normalized using the Softmax function, the resulting probability distribution value is the attention weight, as shown in the following formula: ;in It is a query vector With key vector K The dot product; Weighted fusion and residual connections are performed based on attention weights to output a weighted enhanced feature map.
[0054] Furthermore, an artificial intelligence-based method for distribution network topology identification also includes: The RF-CNN-Attention hybrid model is trained using the cross-entropy loss function, which measures the difference between the model's predicted labels and the true labels. The cross-entropy loss function is as follows:
[0055] In the formula, For real labels, To predict probabilities.
[0056] In this embodiment, to further compare and analyze the identification accuracy of the proposed method, the samples were randomly divided into a training set (80%) and a test set (20%) in the IEEE 33-node system test. To simulate the signal disturbances that may exist under the error of distribution network measurement equipment and disaster scenarios, different levels of noise were added to both the test set and the training set for model training and verification. During the model training phase, all neural network models were set to 1000 iterations to ensure that the models could be fully trained. The best-performing model obtained during the training process was retained and called. Random forest and other traditional machine learning algorithms used for comparative analysis were set with default parameters. Different noise levels (0%, 2%, 5%) were set to compare the accuracy of different algorithms in the IEEE 33-node system using 33, 16, and 12 voltage features, respectively. The results are shown in Figures 3(a), 3(b), and 3(c).
[0057] Simulation results show that: ① Both CNN-Attention and CNN outperform traditional machine learning algorithms. As the dimensionality of input features decreases and noise increases, the role of the attention mechanism becomes increasingly prominent, proving its crucial role in improving model robustness; ② In a horizontal comparison, CNN-Attention outperforms deep learning algorithms such as CNN and DNN, and significantly outperforms other machine learning algorithms; ③ When performing topology identification on more complex distribution networks, i.e., when the input voltage features increase, the lead of CNN-Attention over CNN is not significant.
[0058] By calling the CNN-Attention model, in different simulation examples, the time taken for topology identification using voltage data at a single time point is less than 10 milliseconds, proving that it can meet the requirements of online application of distribution network topology identification in terms of both accuracy and speed.
[0059] On the other hand, embodiments of the present invention also propose an artificial intelligence-based distribution network topology identification system for implementing the aforementioned artificial intelligence-based distribution network topology identification method, comprising: The data processing module is used to acquire distribution network measurement data and preprocess the distribution network measurement data. The model building module is used to build a hybrid RF-CNN-Attention model, which consists of a sequentially connected RF model and a CNN-Attention model, including: The feature filtering unit uses the RF model to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics. The topology identification unit is used to input the initial voltage features into the CNN-Attention model, extract feature maps, assign different weights to different feature maps through the attention mechanism, and finally output the topology classification result.
[0060] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0061] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A distribution network topology identification method based on artificial intelligence, characterized in that, Includes the following steps: Acquire distribution network measurement data and preprocess the distribution network measurement data; Construct a hybrid RF-CNN-Attention model, consisting of a sequentially connected RF model and a CNN-Attention model; The RF model is used to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics; The preliminary voltage features are input into the CNN-Attention model to extract feature maps. Different weights are assigned to different feature maps through the attention mechanism, and finally the topology classification result is output.
2. The distribution network topology identification method based on artificial intelligence according to claim 1, characterized in that, Preprocessing of the power distribution network measurement data includes: The KNN algorithm is used to calculate the mean of neighboring samples of missing voltage data in the distribution network measurement data based on Euclidean distance, and the missing values are filled in. The K-means algorithm was used to perform cluster analysis on the photovoltaic power generation in the power distribution network measurement data to generate a typical power output scenario dataset. The voltage amplitude in the power distribution network measurement data is mapped to the [0,1] interval by min-max normalization.
3. The distribution network topology identification method based on artificial intelligence according to claim 2, characterized in that, The KNN algorithm is used to calculate the mean of neighboring samples of missing voltage data in the distribution network measurement data based on Euclidean distance, and to complete the missing value imputation, including: Data samples were determined based on Euclidean distance. of K One neighbor; Pick K The average of the neighbors for the data sample The j Each feature is filled in, and the filled value is... for: ； in, yes K The first of the neighbors k The first neighbor's j Each feature value.
4. The distribution network topology identification method based on artificial intelligence according to claim 1, characterized in that, The RF model is used to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage characteristics, including: Calculate the first using random forest. j The importance of each feature The formula is as follows: ； In the formula, The number of decision trees; Features of all samples of the out-of-bag data at random Add noise interference and recalculate the error of the out-of-bag data; The prediction error rate calculated using the raw out-of-bag data; The features are sorted in descending order of importance, and the top p% are selected as the initial voltage features.
5. The distribution network topology identification method based on artificial intelligence according to claim 1, characterized in that, The CNN-Attention model comprises an input layer, a convolutional layer, an attention layer, a pooling layer, a flattening layer, a first fully connected layer, a Dropout layer, a second fully connected layer, and an output layer, all connected in sequence. The input layer outputs a single-channel two-dimensional tensor adapted for convolution operations, which serves as the input to the convolutional layer; the convolutional layer outputs a local feature map activated by ReLU, which is then input to the attention layer; the attention layer generates a query matrix, a key matrix, and a value matrix, calculates attention weights, outputs a weighted and enhanced feature map, and feeds it into the pooling layer. The pooling layer outputs a dimensionality-reduced feature map; the flattening layer flattens the dimensionality-reduced feature map into a vector, which is then input into the first fully connected layer; the first fully connected layer outputs a high-dimensional vector after ReLU mapping, which is then input into the Dropout layer; the Dropout layer outputs a regularized vector, which is then input into the second fully connected layer; the second fully connected layer outputs a topology category space vector, which is then input into the output layer; the output layer outputs the topology classification probability result after Softmax processing.
6. The distribution network topology identification method based on artificial intelligence according to claim 5, characterized in that, The attention layer generates a query matrix, a key matrix, and a value matrix, calculates attention weights, and outputs a weighted and enhanced feature map, including: Each time step t The query vector is Q t The key matrix is K The value matrix is V Then the attention score in the attention mechanism is: In the formula, d k The dimension of the key vector; After the attention score is normalized using the Softmax function, the resulting probability distribution value is the attention weight, as shown in the following formula: ;in It is a query vector With key vector K The dot product; Weighted fusion and residual connection are performed based on the attention weights to output a weighted enhanced feature map.
7. The distribution network topology identification method based on artificial intelligence according to claim 1, characterized in that, Also includes: The RF-CNN-Attention hybrid model is trained using the cross-entropy loss function, which measures the difference between the model's predicted labels and the true labels.
8. An artificial intelligence-based distribution network topology identification system, used to implement the artificial intelligence-based distribution network topology identification method as described in any one of claims 1-7, characterized in that, include: The data processing module is used to acquire power distribution network measurement data and preprocess the power distribution network measurement data. The model building module is used to build a hybrid RF-CNN-Attention model, which consists of a sequentially connected RF model and a CNN-Attention model, including: The feature filtering unit uses the RF model to perform feature filtering on the preprocessed distribution network measurement data to obtain preliminary voltage features. The topology identification unit is used to input the preliminary voltage features into the CNN-Attention model, extract feature maps, assign different weights to different feature maps through the attention mechanism, and finally output the topology classification result.