Power distribution network modular design method and system application

Abstract

The application relates to the technical field of power distribution network design, and discloses a power distribution network modularization design method and system application. The method first acquires initial network parameters and load distribution data of a power distribution network, divides the power distribution network into multiple independent module units according to the initial network parameters, constructs a power flow characteristic operator model, and processes feature coding of the module units by using a pre-defined characteristic function; an initial optimization model pre-trained is acquired, the model is obtained through comparative learning of feature coding results by an integrated training framework; based on the load distribution data, the initial optimization model is subjected to multi-objective optimization training to obtain a trained power distribution network design model; and the model is applied to modularization design of a power distribution network scene to be designed to obtain an optimized design result. The method is adapted to different power distribution network scenes by means of modularization decomposition, feature coding, a pre-trained model and multi-objective optimization, helps to realize efficient and accurate design, and improves flexibility and expansibility of design.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution network design, in particular to a power distribution network modular design method and system application. BACKGROUND

[0002] With the rapid development of social economy, the demand for electricity continues to rise. As a key link between the power transmission network and users in the power system, the complexity of the structure of the power distribution network is increasing day by day. Traditional power distribution network design relies on experience-based planning, and various factors such as line layout, load capacity, and equipment selection need to be considered in the design process, often resulting in long design cycles and poor adaptability.

[0003] In the prior art, the design of the power distribution network often uses a whole planning mode, that is, the entire network is modeled and optimized uniformly. This mode, when faced with large-scale power distribution networks, will result in low computational efficiency due to complex network topology and high variable dimension, making it difficult to quickly respond to load fluctuations and network structure adjustment requirements. At the same time, due to the significant differences in load characteristics in different regions, the overall design method is difficult to accurately optimize the characteristics of local areas, which can easily result in unreasonable resource allocation, reduced power supply reliability, and other problems.

[0004] Traditional design methods rely heavily on the experience of designers, lack standardized design processes and modular design concepts, and result in the inability to reuse design results from different projects, leading to low design efficiency. With the development of smart grid technology, power distribution networks need to have greater flexibility and scalability to accommodate the emergence of new types of loads such as distributed power sources and electric vehicle charging. Traditional design methods have difficulty meeting these new demands. Therefore, exploring a modular design method that can achieve efficient, accurate, and flexible design of power distribution networks has become an important direction of current research. SUMMARY

[0005] The purpose of the present application is to provide a power distribution network modular design method and system application to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides a power distribution network modular design method, which comprises:

[0007] Obtaining initial network parameters and load distribution data of the power distribution network; based on the initial network parameters, dividing the power distribution network into a plurality of independent module units;

[0008] Constructing a power flow feature operator model, the power flow feature operator model using a pre-defined feature function to perform feature encoding processing on the plurality of independent module units;

[0009] obtaining a pre-trained initial optimization model, wherein the pre-trained initial optimization model is obtained through contrastive learning on the feature encoding processing result by using an integrated training framework;

[0010] based on the load distribution data, performing multi-objective optimization training on the pre-trained initial optimization model to obtain a trained power distribution network design model;

[0011] applying the trained power distribution network design model to the power distribution network scene to be designed for modular design processing to obtain an optimized power distribution network modular design result.

[0012] Preferably, the running state data of the plurality of independent module units is obtained; the topology structure of each independent module unit is determined according to the running state data; the feature vector instance of each independent module unit is obtained through feature extraction on the topology structure of each independent module unit based on a preset basic model; the design inference result of each independent module unit is obtained through identification of each feature vector instance based on a trained feature recognition component; and the preset basic model is integrated and optimized to obtain the pre-trained initial optimization model.

[0013] Preferably, the topology structure of each independent module unit is determined according to the running state data, comprising:

[0014] a plurality of module topology patterns are obtained; the topology structure of each independent module unit is associated and matched with the plurality of module topology patterns to obtain the associated parameters of the topology structure of each independent module unit; and the contrast feature code of each independent module unit is determined according to the associated parameters of the topology structure of each independent module unit.

[0015] Preferably, the topology structure of each independent module unit is associated and matched with the plurality of module topology patterns to obtain the associated parameters of the topology structure of each independent module unit, comprising:

[0016] the topology structure of the target independent module unit is associated and matched with each module topology pattern to obtain the associated value corresponding to each module topology pattern; the module topology pattern whose associated value exceeds the associated value threshold is taken as the target module topology pattern corresponding to the topology structure of the target independent module unit; and the target module topology pattern corresponding to the topology structure of the target independent module unit is taken as the associated parameter of the topology structure of the target independent module unit.

[0017] Preferably, the contrast feature code of each independent module unit is determined according to the associated parameters of the topology structure of each independent module unit, comprising:

[0018] obtaining a core feature code of a target module topology mode corresponding to the target independent module unit topology structure; determining a contrast feature code of a module topology mode in the plurality of module topology modes that is not associated with the target independent module unit topology structure according to the core feature code; and performing feature fusion processing on the core feature code and the contrast feature code to obtain a comparison feature code of the target independent module unit topology structure.

[0019] Preferably, the preset basic model comprises a plurality of network interaction layers and at least one network fusion layer, and the feature extraction is performed on the topology structure of each independent module unit based on the preset basic model to obtain a feature vector instance of each independent module unit, which comprises:

[0020] According to the running state data of the target independent module unit, an electric power parameter vector of each node unit in the target independent module unit topology structure, a strength vector of an associated node unit of each node unit and a connection relationship between each node unit and the associated node unit are obtained; a feature conversion processing is performed on the electric power parameter vector of each node unit to obtain a basic parameter vector of each node unit; a feature integration is performed on the basic parameter vector of the associated node unit and the strength vector to obtain an integrated vector; a first weight matrix of a first aggregation module used for interaction aggregation processing is obtained, a linear transformation processing is performed on the first weight matrix and the integrated vector to obtain a first transformation parameter matrix; a first-order interaction aggregation vector of the node unit is obtained by processing the first transformation parameter matrix based on a preset first nonlinear function; a parameter aggregation processing is performed on the input parameter vector of each node unit and the first-order interaction aggregation vector to obtain a first-order parameter vector of each node unit; a target-order interaction aggregation vector of each node unit is obtained by performing interaction aggregation processing on a previous-order parameter vector of the associated node unit and the strength vector based on a target network interaction layer; a parameter aggregation processing is performed on the previous-order parameter vector of each node unit and the target-order interaction aggregation vector to obtain a target-order parameter vector of each node unit; the target-order parameter vector of each node unit is taken as a feature vector of each node unit; and a feature vector instance corresponding to the target independent module unit topology structure is obtained by processing the feature vector of each node unit based on the network fusion layer.

[0021] Preferably, the parameter aggregation processing performed on the input parameter vector of each node unit and the first-order interaction aggregation vector to obtain the first-order parameter vector of each node unit comprises:

[0022] obtaining a second weight matrix of a second aggregation module for parameter aggregation processing; performing linear transformation processing on the second weight matrix and the first-order interaction aggregation vector to obtain a second transformation parameter matrix; and processing the input parameter vector of each node unit and the second transformation parameter matrix based on a preset second nonlinear function to obtain a first-order parameter vector of each node unit.

[0023] Preferably, the power parameter vector of each node unit in the target independent module unit topology, the associated node unit of each node unit and the strength vector of the connection relationship between each node unit and the associated node unit are obtained according to the running state data of the target independent module unit, including:

[0024] According to the running state data of the target independent module unit, the power index data of each subunit in the target independent module unit is obtained; the power index data of each subunit is processed by feature coupling to obtain the power parameter vector of each node unit in the topology structure of the target independent module unit; the contrast subunit having an electric energy connection path with each subunit is taken as the associated subunit of each subunit; the electric energy index data of the electric energy connection path between each subunit and the associated subunit is obtained; and the electric energy index data of the electric energy connection path is processed by feature coupling to obtain the strength vector of the connection relationship between each node unit and the associated node unit.

[0025] Preferably, the multi-objective optimization training of the pre-trained initial optimization model based on the load distribution data is performed to obtain the trained power distribution network design model, including:

[0026] Obtaining the load parameter of each load point in the load distribution data; analyzing and processing each load parameter based on the pre-trained initial optimization model to obtain an optimization feature instance corresponding to each load parameter; identifying each optimization feature instance based on the trained feature recognition component to obtain a load inference result of each load parameter; and performing integrated optimization training on the pre-trained initial optimization model according to the load inference result of each load parameter to obtain the trained power distribution network design model.

[0027] Preferably, the application further includes a power distribution network modular design system application, and the system application includes a server, and the server is used for executing the method as described above.

[0028] Compared with the prior art, the application has the following beneficial effects:

[0029] By dividing the power distribution network into multiple independent module units, the complex network is decomposed, and the design difficulty is reduced. Each module unit can be independently processed according to its load characteristics and network parameters, avoiding the calculation complexity problem caused by too many variables in overall design, which helps to improve the efficiency of the design process.

[0030] With the help of the power flow characteristic operator model, the module units are encoded, which can effectively extract the key features of each module and provide a reliable basis for subsequent optimization training. This feature encoding method takes into account both the local characteristics of the module and the correlation of the overall network, so that the design result can meet the power supply demand of the local area and ensure the coordinated operation of the entire power distribution network.

[0031] Using the pre-trained initial optimization model and the contrast learning results of the integrated training framework, the time cost of model training is reduced. The knowledge accumulated in the pre-training process can help the model adapt to new design scenarios faster, improving the generalization ability of the model, so that this method can play a role in power distribution network design of different scales and different load characteristics.

[0032] Based on the load distribution data, multi-objective optimization training is carried out, so that the power distribution network design model can consider multiple design objectives such as power supply stability and resource utilization. This multi-objective optimization idea avoids the limitations that may be brought by single-objective design, and helps to find a better design scheme under various constraints, improving the overall performance of the power distribution network.

[0033] Applying the trained model to modular design realizes the standardization and automation of the design process, reducing the dependence on manual experience. The design results have good reusability, and the design schemes of different modules can be directly applied or slightly adjusted in similar scenarios, improving the flexibility and expandability of the design work, and better adapting to various new demands brought by the development of smart grid. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 The timing diagram of the power distribution network modular design method described in the present application;

[0035] Figure 2 The flowchart for pre-training of the initial optimization model;

[0036] Figure 3 The flowchart for generating contrast feature encoding;

[0037] Figure 4 The flowchart for generating power parameters and intensity vectors;

[0038] Figure 5 The flowchart for multi-objective optimization training. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the protection scope of the present application.

[0040] Please refer to Figure 1 The present application provides a power distribution network modular design method, the method comprises:

[0041] Obtain initial network parameters and load distribution data of the power distribution network, wherein the initial network parameters cover physical attribute information such as line length, conductor type, transformer capacity, switch device parameters, and the load distribution data contains operation information such as load type, power value, peak-valley period characteristics of each node. Based on the initial network parameters, the power distribution network is divided into multiple independent module units through topology analysis or electrical distance division method, and each module unit has a relatively independent power supply range or functional positioning. A power flow characteristic operator model is constructed, which uses a pre-defined characteristic function to perform feature coding processing on multiple independent module units. The characteristic function can include a power flow equation, a sensitivity matrix or an adjacency matrix conversion rule in graph theory, to extract the electrical connection characteristics, power flow characteristics and topological structure characteristics of the module units. An initial optimization model pre-trained is obtained, which obtains the comparison learning of the feature coding processing results through an integrated training framework. The comparison learning process constructs a positive sample pair and a negative sample pair using historical module data, and optimizes the discrimination ability of the model for similar modules through a loss function. Based on the load distribution data, the initial optimization model pre-trained is subjected to multi-objective optimization training, and the multi-objectives include economy, reliability and power supply quality. The output results are balanced among the multi-objectives by adjusting the model parameters, and a trained power distribution network design model is obtained. The trained power distribution network design model is applied to the modular design processing of the to-be-designed power distribution network scene, the network parameters and load demand of the to-be-designed scene are input, and the optimized module division scheme, module connection mode and module parameter configuration result are output.

[0042] Embodiment 1: Please refer to Figure 2, obtaining the running state data of a plurality of independent module units is a basic link for constructing a pre-training initial optimization model. The data is collected by various sensors and intelligent terminals deployed in the module units, covering real-time monitoring information such as voltage, current, active power, reactive power, power factor, device surface temperature, switch opening and closing state, etc. The sensors upload data to the data processing platform through wired or wireless communication methods (such as RS485, LoRa, 5G, etc.), the platform cleans the raw data, removes abnormal values caused by communication interference, and ensures data integrity through a check algorithm (such as CRC check); normalization processing is performed to convert physical quantities of different dimensions to the same scale range (such as 0-1 or -1-1).

[0043] When determining the topology structure of each independent module unit according to the running state data, a directed graph representation of the module unit is constructed based on graph theory. Among them, the nodes of the graph are defined as the key devices in the module, such as distribution transformers, feeder switches, distributed power sources, load nodes, etc., each node stores attribute information of the device such as type (such as oil-immersed transformer, vacuum circuit breaker), rated capacity (such as 630kVA, 1250A), running state (such as operation, maintenance), etc.; the edges of the graph are defined as the electrical connection relationship between devices, and the weight of the edge includes line impedance (such as 0.4Ω / km), susceptance (such as 0.01S / km), connection type (such as overhead line, cable) and other parameters. After construction, the graph structure is traversed through depth-first search (DFS) or breadth-first search (BFS), and topological features such as the degree of the node (i.e. the number of edges directly connected to the node), the load rate of the edge (the ratio of the actual transmission power to the rated capacity), the number of connected components (the number of independent power supply areas in the module) are extracted, forming a quantitative description of the topology structure of the module unit. For example, a module unit contains 3 transformers and 5 feeders, and the DFS traversal finds that the connected component is 1, which can be described as "3 transformers operating in parallel, 5 feeders in single connection structure".

[0044] When the topology of each independent module unit is extracted based on the preset base model, the preset base model adopts a graph neural network (GNN) architecture, including an input layer, three hidden interaction layers, and an output layer. The input layer receives node features (device type, rated capacity, and running state coded as a one-hot vector) and edge features (connection impedance, susceptance, and connection type coded as a numerical vector) of the topology structure, and converts high-dimensional sparse original features into low-dimensional dense embedding vectors. The hidden interaction layer updates the node features through a neighborhood aggregation operation: each node collects the embedding vectors of all adjacent nodes, concatenates them with its own embedding vector, and then processes them through a weight matrix transformation (the weight matrix is a model training parameter, and the initial value is randomly generated) and a nonlinear activation function to generate a new node embedding vector. This process is performed layer by layer, with the first layer of hidden interaction layer capturing direct adjacency information of nodes, the second layer capturing secondary adjacency information, and the third layer capturing global topology information. The output layer globally pools the final embedding vectors of each node to generate a feature vector instance reflecting the overall topology features of the module unit. This vector contains information such as connection density, key device association strength, and topology redundancy of the module. For example, a module with high connection density may have a larger numerical value at the corresponding position in the vector.

[0045] When each feature vector instance is identified based on the trained feature recognition component, the feature recognition component is a random forest classifier, and its training data comes from typical module units in historical operation. These typical module units have been labeled with design labels in the design stage, such as "standard radial module" (single power supply, no connection), "multi-connection ring network module" (dual power supply, multiple connection lines), and "high reliability module" (including distributed power supply and energy storage device). The random forest builds multiple decision trees, each tree splits based on different feature subsets of the feature vector instance, and determines the design label of each feature vector instance through a voting mechanism. For example, a feature vector instance may be classified as a "standard radial module" if it shows low node degree, one connected component, and high impedance of the connection line in the edge weight; or it may be classified as a "multi-connection ring network module" if it shows high node degree, two connected components, and multiple low-impedance connection lines. The identification result outputs the design inference result of each module unit, including the module type, key parameter recommended value, and optimization direction.

[0046] According to the design inference result of each independent module unit, the preset basic model is integrated and optimized for training, and the design inference result is used as a supervision signal to adjust the parameters of the model by using a transfer learning method. First, the weight matrix of the input layer and the first hidden interaction layer is frozen, and only the weight matrix of the second and third hidden interaction layers and the parameters of the output layer are fine-tuned, so as to retain the basic topological feature extraction capability learned in the pre-training stage. At the same time, a contrastive learning loss function is introduced to construct a positive sample pair and a negative sample pair: the positive sample pair is the feature vector instance of two module units with the same design label, and the negative sample pair is the feature vector instance of two module units with different design labels. The goal of the loss function is to minimize the distance between the positive sample pair in the embedding space and maximize the distance between the negative sample pair. Through the back propagation algorithm, the model parameters are adjusted so that the feature vectors of the module units with similar design labels are adjacent in the embedding space, and the feature vectors of the module units with non-similar labels are far away. During the training process, the model continuously receives new module unit data, calculates the feature vectors through forward propagation, evaluates the prediction effect through the contrastive loss function, updates the parameters through back propagation, and finally obtains the pre-trained initial optimization model until the loss function converges (i.e. the loss value changes less than a set threshold in continuous multiple rounds of training). The final optimized model can effectively extract the topological features of the module units and make preliminary inferences on the module types and parameters based on historical design experience, laying a foundation for subsequent multi-objective optimization training.

[0047] Embodiment 2: refer to Figure 3 , the association matching process of the topological structure of the plurality of independent module units and the preset module topological mode needs to clearly define the specific type and standard template of the preset module topological mode. The association matching process is realized by using the method of “feature vector similarity matching + constraint condition verification”. A module topological mode feature library is constructed, which contains 4 typical modes of radial, single connection, double connection and multi-segment multi-connection. The feature vector of each mode contains 5 dimensions: node degree mean (Dmean) , number of connection lines (Lnum) , number of connected components (Cnum) , standard deviation of line load rate (Sdev) , number of power points (Pnum) , such as the feature vector of the radial mode , the feature vector of the single connection mode ; the feature vector of the to-be-matched module topological structure is calculated, Dmean is obtained by counting the number of connections of the nodes in the module, Lnum is obtained by counting the number of lines connected to different power points at both ends, Cnum is obtained by depth-first search, Sdev is obtained by calculating the deviation standard deviation of the actual load rate and the rated load rate of all lines, and Pnum ; the Euclidean distance between the feature vector to be matched and the feature vector of each mode in the mode library is calculated according to the following formula:

[0048]

[0049] wherein, is the feature to be matched, is the mode feature, and the mode with the smallest distance is selected as the preliminary matching result; finally, the constraint condition is checked, and the checking conditions include that the deviation between the total load of the module and the rated load capacity of the mode is not more than 15%, and the deviation between the length of the module line and the recommended length of the mode is not more than 20%, if the constraint is satisfied, the mode is determined as the associated matching result, if not, the second smallest distance mode is recalculated and checked, until the matching mode that meets the constraint is found. The preset module topology mode is usually determined based on the typical design specification of the power distribution network, including typical structures such as radial, single connection, double connection, multi-section multi-connection, etc. Each mode corresponds to a standardized topology template, which defines the specifications of node connection mode, the number range of edges, key device configuration requirements, etc. These templates are formed through historical operation data and design experience summary, and are stored in the model training database as a reference benchmark for associated matching.

[0050] When the topology structure of the target independent module unit is associated with each preset module topology mode, a method combining structural similarity calculation and statistical feature matching is adopted. The structural similarity calculation is based on the adjacency matrix in graph theory: an adjacency matrix is constructed for the target module unit, the rows and columns of the matrix correspond to the nodes in the module, and the element value of the matrix is the attribute of the edge between the nodes, if two nodes are directly connected, the element value is the impedance value of the corresponding edge, otherwise it is 0. For each preset module topology mode, a standard adjacency matrix template is also constructed, and the edge impedance value in the template is the average value of all similar edges in the mode. The cosine similarity of the adjacency matrix of the target module and the standard adjacency matrix of each mode is calculated, the formula is the sum of the product of the corresponding elements of the two matrices divided by the product of the lengths of the two matrices, and the structural similarity value corresponding to each mode is obtained, the value ranges between 0 and 1, and the closer to 1 indicates that the structure is more similar.

[0051] Statistical feature matching is realized by extracting the statistical features of the target module and each mode and calculating the Euclidean distance. The statistical features of the target module include node degree distribution, edge impedance mean value (average of all edge impedance values), number of connected components (number of independent power supply areas in the module), etc.; the statistical features of each preset mode are also obtained based on the historical data statistics in the template library. The Euclidean distance between the statistical feature vector of the target module and the statistical feature vector of each mode is calculated, and the statistical distance value corresponding to each mode is obtained, and the smaller the distance, the closer the statistical features.

[0052] The structural similarity value and the statistical distance value are weighted and fused to obtain a final correlation value. The weight distribution is determined according to the importance of historical data, for example, the structural similarity weight is set to 0.6, and the statistical distance weight is set to 0.4. The fusion formula is: final correlation value = structural similarity value x 0.6 + (1-statistical distance value) x 0.4 (the smaller the statistical distance value, the larger 1-statistical distance value, highlighting the positive contribution of statistical similarity). The correlation value threshold is determined according to the quantile of historical data, and the top 20% quantile of the correlation value of all historical module units is usually selected as the threshold (for example, if the correlation value of the historical data is concentrated between 0.4-0.8, the threshold is set to 0.6), to ensure that only the mode with high structural similarity to the target module is selected as the correlation object.

[0053] The preset module topology mode with a correlation value exceeding the threshold is taken as the target module topology mode corresponding to the topology structure of the target independent module unit, and the modes not selected are taken as the control module topology mode. For example, if the structural similarity value of the target module is 0.75, the statistical distance value is 0.3, the fused correlation value is 0.75 x 0.6 + (1-0.3) x 0.4 = 0.45 + 0.28 = 0.73, which is higher than the threshold 0.6, then the single-connection mode with the highest correlation value is selected as the target module topology mode, and other modes (such as radial mode and double-connection mode) are taken as the control mode.

[0054] When determining the comparison feature code according to the correlation parameters, the core feature code of the target module topology mode is extracted. The core features include the key connection mode defined by the mode, the device configuration parameters, and the topology constraints. These core features are extracted from the mode template through natural language processing technology and converted into numerical codes.

[0055] Then, the control module topology mode with significant differences from the target mode is screened, and its control feature code is extracted. The control features are the key differences between the control mode and the target mode, for example, when the target mode is a single-connection mode, the control mode can be a radial mode (without connection line) or a double-connection mode (with two connection lines). The control features of the radial mode include "no connection line" (coded as 0) and "existence of power outage area under N-1 fault" (coded as 1); the control features of the double-connection mode include "two connection lines" (coded as 2) and "still maintaining connection under N-1 fault" (coded as 1). These control features are also obtained through text extraction and numerical conversion.

[0056] Feature fusion processing is performed on the core feature codes and the control feature codes using a three-step fusion method: weight allocation, feature concatenation, and dimensionality compression. The first step determines the fusion weights for the core feature codes and the control feature codes. These weights are determined through feature importance evaluation, calculated using a random forest algorithm. The core feature codes and the control feature codes are used as inputs, and the module design effectiveness (e.g., power supply reliability, economic efficiency) is used as the output. After training the random forest model, the node splitting contribution of each feature is extracted as the importance weight. The average weight of the core feature codes is set to... The average weight of the feature encoding is set to The second step is to perform feature weighted concatenation to encode the core features. ( (Core feature dimension) and corresponding feature encoding ( To obtain the weighted features, multiply each of the corresponding feature dimensions by its corresponding weight. , Then, the two are spliced ​​together in sequence to obtain the initial fusion feature. The third step is to perform dimensionality compression. Principal Component Analysis (PCA) is used to reduce the dimensionality of the initial fused features. The formula for calculating the covariance matrix of the initial fused features is as follows:

[0057]

[0058] in, For the sample size, To obtain the initial fused feature mean, solve for the eigenvalues ​​and eigenvectors of the covariance matrix, and select the top features with a cumulative variance contribution rate exceeding 95%. Using the 10 feature vectors as principal component vectors, the initial fused features are projected onto the principal component vector space to obtain the final fused feature encoding. This encoding retains the key information of the core features, incorporates the difference information of the comparison features, and reduces the feature dimensionality, making it easier for subsequent model processing.

[0059] A method combining one-hot encoding concatenation and a weighted summation method based on an attention mechanism is employed. One-hot encoding concatenates the encoded vectors of core features and contrast features sequentially to form a high-dimensional feature vector. The attention mechanism assigns weights based on the contribution of features to the discriminative power of topological patterns, and converts the one-hot encoded vector into a low-dimensional, compact contrast feature encoding through weighted summation. This encoding contains both the core feature information of the target pattern and highlights the differences from the contrast pattern, effectively reflecting the discriminative power of different topological patterns.

[0060] After obtaining the running state data of each independent module unit, the key information related to the topology structure needs to be extracted. These data include the running parameters of each sub-unit (such as transformer, switch, line) in the module, such as the load rate of the transformer, the on-off state of the switch, the current value of the line, etc. Through the preprocessing of these data, the power parameter vector of each node unit (i.e. key device) can be obtained, which is composed of normalized voltage amplitude, current effective value, active power, reactive power and other parameters, reflecting the real-time electrical characteristics of the node unit. According to the connection relationship information in the running state data, the associated node units (i.e. other devices directly connected through lines) of each node unit and the strength vector of the connection relationship are determined, which is composed of line impedance, susceptance, historical maximum transmission capacity and other parameters, quantifying the closeness of the electrical connection between nodes.

[0061] The preset basic model adopts a graph neural network architecture, including multiple network interaction layers and at least one network fusion layer. The network interaction layer adopts a graph attention network (GAT) structure, the core of which is to dynamically capture the dependency relationship between nodes through attention mechanism, so as to more accurately extract local and global features in the topology structure. The network fusion layer adopts a gated recurrent unit (GRU) structure, which is used for time series modeling of the node feature sequence output by the network interaction layer, capturing the time series dependency relationship between nodes in the topology structure, and further optimizing the feature expression.

[0062] When the preset basic model is used to extract features from the topology structure of each independent module unit, the input data is processed. The input data includes the power parameter vector of each node unit in the topology structure of the target independent module unit, the list of associated node units, and the strength vector of the connection relationship. Among them, the power parameter vector is processed by normalization to eliminate the dimensional influence, such as converting the voltage value from 0-10kV to the standardized value of 0-1; the list of associated node units is determined by adjacency relationship, i.e. other node units directly connected with the current node unit; the strength vector of the connection relationship is obtained by logarithmic transformation of the line impedance, susceptance and other parameters, to compress the data range and highlight the key differences.

[0063] The first layer of the network interaction layer processes from the input data. The power parameter vector of each node unit is first subjected to a linear transformation, projecting the high-dimensional sparse original features to a low-dimensional dense space through a weight matrix to obtain a basic parameter vector. For each node unit, the basic parameter vectors of all its associated node units are collected and spliced with the basic parameter vector of the current node unit to form an interaction feature pair. An attention mechanism is introduced to dynamically calculate the attention degree of the current node unit to each associated node unit, which is specifically realized by calculating an attention coefficient. The calculation of the attention coefficient is based on the combination of the features of the current node unit and the associated node units, using a modified linear unit with leakage as the activation function, as follows:

[0064]

[0065] wherein: denotes the node The attention coefficient of the associated node of the node reflects the attention degree of the node to the features of the node denotes the natural exponential function; is a modified linear unit with leakage activation function, whose mathematical form is , which is used to introduce nonlinear characteristics and avoid gradient disappearance; is a learnable weight vector of the attention mechanism, which is optimized through model training; is the weight matrix of the linear transformation, which is also a model training parameter; and denote the basic parameter vectors of the node and the node ; denotes the vector splicing operation; denotes the set of all associated node units of the node ; denotes the summation operation on all adjacent nodes of the node ; denotes the transpose operation of the matrix. Through this formula, the model can dynamically allocate attention weights to the associated nodes for each node unit, thereby more accurately capturing the interaction importance between different nodes in the topology structure.

[0066] The calculated attention coefficient and the basic parameter vector of the associated node unit The first-order interaction aggregation vector of the current node unit is obtained by multiplying and summing. The vector fuses the basic features of the current node unit itself and the features of the associated node units, reflecting the interaction of the node in the local topology. Subsequently, a second nonlinear activation function (such as the Sigmoid function) is introduced to process the first-order interaction aggregation vector to generate the first-order parameter vector, further suppressing noise and enhancing the discriminability of the features.

[0067] The second layer and higher layers of the network interaction layer process the above-mentioned process repeatedly, but the input changes from the basic parameter vector to the first-order parameter vector output by the previous layer. Each layer dynamically aggregates the features of adjacent nodes through the attention mechanism, gradually extracting higher-order topological features. For example, the second layer interaction aggregation vector fuses the local adjacency information and the secondary adjacency information of the node, and the third layer further captures the global topological information. With the increase of the number of layers, the node features gradually transition from local interaction information to global topological features, and the final output layer aggregates the target-order parameter vectors of each node into a feature vector instance of the module unit through global average pooling operation. The instance contains the overall topological features of the module unit, such as connection density, correlation strength of key devices, and redundancy of the topology.

[0068] The network fusion layer receives the feature vector instance sequence (arranged in the node traversal order) output by the network interaction layer, and models the time sequence through a gated recurrent unit (GRU). The GRU includes an update gate and a reset gate, which control the preservation of historical information and the introduction of new information, respectively. The update gate determines the fusion ratio of the candidate state and the historical state at the current time, and the reset gate controls the influence degree of the historical state on the candidate state. Through the synergistic effect of the two gates, the GRU can capture the long-term dependence relationship in the node feature sequence, such as the time sequence change pattern of the device connection within the module. Finally, the network fusion layer outputs the feature vector instance corresponding to the target independent module unit topology, which integrates the local interaction features and global time sequence dependence features of the module, providing rich topological information representation for subsequent feature recognition and model training.

[0069] Embodiment 4: see Figure 4, consider a 10kV distribution network radiation module unit as an example, which mainly serves a small industrial park, including 1 distribution transformer (T-1), 2 feeder switches (S-1, S-2), 3 feeder lines (L-1, L-2, L-3) and 4 load nodes (N-1 to N-4). The module topology structure is: the high-voltage side of the transformer T-1 is connected to the 10kV bus, and the low-voltage side is connected to the line L-1 (with N-1, N-2 at the end) and the line L-2 (with N-3 at the end) through the switch S-1. Switch S-2 is used as a backup tie switch to connect the low-voltage side of T-1 and line L-3 (with N-4 at the end). The feature extraction process is described in detail below in conjunction with this example. Obtain the operating state data of the module unit, collect real-time monitoring information through intelligent sensors deployed on the equipment, and refer to Table 1 for specific data.

[0070] Table 1: The specific data of real-time monitoring information is as follows.

[0071]

[0072] When determining the associated node units of each node unit, the module topology connection relationship is used: the low-voltage side of the transformer T-1 is connected to the switches S-1 and S-2, so the associated nodes of T-1 are S-1 and S-2; one end of switch S-1 is connected to T-1, and the other end is connected to lines L-1 and L-2, so the associated nodes of S-1 are T-1, L-1 and L-2; one end of line L-1 is connected to S-1, and the other end is connected to loads N-1 and N-2, so the associated nodes of L-1 are S-1, N-1 and N-2; load N-1 is only connected to L-1, so the associated nodes of N-1 are L-1.

[0073] When obtaining the strength vector of the connection relationship, for each pair of associated nodes, the impedance, contact resistance and other parameters are extracted and standardized. For example:

[0074] The connection between T-1 and S-1 is a low-voltage bus connection, with a wire length of 0.2km, a type of JKLYJ-10-50, impedance parameters of resistance 0.05Ω / km×0.2km=0.01Ω, reactance 0.3Ω / km×0.2km=0.06Ω, and contact resistance 0.1mΩ. The average resistance of the same type of connection is 0.04Ω, with a standard deviation of 0.01Ω; the average reactance is 0.05Ω, with a standard deviation of 0.01Ω; the average contact resistance is 0.08mΩ, with a standard deviation of 0.02mΩ.

[0075] S-1 and L-1 are overhead lines, with a length of 1.8 km, a model of JKLYJ-10-95, and impedance parameters of resistance 0.2 Ω / km x 1.8 km = 0.36 Ω and reactance 0.4 Ω / km x 1.8 km = 0.72 Ω. The average resistance of the historical similar lines is 0.25 Ω, with a standard deviation of 0.05 Ω; the average reactance is 0.35 Ω, with a standard deviation of 0.05 Ω. The standardized impedance vector is [(0.36-0.25) / 0.05, (0.72-0.35) / 0.05] = [2.2, 7.4].

[0076] When calculating the first-order parameter vector, the power parameter vector of each node unit is first processed by feature conversion. For example, the original power parameter vector of transformer T-1 [1.5, 1.0, 1.0] is projected to a 4-dimensional basic parameter vector through a linear transformation matrix (model training parameter), eliminating redundant information and retaining core features such as load rate, voltage deviation, and reactive power compensation level, to obtain the basic parameter vector [1.2, 0.9, 1.1, 0.8].

[0077] The basic parameter vectors of the associated node units and the strength vectors of the connection relationships are integrated. Taking transformer T-1 as an example, its associated nodes are S-1 and S-2:

[0078] The basic parameter vector of S-1 (after feature conversion) is [2.0, 1.1, 0.9] (assuming dimension 3), which is spliced with the strength vector of T-1-S-1 [-3, 1, 1] to obtain the integrated vector [2.0, 1.1, 0.9, -3, 1, 1];

[0079] The basic parameter vector of S-2 (assuming [1.8, 0.8, 1.0]) is spliced with the strength vector of T-1-S-2 (assuming [-2, 0.5, 0.8]) to obtain the integrated vector [1.8, 0.8, 1.0, -2, 0.5, 0.8];

[0080] The two integrated vectors are added according to the weights (determined by model training, such as 0.6 and 0.4), to obtain the first-order interaction aggregation vector of T-1 [(2.0 x 0.6 + 1.8 x 0.4), (1.1 x 0.6 + 0.8 x 0.4),..., (-3 x 0.6 - 2 x 0.4)].

[0081] The input parameter vector of the current node unit (the result of the original power parameter vector after standardization [1.5, 1.0, 1.0]) is fused with the first-order interaction aggregation vector through a gating mechanism. The gating mechanism calculates the fusion coefficient (range 0-1) through a sigmoid function. For example, the input parameter vector weight is 0.7, the first-order interaction aggregation vector weight is 0.3, and the first-order parameter vector is generated after weighted summation. The vector integrates the running state of the node itself (such as transformer load rate) and the interaction information with the associated node (such as switch contact resistance, line impedance), reflecting the electrical characteristics and connection effect of the node in the topology structure.

[0082] Example 5: see Figure 5 The load parameters of each load point in the load distribution data are the basic input of the multi-objective optimization training. The type and range of the load parameters are determined by the actual operation scenario of the power distribution network, and usually include active power, reactive power, power factor, load growth rate, and time distribution of peak and valley periods, etc. These data mainly come from two parts: one is historical monitoring data, which is collected by intelligent meters, power distribution transformer monitoring terminals and other devices deployed at the user side at a frequency of minutes or hours, covering typical days (such as weekdays, weekends), typical weeks (such as summer, winter) and typical years (such as wet season, dry season). The operation situation; the second is the output result of the load prediction model, the commonly used prediction model includes long short-term memory network (LSTM), which learns the time dependence of historical data to predict the load change trend in the future period, for example, to predict the active power maximum value, reactive power fluctuation range of each hour in the next typical day. For example, the load point of a residential area shows in the historical data that the active power is concentrated between 120kW-180kW from 8am to 10pm on weekdays, the power factor is maintained at 0.85-0.90, and the reactive power is fluctuated between 90kVar-130kVar. The load peak on weekends appears from 10am to 4pm, the active power is reduced by about 30% compared with weekdays, and the power factor is increased to about 0.92. After sorting these data, the time series data set of each load point is formed as the input of subsequent model processing.

[0083] When analyzing each load parameter based on the pre-trained initial optimization model, the input layer of the model first receives the time series data of the load parameter. The initial optimization model adopts a bidirectional LSTM network structure, and the hidden layer thereof is stacked by multiple LSTM units, which can simultaneously capture the forward and backward dependencies of the time series. For example, for a sequence containing 24 hours of load data (one data point per hour), the forward LSTM unit of the bidirectional LSTM processes data from the 1st hour to the 24th hour to extract the load change characteristics from early to late; the backward LSTM unit processes data from the 24th hour to the 1st hour to extract the load change characteristics from late to early. The outputs of the LSTM units in the two directions are combined by splicing to form a feature representation that can reflect the change law of the load in the whole time period. The output of the hidden layer is further transmitted to the output layer, and the output layer maps the high-dimensional features to a low-dimensional space through a fully connected layer to generate an optimization feature instance corresponding to each load parameter. The instance is a low-dimensional vector, which contains information such as the time series change characteristics (such as the hourly load fluctuation amplitude) of the load, the peak-valley difference (such as the difference between the maximum load and the minimum load), the maximum load probability (such as the probability of the maximum load occurring in a certain period), and the like. For example, after the load parameter of a certain commercial building is processed by the model, the output optimization feature instance may show that the load of the building peaks from 18:00 to 21:00, the peak-valley difference is 80 kW, and the maximum load probability is 0.92. These features provide a key basis for subsequent feature recognition.

[0084] When the trained feature recognition component identifies each optimized feature instance, the feature recognition component adopts a convolutional neural network (CNN) structure, with the input being a one-dimensional form of the optimized feature instance. The hidden layer of the CNN contains multiple convolutional layers and pooling layers, which extract local key features by sliding window scanning of the input features through one-dimensional convolutional kernels. For example, the convolutional kernel size is set to 3, with a step size of 1. When each convolutional kernel slides over the input features, the weighted sum of the features at the current position and the previous and subsequent positions is calculated to generate a new feature map. Subsequently, the maximum pooling layer down-samples the feature map, retaining the maximum value in each local region and reducing the feature dimension while preserving the main information. After multiple layers of convolution and pooling operations, the output of the hidden layer is passed to the fully connected layer, which classifies the features through a softmax activation function and outputs the load inference result for each load parameter. The load inference result includes the load type (e.g., residential, commercial, industrial), the importance level (e.g., first, second, third), and the sensitivity to power supply reliability (e.g., high, medium, low). For example, if an optimized feature instance shows a small peak-valley difference (less than 30 kW), a low maximum load probability (less than 0.7), and a smooth load fluctuation (hourly fluctuation amplitude less than 10%), it may be classified as a "residential third-level load" with low sensitivity to power supply reliability. If the peak-valley difference is large (more than 100 kW), the maximum load probability is high (more than 0.9), and the load fluctuation is severe (hourly fluctuation amplitude more than 30%), it may be classified as an "industrial first-level load" with high requirements for power supply reliability.

[0085] When integrating and optimizing the pre-trained initial optimization model based on the load inference result of each load parameter, the multi-objective optimization is taken as the core target, and a multi-objective loss function containing economy, reliability, and power supply quality is constructed.

[0086] The economy objective function is as follows:

[0087]

[0088] wherein is the line length, is the unit length line cost, selected from the standard cost library according to the conductor model, such as JKLYJ-10-95 conductor Yuan / meter, is the number of transformers, is the cost of a single transformer, such as S13-500kVA transformer ten thousand Yuan, is the average annual power loss of the module unit, obtained by power flow calculation, Yuan / kWh is the electricity price. ​​​

[0089] The reliability objective function is as follows:

[0090]

[0091] where is the system average interruption duration index (unit: min / household), is the system average interruption frequency index (unit: times / household), , are the maximum allowable values specified in the distribution network design standard (such as min / household, times / household).

[0092] The power supply quality objective function is as follows:

[0093]

[0094] where is the number of module unit nodes, is the actual voltage of node , is the rated voltage (10 kV).

[0095] The multi-objective optimization algorithm adopts the non-dominated sorting genetic algorithm II (NSGA-II), and the specific implementation steps are as follows: 1. Initialize the population, and set the population size to 100. The individual is the design parameter vector of the module unit (including line cross section, transformer capacity, and switch configuration), and the parameter value range meets the distribution network design standard; 2. Calculate the values of the three objective functions of each individual, and divide the individual into different Pareto levels by using fast non-dominated sorting; 3. Calculate the crowding distance of the individual, which represents the distribution density of the individual on the Pareto frontier; 4. Generate the offspring population by roulette selection, single-point crossover (crossover probability 0.8), and polynomial mutation (mutation probability 0.05); 5. Merge the parent and offspring populations, repeat steps 2-3, and select the first 100 individuals to form the new generation population; 6. Iterate the above process to 100 generations, and finally select the smallest individual according to the engineering actual demand (such as giving priority to economy) from the optimal Pareto frontier of the 100th generation population, or select the The optimal design parameters are determined to complete the training of the power distribution network design model. The economic loss mainly considers the line investment cost, transformer loss cost, etc. For example, the longer the line length and the larger the conductor cross section, the higher the investment cost; the farther the transformer load rate deviates from the rated capacity, the higher the loss cost. The reliability loss mainly considers the power outage time expectation, power supply shortage, etc. For example, the loss caused by the power outage of the load with high sensitivity is much larger than that of the residential load. The power supply quality loss mainly considers the voltage deviation exceeding the standard rate and the harmonic distortion rate, etc. For example, the higher the proportion of time that the voltage deviation exceeds ±5%, the greater the power supply quality loss. The multi-objective loss function quantifies these indicators into numerical form, and the optimization effect of the model is comprehensively evaluated by weighted summation or the Pareto frontier method.

[0096] The integrated optimization training adopts a method combining a multi-objective genetic algorithm (NSGA-II) and back propagation. First, a population containing multiple model parameter combinations is initialized, and each parameter combination represents a possible model configuration (such as the number of LSTM units, the size of the convolution kernel, the initial value of the weight matrix, etc.). Each individual in the population calculates its corresponding optimization feature instance through forward propagation, and then obtains the load inference result through the feature recognition component, and further calculates the value of the multi-objective loss function as the fitness of the individual. After the fitness evaluation is completed, NSGA-II generates a new generation of population through selection, crossover and mutation operations: the selection operation preferentially retains individuals with high fitness (i.e., model configurations with small loss function values); the crossover operation exchanges the parameter combinations of two parent individuals to generate offspring individuals; the mutation operation randomly adjusts part of the parameters of the offspring individuals to increase the diversity of the population. The new generation of population repeats the above fitness evaluation, selection, crossover and mutation process until the population converges (i.e., the fitness of the individuals changes by less than a set threshold for consecutive generations). Finally, the optimized model parameter combination determines the trained power distribution network design model, which can output the optimal modular design scheme under multi-objective constraints according to the load distribution data of the to-be-designed scenario, such as configuring high-reliability tie lines for industrial first-level loads and configuring economic radial modules for residential third-level loads, thereby reducing investment costs and losses while meeting power supply reliability.

[0097] Example 6: Based on the upgrading and reconstruction project of a certain county 10 kV distribution network of State Grid Gansu Electric Power Company, the modular design method of distribution network is applied to carry out specific design work. Through the survey module of the distribution network field management and control system, the initial network parameters of the county distribution network are collected, including 12 existing 10 kV lines with a total length of 85 km, among which the wire types cover JKLYJ-10-70, JKLYJ-10-95 and JKLYJ-10-120, corresponding to the standard wires numbered DL001, DL002 and DL003 in the State Grid Gansu Electric Power Distribution Rural Network Engineering Standard Material; 42 distribution transformer areas, with S13-315 kVA and S13-500 kVA as the main transformer types, all selected from the solidified standard material library; switchgear includes 28 pole circuit breakers and 15 load switches, whose parameters meet the requirements of the provincial typical design scheme. Through the smart meter and distribution terminal, the load distribution data are collected, covering three categories of county urban residential load, industrial park industrial load and township agricultural load, among which the urban residential load is concentrated in 18:00-22:00, with an average active power of 650 kW and a power factor of 0.9; the industrial park load is relatively stable throughout the day, with an average active power of 1200 kW and a power factor of 0.88; the township agricultural load is significantly affected by the irrigation season, with an average active power of 480 kW and a power factor of 0.85 in the peak season.

[0098] Based on the above initial network parameters, according to the principle of matching between power supply area division and load characteristics, the county distribution network is divided into three independent module units: urban residential power supply module covering 6 lines and 20 areas in the core area of the county; industrial park power supply module covering 3 lines and 8 areas in the industrial park; and township agricultural power supply module covering 3 lines and 14 areas in the surrounding five townships. The two-step division method of "electrical distance + load clustering" is used to realize the division of independent module units. First, the electrical distance between the nodes of the distribution network is calculated, and the electrical distance formula is:

[0099]

[0100] wherein, , are the resistance and reactance of the line between nodes and , is the rated apparent power of the line, and the electrical distance matrix between all nodes is calculated through the formula; secondly, based on the electrical distance matrix, the load is clustered, and the K-means clustering algorithm is used, with the clustering number determined according to the power supply range and load density of the distribution network (such as the urban distribution network taking 5-8, and the township distribution network taking 3-5), and the clustering characteristics are the electrical distance and load power (active power , reactive power ), by iterative calculation to make the same cluster within the node electrical distance mean minimum, load power similarity maximum, clustering is completed, each cluster corresponds to an independent module unit, while ensuring that each module unit contains at least one power supply point (such as distribution transformer, ring network cabinet), and the number of inter-module unit tie lines is not more than 2, meet the electrical independence and operational flexibility requirements. Each module unit refers to the topological mode in the "State Grid Gansu Electric Power Distribution Rural Power Grid Engineering Typical Design", among which the urban module adopts single tie topology, the industrial park module adopts double tie topology, and the township module adopts radial topology, ensuring that the module division meets the electrical distance requirements and adapts to the load characteristics of different regions.

[0101] A power flow characteristic operator model is constructed, and a pre-defined power flow calculation function and a topological characteristic function are used to process the characteristics of three module units. The pre-defined characteristic function includes three types of core functions, namely electrical parameter mapping function, topological structure representation function and power flow correlation function. Among them, the electrical parameter mapping function adopts the following standardization normalization formula:

[0102]

[0103] Where x is the original electrical parameter of the module unit (such as line impedance, transformer load rate, node voltage), xmin and xmax are the minimum and maximum values of this type of parameter in the distribution network design standard, and the function is used to convert electrical parameters of different dimensions into standardized values in the interval [0, 1]; the topological structure representation function adopts the graph theory adjacency matrix conversion function as follows:

[0104]

[0105] Where A is the adjacency matrix of the module unit topological structure (1 if connected between nodes, 0 if not connected), D is the degree matrix (diagonal elements are the number of connections of the corresponding node, and the rest are 0), tr( ⋅ ) is the matrix trace operation, and the function converts the topological structure into a normalized feature matrix; the power flow correlation function adopts the sensitivity function derived from the power flow equation as follows:

[0106]

[0107] where P, Q are the active power and reactive power vectors of the module unit, and U is the node voltage vector. The function quantifies the correlation characteristics of power and voltage. During feature encoding, the line impedance, transformer capacity, node load, and other parameters of the module unit are standardized by the electrical parameter mapping function to obtain a 12-dimensional electrical feature vector; the adjacency matrix of the module topology is converted into an 8-dimensional topological feature matrix by the topological structure representation function; and the sensitivity matrix of power and voltage is calculated by the power flow correlation function to extract a 6-dimensional correlation feature vector. The three types of features are concatenated in the order of "electrical feature vector-topological feature matrix-correlation feature vector" to form a 26-dimensional module unit feature encoding result, which can be directly input into the subsequent model for training and reasoning. For urban residential power supply modules, the line connection mode, transformer load rate, and node voltage deviation are extracted, and parameters such as wire type JKLYJ-10-120 and pole-mounted circuit breaker ZW20-12 are converted into numerical codes in combination with standard material codes; for industrial park modules, the switching logic of double connection lines, power flow characteristics of heavy load nodes, and distributed power supply access interface parameters are encoded; and for rural agricultural modules, the end voltage loss of radial lines and seasonal load fluctuation coefficient are encoded. During the encoding process, the standard legend library in the design and cost integration module is called synchronously to associate the module topological structure with the graphic element information, forming a comprehensive feature encoding result containing electrical parameters, topological features, and material information.

[0108] An initial optimization model is obtained by pre-training, and the training data of the model is derived from the module data of 20 similar county distribution network reconstruction projects completed in Gansu Province in the past three years. Comparative learning is performed through an integrated training framework. The integrated training framework adopts a three-layer architecture of "basic model layer-feature fusion layer-decision output layer". The basic model layer integrates three core models: a graph neural network (GNN) model (used to extract module topological features), a long short-term memory (LSTM) model (used to capture time sequence features of module running states), and a gradient boosting tree (XGBoost) model (used to mine nonlinear correlations between module parameters and design effects). The feature fusion layer uses an attention mechanism to fuse the output features of the three basic models. The attention weight matrix ( is used to perform linear transformation and weighted summation on the topological feature vector of the GNN model, the time sequence feature vector of the LSTM model, and the correlation feature vector of the XGBoost model to obtain the fused feature vector as follows:

[0109]

[0110] The decision output layer outputs the preliminary inference result through a full connection network output module. The specific implementation of contrastive learning is as follows: first, construct a training sample pair, and encode the feature of two module units with a topological structure similarity greater than 0.8 in the same power distribution network scene as a positive sample pair , and encode the feature of two module units with a topological structure similarity less than 0.3 in different power distribution network scenes as a negative sample pair . The similarity is calculated by a cosine similarity formula as follows:

[0111]

[0112] wherein , are the feature encodings of the two module units respectively, and then define the contrastive loss function as follows:

[0113]

[0114] wherein is the number of positive sample pairs, is the number of negative sample pairs, is a temperature coefficient; finally, the contrastive loss function and the classification loss function (cross-entropy loss) of the basic model are weighted and summed to obtain the total loss function formula as follows:

[0115]

[0116] The parameters of the integrated training framework are iteratively optimized by a gradient descent algorithm, and when the total loss function decreases by less than in 10 consecutive iterations, the training is stopped, and a pre-trained initial optimization model is obtained. During the training, the data of a certain county module in Lanzhou with a similar topological structure to the current county module in the urban area is taken as a positive sample, and the data of a certain county and township module in Jiuquan with a large topological difference is taken as a negative sample, so as to optimize the discrimination ability of the model to the adaptability of different module topologies and loads through the comparison of the feature encoding results, and finally form an initial optimization model which has covered the common power distribution module types and load scenes in the province.

[0117] Based on the collected load distribution data, the initial optimization model is trained for multi-objective optimization. For the urban module, the line section selection and transformer distribution are optimized in combination with the peak-valley difference of residential load to ensure voltage stability at peak time and avoid equipment overload. For the industrial park module, the switching time and standby capacity configuration of the tie line are optimized around the continuity requirement of industrial load, while the 2022 version of the quota extraction rules of the cost software is interfaced to control the engineering investment while meeting the power supply reliability. For the rural agricultural module, the configuration position and capacity of the reactive power compensation device are optimized according to the load fluctuation in the irrigation season, taking into account the power supply quality and economy. During the optimization training process, the material price data and construction quota standards of the cost software are real-time retrieved through the design and cost integration module to link the adjustment of equipment selection and line layout with cost accounting, ensuring that each optimization direction meets the engineering cost control requirements.

[0118] The trained power distribution network design model is applied to the modular design of the county power distribution network reconstruction project, and the detailed design scheme of each module is output. The urban residential module finally determines to add 2 S13-500kVA transformers, upgrade 3 lines to JKLYJ-10-12 conductors, and add 1 tie line; the industrial park module adds 1 10kV box transformer, optimizes the protection setting value of 2 tie lines, and reserves 2 distributed photovoltaic access ports; the rural agricultural module installs reactive power compensation devices in individual transformer areas and replaces 2 aging lines with JKLYJ-10-95 conductors. After the design scheme is generated, it is pushed to the project management department of the county company through the design and cost integration module, and the county company organizes the owner, design, and price audit personnel to carry out online internal audit, checks the correspondence between design drawings and feature codes through the platform, and marks the transformer distribution problems that need to be modified; the modified scheme is uploaded to the city company review system, the city company mainly checks the module topology and load matching and cost reasonableness, and confirms that there is no error before pushing it to the provincial company; the provincial company combines the provincial power distribution network investment plan for final review, and after passing, the design drawings, material list, and cost data are synchronized to the power distribution site management and control system. The construction unit checks the point coordinates of the construction drawings through the mobile terminal of the site management and control system and constructs according to the drawings; after completion, the design unit generates the as-built drawing through the survey software and directly imports it into the cost software to carry out settlement. During the settlement process, the integration module retrieves the comparison information between the construction drawing and the as-built drawing to verify the material usage and design deviation, ensuring the accuracy of the settlement data. The whole design process realizes the full-process connection of module division, feature coding, model optimization, design review, and cost accounting, fully adapting to the actual needs of the county power distribution network.

[0119] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0120] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A modular design method for power distribution networks, characterized in that, include: Obtain initial network parameters and load distribution data of the power distribution network; Based on the initial network parameters, the distribution network is divided into multiple independent modular units; The process involves: acquiring operational status data of the multiple independent module units; determining the topology of each independent module unit based on the operational status data; extracting features from the topology of each independent module unit using a pre-defined base model employing a graph neural network (GNN) architecture to obtain feature vector instances for each independent module unit; recognizing each feature vector instance based on a trained feature recognition component to obtain design inference results for each independent module unit; and performing integrated optimization training on the pre-defined base model based on the design inference results for each independent module unit to obtain a pre-trained initial optimized model. The step of determining the topology of each independent module unit based on the operating status data includes: acquiring multiple pre-set module topology patterns; associating and matching the topology of each independent module unit with the multiple module topology patterns to obtain association parameters of the topology of each independent module unit; and determining the contrast feature code of each independent module unit based on the association parameters of the topology of each independent module unit. A pre-trained initial optimization model is obtained through comparative learning using an ensemble training framework. The ensemble training framework employs a three-layer architecture: a base model layer, a feature fusion layer, and a decision output layer. The base model layer integrates three core models: a Graph Neural Network (GNN) model for extracting module topology features; a Long Short-Term Memory (LSTM) model for capturing temporal features of module operation; and a Gradient Boosting Tree (XGBoost) model for uncovering the nonlinear correlation between module parameters and design performance. The feature fusion layer uses an attention mechanism to fuse the output features of the three base models, performing linear transformation and weighted summation through the attention weight matrix to obtain a fused feature vector. The decision output layer outputs preliminary inferences about the module design through a fully connected network. Based on the load distribution data, the pre-trained initial optimization model is subjected to multi-objective optimization training to obtain a fully trained distribution network design model. The trained distribution network design model is applied to perform modular design processing on the distribution network scenario to be designed, and the optimized modular design result of the distribution network is obtained.

2. The modular design method for power distribution networks according to claim 1, characterized in that, The step of associating and matching the topology of each independent module unit with the multiple module topology patterns to obtain the association parameters of the topology of each independent module unit includes: The target independent module unit topology structure is associated and matched with each module topology pattern to obtain the association value corresponding to each module topology pattern; the module topology pattern with the association value exceeding the association value threshold is taken as the target module topology pattern corresponding to the target independent module unit topology structure; the target module topology pattern corresponding to the target independent module unit topology structure is taken as the association parameter of the target independent module unit topology structure.

3. The modular design method for power distribution networks according to claim 2, characterized in that, The step of determining the contrast feature encoding of each independent module unit based on the association parameters of the topology of each independent module unit includes: Obtain the core feature code of the target module topology pattern corresponding to the target independent module unit topology structure; determine the corresponding feature code that is not associated with the target independent module unit topology structure among the multiple module topology patterns based on the core feature code; perform feature fusion processing on the core feature code and the corresponding feature code to obtain the comparison feature code of the target independent module unit topology structure.

4. The modular design method for power distribution networks according to claim 1, characterized in that, The preset basic model includes multiple network interaction layers and at least one network fusion layer. Based on the preset basic model using a graph neural network (GNN) architecture, feature extraction is performed on the topology of each independent module unit to obtain feature vector instances for each independent module unit, including: Based on the operational status data of the target independent module unit, obtain the power parameter vector of each node unit in the topology of the target independent module unit, the associated node units of each node unit, and the strength vector of the connection relationship between each node unit and the associated node units; perform feature transformation processing on the power parameter vector of each node unit to obtain the basic parameter vector of each node unit; integrate the basic parameter vector of the associated node units and the strength vector to obtain the integrated vector; obtain the first weight matrix of the first aggregation module for interactive aggregation processing, and perform linear transformation processing on the first weight matrix and the integrated vector to obtain the first transformation parameter matrix; process the first transformation parameter matrix based on a preset first nonlinear function to obtain the first order of the node unit. Interactive aggregation vector; For each node unit, the input parameter vector and the first-order interactive aggregation vector are subjected to parameter aggregation processing to obtain the first-order parameter vector of each node unit; Based on the target network interaction layer, the preceding-order parameter vector and the intensity vector of the associated node unit are subjected to interactive aggregation processing to obtain the target-order interactive aggregation vector of each node unit; The preceding-order parameter vector and the target-order interactive aggregation vector of each node unit are subjected to parameter aggregation processing to obtain the target-order parameter vector of each node unit; The target-order parameter vector of each node unit is used as the feature vector of each node unit; Based on the network fusion layer, the feature vector of each node unit is processed to obtain the feature vector instance corresponding to the topology of the target independent module unit.

5. The modular design method for power distribution networks according to claim 4, characterized in that, The step of performing parameter aggregation processing on the input parameter vector of each node unit and the first-order interaction aggregation vector to obtain the first-order parameter vector of each node unit includes: Obtain the second weight matrix of the second aggregation module used for parameter aggregation processing; perform linear transformation processing on the second weight matrix and the first-order interactive aggregation vector to obtain the second transformation parameter matrix; process the input parameter vector of each node unit and the second transformation parameter matrix based on the preset second nonlinear function to obtain the first-order parameter vector of each node unit.

6. The modular design method for power distribution networks according to claim 4, characterized in that, The step of obtaining the power parameter vector of each node unit in the topology of the target independent module unit, the associated node units of each node unit, and the strength vector of the connection relationship between each node unit and its associated node units based on the operating status data of the target independent module unit includes: Based on the operating status data of the target independent module unit, obtain the power index data of each sub-unit in the target independent module unit; perform feature coupling processing on the power index data of each sub-unit to obtain the power parameter vector of each node unit in the topology of the target independent module unit; take the reference sub-unit that has a power connection path with each sub-unit as the associated sub-unit of each sub-unit; obtain the power index data of the power connection path between each sub-unit and the associated sub-unit; perform feature coupling processing on the power index data of the power connection path to obtain the strength vector of the connection relationship between each node unit and the associated node unit.

7. The modular design method for power distribution networks according to claim 1, characterized in that, The step of performing multi-objective optimization training on the pre-trained initial optimization model based on the load distribution data to obtain a fully trained distribution network design model includes: The load parameters of each load point in the load distribution data are obtained; each load parameter is analyzed and processed based on the pre-trained initial optimization model to obtain the optimization feature instance corresponding to each load parameter; each optimization feature instance is identified based on the completed training feature recognition component to obtain the load inference result of each load parameter; the pre-trained initial optimization model is integrated and optimized based on the load inference result of each load parameter to obtain the completed training distribution network design model.

8. A modular design system for power distribution networks, characterized in that, Includes a server, the server being used to perform the method according to any one of claims 1-7.

Images