A distributed power distribution management method and system for power distribution cabinets

Abstract

This invention discloses a distributed power distribution management method and system for power distribution cabinets, relating to the field of intelligent power management technology. The method includes collecting a heterogeneous set of operating status parameters, mapping the three-dimensional coordinates of the power distribution cabinets to nodes in the topology, mapping the electrical connections of the power distribution cabinets to edges in the topology, generating a power distribution network topology graph, and using an improved multi-scale graph diffusion algorithm to search for anomaly propagation patterns based on the topology graph to obtain fault propagation paths. The spatiotemporal propagation features in the fault propagation paths are extracted and fused to form a fault feature vector. The fault feature vector and the power distribution network topology graph are input into a spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer. This invention improves the dynamic response capability and fault identification accuracy of the power distribution process, and enhances the overall operational reliability of the power distribution cabinet, through the spatiotemporal graph convolutional model and the dynamic power distribution network topology graph.

Description

Technical Field

[0001] This invention relates to the field of intelligent power management technology, and in particular to a distributed power distribution management method and system for a power distribution cabinet. Background Technology

[0002] With the rapid development of smart grids, power distribution systems are gradually evolving towards greater intelligence and distribution. As a key node in power distribution and control, the distribution cabinet plays a crucial role in the power distribution system. It is responsible not only for the safe and stable distribution of power from substations to end users, but also for real-time monitoring and protection functions, ensuring the normal operation of the entire distribution network. In the face of power outages caused by sudden load fluctuations, the distribution cabinet can quickly adjust its operating mode, ensuring uninterrupted power supply to critical users through automatic reclosing, load transfer, and other measures.

[0003] Existing power distribution management methods still have some limitations. On the one hand, traditional GCN models struggle to accurately predict state evolution trends after load transfer, thus affecting the dynamic response capability of the power distribution process and resulting in insufficient control precision of the power distribution scheme. On the other hand, traditional topology diagrams typically use static diagram structures to describe the connection relationships between distribution cabinets, failing to fully consider the dynamic evolution of the three-dimensional spatial location and electrical attributes of equipment, leading to inaccurate topology representation and limiting the accuracy of fault mode identification. Summary of the Invention

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

[0005] Therefore, the present invention provides a distributed power distribution management method and system for power distribution cabinets, which solves the problems of insufficient dynamic response capability and low fault identification accuracy in the power distribution process.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a distributed power distribution management method for a power distribution cabinet, which includes: collecting a heterogeneous set of operating status parameters, mapping the three-dimensional coordinates of the power distribution cabinet to nodes of the topology, mapping the electrical connection relationship of the power distribution cabinet to edges of the topology, and generating a power distribution network topology diagram. Based on the power distribution network topology, an improved multi-scale graph diffusion algorithm is used to search for anomaly propagation patterns, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths and fuse them to form a fault feature vector. The fault feature vector and the distribution network topology are input into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme. The automatic differential engine converts the optimal power distribution adjustment scheme into control commands, drives the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collects real-time operation feedback data, and optimizes the power distribution network topology.

[0007] As a preferred embodiment of the distributed power distribution management method for the power distribution cabinet described in this invention, the heterogeneous operating status parameter set includes the three-dimensional coordinates of the power distribution cabinet, the electrical connection relationship of the power distribution cabinet, real-time electrical parameters, and real-time load data.

[0008] As a preferred embodiment of the distributed power distribution management method for the power distribution cabinet described in this invention, the specific steps for generating the power distribution network topology diagram are as follows: The IEC protocol stack is used to map the three-dimensional coordinates of the power distribution cabinet to the equipment type attributes to obtain the nodes of the topology; the topology tracing algorithm is used to map the electrical connection relationships of the power distribution cabinet to the equipment association attributes to generate the edges of the topology. The nodes and edges are optimized by two-way weighting and multi-attribute fusion to generate a power distribution network topology.

[0009] In a preferred embodiment of the distributed power distribution management method for the power distribution cabinet described in this invention, the specific steps for obtaining the fault propagation path are as follows: An improved multi-scale graph diffusion algorithm is used to perform dynamic weight diffusion and multi-scale subgraph segmentation on the power distribution network topology to obtain the abnormal state distribution; the abnormal state distribution is then hierarchically coordinated and iterated to form an abnormal propagation pattern. A Gaussian kernel function is applied to perform dynamic neighborhood sampling on the anomaly propagation pattern to obtain candidate fault nodes; polynomial fitting is then performed on the candidate fault nodes to generate the fault propagation path.

[0010] In a preferred embodiment of the distributed power distribution management method for the distribution cabinet described in this invention, the specific steps for forming the fault feature vector are as follows: The fault propagation path is decomposed into multiple scales using Fast Fourier Transform to extract spatiotemporal propagation features; Cross-modal coupling of spatiotemporal propagation features is performed through Tucker decomposition to obtain a high-dimensional correlation tensor; the high-dimensional correlation tensor is then concatenated to form a fault feature vector.

[0011] As a preferred embodiment of the distributed power distribution management method for the power distribution cabinet described in this invention, the specific steps for generating the optimal power distribution adjustment scheme are as follows: In the PyTorch framework, spatial and temporal convolutional layers are invoked, and residual connections are used to stack the spatial and temporal convolutional layers across multiple scales to construct a spatiotemporal graph convolutional model. The fault feature vector and the distribution network topology are input into the spatiotemporal graph convolutional model. The spatial convolutional layer uses the graph attention mechanism to capture the neighborhood influence of the fault node and generate the neighborhood propagation feature matrix. The temporal convolutional layer predicts the state evolution after load transfer through causal dilated convolution and obtains the temporal evolution trend vector; The neighborhood propagation feature matrix and the temporal evolution trend vector are concatenated through cross-modal feature channels to output the optimal power distribution adjustment scheme.

[0012] As a preferred embodiment of the distributed power distribution management method for the power distribution cabinet described in this invention, the specific steps for optimizing the power distribution network topology are as follows: The optimal power distribution adjustment scheme is backpropagated through an automatic differential engine to obtain the switch adjustment differential parameters, which are then mapped into control commands through an IEC protocol converter. Based on control commands, the circuit breakers in the power distribution cabinet are driven to perform sectional switching and opening and closing, and dynamic power distribution is carried out through a dynamic load balancing algorithm. Simultaneously, RTU equipment is used to collect real-time operation feedback data. Real-time operational feedback data is input into the power distribution network topology diagram, and node voltage similarity analysis is performed through dynamic time warping to obtain impedance correction coefficients; Based on the impedance correction coefficient, the voltage sensitivity weights of the nodes are iteratively updated, and the impedance parameters of the edges are optimized.

[0013] Secondly, the present invention provides a distributed power distribution management system for a power distribution cabinet, including a topology mapping module, a fault tracing module, a scheme generation module, and a closed-loop control module. The topology mapping module is used to collect heterogeneous operating status parameter sets, map the three-dimensional coordinates of the power distribution cabinets to nodes of the topology structure, map the electrical connection relationships of the power distribution cabinets to edges of the topology structure, and generate a power distribution network topology diagram. The fault tracing module is used to search for abnormal propagation patterns based on the power distribution network topology using an improved multi-scale graph diffusion algorithm, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths and fuse them to form a fault feature vector. The scheme generation module is used to input the fault feature vector and the distribution network topology into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme. The closed-loop control module is used to convert the optimal power distribution adjustment scheme into control commands through the automatic differential engine, drive the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collect real-time operation feedback data, and optimize the power distribution network topology.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the distributed power distribution management method for the power distribution cabinet as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the distributed power distribution management method for a power distribution cabinet as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by constructing a spatiotemporal graph convolution model, it is possible to accurately model the state evolution trend after load transfer, which significantly improves the dynamic response capability and control accuracy of the power distribution process; at the same time, by combining the three-dimensional spatial position and electrical attributes of the equipment to construct a dynamic power distribution network topology, the ability to characterize the fault propagation path is enhanced, which effectively improves the accuracy of fault identification and the overall operational reliability of the power distribution cabinet. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a distributed power distribution management method for a power distribution cabinet.

[0019] Figure 2 This is a schematic diagram of a distributed power distribution management system for a power distribution cabinet.

[0020] Figure 3 A flowchart generated for a power distribution network topology diagram.

[0021] Figure 4 A flowchart for generating fault feature vectors. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figure 1 As one embodiment of the present invention, this embodiment provides a distributed power distribution management method for a power distribution cabinet, comprising the following steps: S1 collects heterogeneous operating status parameter sets, maps the three-dimensional coordinates of the distribution cabinet to nodes of the topology, maps the electrical connection relationships of the distribution cabinet to edges of the topology, and generates a power distribution network topology diagram.

[0026] Reference Figure 3 Specifically, the operations include the following: S1.1, Collect heterogeneous operating status parameter set, which includes the three-dimensional coordinates of the power distribution cabinet, the electrical connection relationship of the power distribution cabinet, real-time electrical parameters and real-time load data; The three-dimensional coordinates of the distribution cabinet are longitude, latitude, and altitude. Longitude and latitude are obtained through GNSS positioning units; altitude is collected using a barometer. The three-dimensional coordinates of the distribution cabinet are used to accurately express the distribution location of the distribution cabinet in physical space, providing a spatial basis for constructing a dynamic power distribution network topology map. The electrical connection relationships of the distribution cabinet include the wiring method, branch connection structure, and equipment association attributes. The wiring method is collected using a cable detector, the branch connection structure is obtained through a circuit diagram, and the equipment association attributes are collected using a power monitoring unit. The electrical connection relationships of the distribution cabinet describe the logical connection relationships of the distribution cabinet and are an important basis for generating the edge structure in the topology graph. Real-time electrical parameters include voltage, current, and AC power; voltage is collected using a voltage transformer, current is obtained using a current transformer, and AC power is collected using a power analyzer; real-time electrical parameters reflect the current operating status of the distribution cabinet and are an important data source for fault detection and condition assessment. Real-time load data includes active and reactive loads of each branch; active loads are collected using smart meters, and reactive loads are collected using power quality analyzers; real-time load data is used to analyze the load distribution characteristics of the distribution network and to support the formulation of load transfer and optimized scheduling strategies.

[0027] S1.2, preprocessing the heterogeneous operating state parameter set. Specifically, Daubechies wavelet is used to perform multi-scale time-frequency decomposition on the heterogeneous operating state parameter set to suppress high-frequency noise. A sliding window statistical method is used to limit the heterogeneous operating state parameter set to a 3σ range within a fixed window (e.g., 5 sampling periods) to filter out outliers. Neville interpolation is applied to perform non-uniform sampling on the heterogeneous operating state parameter set to fill in missing values ​​and ensure the continuity of the heterogeneous operating state parameter set. Dynamic time warping is used to perform time series alignment on the heterogeneous operating state parameter set to eliminate sampling time offset. Simultaneously, PTP protocol is used for cross-device clock calibration to ensure timestamp alignment. Finally, Z-score normalization is used for dimensional normalization to eliminate dimensional differences, and the preprocessed heterogeneous operating state parameter set is output. It should be noted that the 3σ range is defined based on the sliding window local statistical properties of the heterogeneous operating state parameter set.

[0028] S1.3, the three-dimensional coordinates of the distribution cabinet are mapped to nodes in the topology, and the electrical connections of the distribution cabinet are mapped to edges in the topology, generating a power distribution network topology diagram. Specifically, the three-dimensional coordinates of the distribution cabinet are analyzed in real-time and normalized using a spatial coordinate analysis method to obtain the spatial location vectors of the equipment. The ASN.1 encoding rules of the IEC protocol stack are used to perform semantic classification and logical labeling of the spatial location vectors to generate equipment type attribute identifiers. A hash algorithm is used to uniquely encode the equipment type attribute identifiers to obtain the equipment type attributes. Simultaneously, the equipment type attributes are mapped to nodes in the topology using the MMS object dictionary service of the IEC protocol stack. A topology tracing algorithm is used to perform a depth-first traversal of the electrical connections of the distribution cabinet and extract impedance parameters. The impedance parameters are then normalized using Z-score standardization to form an initial connection feature set. Principal component analysis is used to reduce the dimensionality of the initial connection feature set and weighted aggregation is performed to obtain low-dimensional feature vectors. These low-dimensional feature vectors are then multidimensionally fused using a weighted concatenation method to generate composite vectors of related attributes. Finally, these composite vectors of related attributes are mapped using the Neo4j graph database index to obtain the edges of the topology. Neighborhood aggregation and feature transformation are performed on the nodes and edges of the topology to obtain initial node features and initial edge features. In the bidirectional weight optimization stage, the initial node features are coupled in multiple dimensions using a linear weighting method to generate a weighted feature vector. The weighted feature vector is then normalized and integrated to form node weights. The node weights are then subjected to nonlinear transformation and polynomial fitting using a polynomial regression algorithm to generate sensitivity coefficients. Based on the sensitivity coefficients, the node weights are backpropagated using gradient descent to obtain optimized node weights. The initial edge features are separated into components using a complex factorization algorithm to extract impedance features. Load factor weighting is applied to the impedance features to obtain edge weights. The edge weights are then normalized to per-unit values ​​using a baseline capacity normalization method to generate standardized edge weights. Mean sampling is applied to sparsify the standardized edge weights, and a weight update matrix is ​​constructed. The alternating direction multiplier method is used to perform distributed iteration on the weight update matrix to obtain updated weights. The updated weights are then amplitude-limited (range [0,1]) to obtain optimized edge weights. It should be noted that per-unit conversion refers to the process of standardizing the edge weights using reference capacity and reference voltage. Per-unit conversion achieves dimensionless conversion of edge weights. In the multi-attribute fusion stage, based on the optimized node weights and optimized edge weights, the adjacency aggregation operator is used to aggregate the neighborhood information of the initial node features and initial edge features to obtain higher-order feature representations. Different attention weights are assigned to the higher-order feature representations, and feature enhancement and dimensionality compression are performed through weighted summation to generate a weighted feature matrix. The weighted feature matrix is ​​then axially concatenated and nonlinear transformation and multidimensional fusion are performed through the softmax function to generate the power distribution network topology.

[0029] S2. Based on the power distribution network topology, an improved multi-scale graph diffusion algorithm is used to search for anomaly propagation patterns, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths, and fuse them to form a fault feature vector.

[0030] Reference Figure 4 Specifically, the operations include the following: S2.1, an improved multi-scale graph diffusion algorithm is used to perform dynamic weight diffusion and multi-scale subgraph segmentation on the distribution network topology to obtain the distribution of abnormal states. Specifically, electrical connection features are extracted from the distribution network topology, and binary adjacency encoding is performed on these features to obtain node connection relationships. Based on these relationships, the improved multi-scale graph diffusion algorithm is used to perform graph traversal and neighborhood state aggregation on the distribution network topology to generate node state vectors. Multiple rounds of iterative weight diffusion are then performed on these node state vectors. During the diffusion process, wavelet transform is used for frequency domain decomposition to obtain multi-frequency band feature components. These multi-frequency band feature components are then weighted and fused using a gating unit to generate a multi-scale feature space. Subsequently, feature decomposition is used to perform vector mapping on the multi-scale feature space to obtain a low-dimensional embedding representation. Cluster centers are generated by performing joint optimization on the low-dimensional embedding representation using silhouette coefficients. Based on these cluster centers, a multi-scale subgraph segmentation is performed on the multi-scale feature space using a spectral clustering algorithm to obtain a hierarchical subgraph structure. Anomaly distribution quantization is performed on the hierarchical subgraph structure using the inverse matrix solution method to obtain Mahalanobis distance. Interquartile range normalization is then performed on the Mahalanobis distance to generate a joint anomaly score. The specific mathematical formula is as follows. ; in, Indicates joint anomaly score, The weighting coefficients represent the Mahalanobis distance. This represents the normalized scaling factor for the Mahalanobis distance. Represents Mahalanobis distance, This represents the mean vector in a hierarchical subgraph structure. This represents the weighting coefficient of the interquartile range normalization term. Indicates the third quartile. Indicates the first quartile. Indicates the reference interquartile range; It should be noted that the silhouette coefficient is defined based on the ratio of cluster center compactness to separation, and its value range is [-1, 1]; the interquartile range is defined based on the median of the global distribution of Mahalanobis distance, and its value range is (0, +∞); the normalization scaling factor is defined based on the reciprocal of the largest eigenvalue of Mahalanobis distance, and its value range is (0, 1]. The joint anomaly scores are nonlinearly smoothed using a Savitzky-Golay filter and time-aligned using dynamic time warping to obtain an aligned anomaly score sequence. Gaussian kernel density estimation is used to perform probability density fitting on the anomaly score sequence to obtain the anomaly state distribution. The improved multi-scale graph diffusion algorithm makes the following improvements based on the existing multi-scale graph diffusion algorithm: on the one hand, it optimizes the dynamic weight diffusion process and improves the dynamic characteristic perception capability of the distribution network topology graph; on the other hand, it couples wavelet transform for real-time frequency domain decomposition, realizing the effective fusion of multi-frequency band feature components.

[0031] S2.2, hierarchical coordination and iteration of the abnormal state distribution to form an abnormal propagation pattern. In specific operations, a hierarchical message passing mechanism is used to perform multiple rounds of dynamic diffusion of the abnormal state distribution to obtain the initial abnormal topology; hierarchical coordination and iteration of the initial abnormal topology is performed to obtain the abnormal propagation map; the abnormal propagation map is sub-graph segmented to generate candidate abnormal regions; a sliding window is applied to resample the candidate abnormal regions at equal intervals to form topologically related abnormal clusters; Haar wavelets are used to perform three-level decomposition and frequency domain alignment on the topologically related abnormal clusters, and low-frequency fundamental components and high-frequency detail components are extracted; the low-frequency fundamental components and high-frequency detail components are then fused with energy weighting to obtain a multi-scale abnormal feature matrix; a nonlinear kernel function (such as RBF or polynomial kernel) is used to perform kernel space projection and spatiotemporal feature transformation on the multi-scale abnormal feature matrix to output the abnormal propagation pattern.

[0032] S2.3, a Gaussian kernel function is applied to perform dynamic neighborhood sampling on the anomaly propagation pattern to obtain candidate fault nodes. Polynomial fitting is then performed on these candidate fault nodes to generate the fault propagation path. Specifically, the Gaussian kernel function is used to calculate the node similarity of the anomaly propagation pattern, and the node similarity is normalized to obtain the Gaussian weights. The specific mathematical formula is as follows. ; in, Represents a node To the node Gaussian weights; Represents a node To the node similarity, Indicates the normalization factor; It should be noted that the normalization factor is defined based on the sum of Gaussian kernel values ​​of node similarity, and its value range is (0.3, 3.5]. The neighborhood topology of the anomaly propagation pattern is expanded by a graph traversal algorithm to obtain the k-hop neighborhood of the anomaly propagation pattern; according to Gaussian weights, probability sampling is performed on the k-hop neighborhood of the anomaly propagation pattern to obtain initial candidate nodes, and density clustering is performed on the initial candidate nodes to form candidate fault nodes. The timestamp sequence of candidate fault nodes is extracted. Based on the timestamp sequence, dynamic time warping is used to align the candidate fault nodes temporally to generate a spatiotemporal trajectory point set. Least square linear regression is used to fit the spatiotemporal trajectory point set with a second-order polynomial to generate a fault propagation parabola. The fault propagation parabola is then mapped inversely using least square projection to obtain the fault propagation path.

[0033] S2.4 Extracts and fuses the spatiotemporal propagation features from the fault propagation path to form a fault feature vector. Specifically, a Fast Fourier Transform (FFT) is used to perform frequency domain energy analysis on the fault propagation path to obtain the original spectrum. The original spectrum is then decomposed into multi-scale frequency bands. The multi-scale frequency bands are framed using a Hanning window to extract the amplitude and phase spectra as spatiotemporal propagation features. Next, the spatiotemporal propagation features are decoupled into higher-order features using Tucker decomposition, and regularized using Graph Laplacian to construct a three-dimensional feature tensor. The first mode of the three-dimensional feature tensor is the spatial node dimension, the second mode is the time scale dimension, and the third mode is the frequency domain component dimension. Alternating least squares is used to iteratively optimize the three-dimensional feature tensor to obtain the core tensor. The core tensor is then coupled across modes to generate a high-dimensional correlation tensor. Principal component analysis is used to compress the feature dimension of high-dimensional correlation tensors, and the tensor is spliced ​​using a variance-weighted fusion method to generate fault feature vectors.

[0034] S3 inputs the fault feature vector and the distribution network topology into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme.

[0035] Specifically, the following operations are included: S3.1, Construct and train a spatiotemporal graph convolutional model. Specifically, in the PyTorch framework, a graph convolutional network is invoked using the `nn.ChebConv` parameter. The input dimension of the graph convolutional network is set to 64, the number of output channels to 128, and the Chebyshev polynomial order to 2. A ReLU function is applied after the graph convolutional network for non-linear activation to enhance feature representation. An attention mechanism is used for dynamic neighborhood weight allocation to complete the construction of the spatial convolutional layer. Simultaneously, a temporal convolutional architecture is invoked using the `nn.Conv1d` parameter, with the kernel size set to 3 and the dilation rate set to doubling layer-by-layer. A LeakyReLU function is applied after the temporal convolutional architecture for non-linear transformation to capture multi-scale temporal patterns. A Dropout layer is used for random deactivation to prevent overfitting, thus completing the construction of the temporal convolutional layer. Next, residual connections are used to linearly project the spatial and temporal convolutional layers to match the number of channels. Simultaneously, strided convolutions are used to adjust dimensionality consistency, resulting in aligned spatial and temporal convolutional layers. The aligned spatial and temporal convolutional layers are then stacked hierarchically using the nn.ModuleList class to form a multi-scale architecture. After each stack, a downsampling convolution is inserted after the multi-scale architecture for feature compression to reduce computational complexity. Max pooling is used for downsampling to ensure that the feature resolution decreases layer by layer. Finally, global temporal pooling is applied to the multi-scale architecture to complete the construction of the spatiotemporal graph convolutional model. Next, the spatiotemporal graph convolutional model is trained. Further, fault feature vectors and the power distribution network topology are loaded in batches using the DataLoader parameterization function to generate graph data batches. These batches are divided into a sample set, a training set, and a validation set. On the sample set, StandardScaler (mean-variance normalization) is used to normalize the graph data batches using Z-score, and feature enhancement is performed through random time window sliding to form enhanced samples. On the training set, the Adam optimizer is used to update the parameters of the enhanced samples, and the Focal Loss function is applied simultaneously to calculate the training loss. On the validation set, the training loss is backpropagated to obtain the validation accuracy. When the validation accuracy exceeds the convergence threshold for five consecutive rounds, training terminates, and the trained spatiotemporal graph convolutional model is output using torch.save. It should be noted that the convergence threshold is based on the definition of the exponentially weighted moving average of the training loss, and its value ranges from [0.001, 0.01].

[0036] S3.2, the spatial convolutional layer uses a graph attention mechanism to capture the neighborhood influence of faulty nodes and generate a neighborhood propagation feature matrix. In specific operations, a graph convolutional network is used to extract features from the graph data to generate initial graph data, and feature mapping is performed on the initial graph data to form initial node feature representations. The graph attention mechanism calculates attention weights on the initial node feature representations through four parallel attention heads. Each attention head contains an attention vector. The specific calculation process is as follows: the initial node feature representations are linearly transformed using the softmax function, and attention scores are calculated using dot product operations. During the calculation process, weighting is adjusted using the inverse of impedance to reflect the influence of different impedance lines in the distribution cabinet on fault propagation. At the same time, the attention scores are normalized using the LeakyReLU activation function to obtain normalized attention weight coefficients. ; in, This represents the normalized attention weight coefficients. Represents the attention vector. This indicates the transpose operation. Indicates the first The linear transformation weight matrix of each node. Indicates the first The linear transformation weight matrix of each node. This represents the reciprocal of the impedance between nodes; It should be noted that the reciprocal of the impedance is obtained by taking the reciprocal of the real-time monitored line impedance value, with a value range of (0.05, 1]. After obtaining the attention weight coefficients, each attention head performs weighted neighborhood aggregation on the initial node feature representation and outputs a weighted neighborhood feature representation. The weighted neighborhood feature representations output by the four attention heads are concatenated along the feature dimension to form a multi-head fusion feature. Finally, the multi-head fusion feature is non-linearly mapped through a GRU (Gated Recurrent Unit) to output a neighborhood propagation feature matrix.

[0037] In S3.3, the temporal convolutional layer predicts the state evolution after load transfer through causal dilated convolution, obtaining a temporal evolution trend vector. Specifically, causal dilated convolution is used to extract temporal features from batches of graph data to obtain multi-scale temporal features. A multi-level receptive field is constructed using progressively increasing dilation rates on these multi-scale features. A gated linear unit is applied to perform a nonlinear transformation on the multi-level receptive field, and LayerNorm is used for normalization to output gated temporal features. Residual connections are used to perform skip connections on the gated temporal features, and a dropout layer is used to perform random feature deactivation to suppress irrelevant temporal patterns, resulting in stable temporal features. A 1×1 convolution is used to superimpose the stable temporal features and batches of graph data across modalities, outputting enhanced temporal features. The enhanced temporal features are then compressed using global average pooling to generate a temporal evolution trend vector.

[0038] S3.4, the neighborhood propagation feature matrix and the temporal evolution trend vector are concatenated through a cross-modal feature channel to output the optimal power distribution adjustment scheme. Specifically, the spatial dimension neighborhood propagation feature matrix and the temporal dimension temporal evolution trend vector are unified using a 1×1 convolution. On this unified feature dimension, different weights are assigned to the neighborhood propagation feature matrix and the temporal evolution trend vector, and they are concatenated using a cross-modal feature channel to form a comprehensive feature tensor. This comprehensive feature tensor reflects the synergistic characteristics of the dynamic spatial topology and temporal evolution law of the power distribution cabinet during the power distribution process. Next, the comprehensive feature tensor is input into a two-layer graph convolutional network for processing: the first layer uses 256 neurons with a ReLU activation function to abstract features and obtain high-order fusion features; the second layer uses 128 neurons to further structure the high-order fusion features and generate scheme generation features. The optimal power distribution adjustment scheme is generated by nonlinear mapping of the scheme generation features using gated convolution and amplitude control through gradient pruning.

[0039] S4 uses an automatic differential engine to convert the optimal power distribution adjustment scheme into control commands, drive the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collect real-time operation feedback data, and optimize the power distribution network topology.

[0040] Specifically, the following operations are included: S4.1, the automatic differentiation engine performs gradient backpropagation on the optimal power distribution adjustment scheme to obtain the switch adjustment differential parameters, which are then mapped into control commands through the IEC protocol converter. Specifically, the forward differentiation function of the automatic differentiation engine performs linear parameter transformation on the optimal power distribution adjustment scheme to obtain the Jacobian matrix parameters; the impedance-weighted loss function is used to perform gradient weighting on the Jacobian matrix parameters to generate a fine-tunable switch state; the gradient calculator of the automatic differentiation engine performs backpropagation on the fine-tunable switch state to obtain the original gradient parameters; the Softmax function is used to probabilistically sample the original gradient parameters to achieve gradient pass-through, obtaining the switch adjustment differential parameters; then, the switch adjustment differential parameters are converted into a specific control point table using the logic node mapping function of the IEC protocol converter, and the specific control point table is standardized into command encoding to output the control commands.

[0041] S4.2, based on control commands, drives the circuit breaker in the distribution cabinet to perform sectional switching operations and dynamically distributes power using a dynamic load balancing algorithm. Simultaneously, RTU devices collect real-time operational feedback data. Specifically, control commands are sent to the distribution cabinet via the IEC MMS protocol, driving the control unit to parse the commands, obtain the target switch action parameters, and perform a deep search of the target switch action parameters using a network topology traversal algorithm to confirm that the switch action will not lead to islanded operation, generating a safe operation command set. Based on the safe operation command set, the circuit breaker coil in the distribution cabinet is driven to perform sectional switching operations. During sectional switching, electrical quantities are sampled using synchronous phasor measurement to obtain the three-phase electrical quantities of the feeder. The dynamic load balancing algorithm is used to perform impedance-weighted optimization of the three-phase electrical quantities of the feeder to obtain optimized power supply parameters. Simultaneously, multi-time-period trend fitting is performed on the optimized power supply parameters to generate the optimal power supply path, and dynamic power distribution is completed based on the optimal power supply path. Meanwhile, RTU devices are deployed in the distribution cabinet to collect real-time data on the three-phase power of the feeder and the switch status at fixed intervals (such as 100ms), and the data is integrated to generate real-time operation feedback data.

[0042] S4.3 Input real-time operational feedback data into the distribution network topology diagram. Perform node voltage similarity analysis using dynamic time warping to obtain impedance correction coefficients. Specifically, the data mapping engine maps the real-time operational feedback data to the corresponding nodes in the distribution network topology diagram, forming a dynamic topology diagram with time-series attributes. Then, the dynamic time warping (DTW) algorithm is used to align the multi-node voltage sequences of the dynamic topology diagram with time-series attributes to obtain a time-series difference matrix. The least squares method is used to perform polynomial regression fitting on the time-series difference matrix to generate voltage time-series curves. The dynamic time bending distance of the voltage time-series curves is then calculated to form node voltage similarity values. The node voltage similarity values ​​reflect the differences in voltage fluctuation trends between nodes and can quantify the spatial-temporal coupling imbalance in the distribution process. The node voltage similarity values ​​are resampled at equal intervals and filtered by moving average through a sliding time window to generate impedance correction coefficients. The specific mathematical formula is as follows: ; in, This represents the impedance correction factor. The inverse representation of node voltage similarity values. Indicates the optimal number of samples. Indicates the rated voltage; It should be noted that the rated voltage is based on the nominal voltage level definition specified in the IEC 60038 standard; the inverse characterization of the node voltage similarity value refers to the normalized result of the dynamic time bending distance of the voltage sequence curve.

[0043] S4.4, based on the impedance correction coefficient, the voltage sensitivity weights of the nodes are iteratively updated, and the impedance parameters of the edges are optimized. Specifically, the distribution network topology is sampled using a graph traversal method to extract branch impedance parameters. These parameters are then matrixed and filled with elements to construct the node admittance matrix. The impedance correction coefficient is incorporated as a weighting factor into the node admittance matrix to form a weighted admittance matrix. Newton's iteration method is used to perform weight iteration and linearization on the weighted admittance matrix to obtain the sensitivity weight matrix. Simultaneously, a partial differential equation is performed on the distribution network topology using a voltage sensitivity algorithm to obtain the node voltage sensitivity weights. Based on the sensitivity weight matrix, the node voltage sensitivity weights are iteratively updated using gradient descent. During the iteration process, the Adam optimizer is used for dynamic coefficient adjustment and parameter limiting to complete the update of the node voltage sensitivity weights. Based on the impedance correction coefficient, the branch impedance parameters are weighted and scaled to obtain the side admittance matrix. Principal component analysis is then used to fit the side admittance matrix to generate side impedance adjustment weights. Dynamic sampling of the distribution network topology is performed using a sliding window to extract path impedance features. Linear interpolation is applied to smooth the path impedance features to form side impedance parameters. Based on the side impedance adjustment weights, the side impedance parameters are dynamically iteratively optimized using conjugate gradients, while a parameter limiter is simultaneously applied for safety limiting, generating optimized impedance parameters and completing the optimization of the side impedance parameters.

[0044] Reference Figure 2 This embodiment also provides a distributed power distribution management system for a power distribution cabinet, including: a topology mapping module, a fault tracing module, a scheme generation module, and a closed-loop control module. The topology mapping module is used to collect heterogeneous operating status parameter sets, map the three-dimensional coordinates of the power distribution cabinets to nodes of the topology structure, map the electrical connection relationships of the power distribution cabinets to edges of the topology structure, and generate a power distribution network topology diagram. The fault tracing module is used to search for abnormal propagation patterns based on the power distribution network topology using an improved multi-scale graph diffusion algorithm, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths and fuse them to form a fault feature vector. The scheme generation module is used to input the fault feature vector and the distribution network topology into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme. The closed-loop control module is used to convert the optimal power distribution adjustment scheme into control commands through the automatic differential engine, drive the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collect real-time operation feedback data, and optimize the power distribution network topology.

[0045] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distributed power distribution management method of the power distribution cabinet proposed in the above embodiment.

[0046] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0047] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the distributed power distribution management method for the distribution cabinet as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0048] In summary, by constructing a spatiotemporal graph convolution model, this invention can accurately model the state evolution trend after load transfer, significantly improving the dynamic response capability and control accuracy of the power distribution process. At the same time, by combining the three-dimensional spatial location and electrical attributes of the equipment to construct a dynamic power distribution network topology, it enhances the ability to depict fault propagation paths and effectively improves the accuracy of fault identification and the overall operational reliability of the power distribution cabinet.

[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A distributed power distribution management method for a power distribution cabinet, characterized in that: include: Collect heterogeneous operating status parameter sets, map the three-dimensional coordinates of the power distribution cabinets to nodes of the topology, map the electrical connection relationships of the power distribution cabinets to edges of the topology, and generate a power distribution network topology graph. Based on the power distribution network topology, an improved multi-scale graph diffusion algorithm is used to search for anomaly propagation patterns, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths and fuse them to form a fault feature vector. The fault feature vector and the distribution network topology are input into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme. The automatic differential engine converts the optimal power distribution adjustment scheme into control commands, drives the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collects real-time operation feedback data, and optimizes the power distribution network topology.

2. The distributed power distribution management method for the distribution cabinet as described in claim 1, characterized in that: The heterogeneous operating status parameter set includes the three-dimensional coordinates of the power distribution cabinet, the electrical connection relationship of the power distribution cabinet, real-time electrical parameters, and real-time load data.

3. The distributed power distribution management method for the distribution cabinet as described in claim 1, characterized in that: The specific steps for generating the power distribution network topology diagram are as follows: The IEC protocol stack is used to map the three-dimensional coordinates of the power distribution cabinet to the equipment type attributes to obtain the nodes of the topology; the topology tracing algorithm is used to map the electrical connection relationships of the power distribution cabinet to the equipment association attributes to generate the edges of the topology. The nodes and edges are optimized by two-way weighting and multi-attribute fusion to generate a power distribution network topology.

4. The distributed power distribution management method for the distribution cabinet as described in claim 3, characterized in that: The specific steps for obtaining the fault propagation path are as follows: An improved multi-scale graph diffusion algorithm is used to perform dynamic weight diffusion and multi-scale subgraph segmentation on the power distribution network topology to obtain the abnormal state distribution; the abnormal state distribution is then hierarchically coordinated and iterated to form an abnormal propagation pattern. A Gaussian kernel function is applied to perform dynamic neighborhood sampling on the anomaly propagation pattern to obtain candidate fault nodes; polynomial fitting is then performed on the candidate fault nodes to generate the fault propagation path.

5. The distributed power distribution management method for the distribution cabinet as described in claim 1, characterized in that: The specific steps for forming the fault feature vector are as follows: The fault propagation path is decomposed into multiple scales using Fast Fourier Transform to extract spatiotemporal propagation features; Cross-modal coupling of spatiotemporal propagation features is performed through Tucker decomposition to obtain a high-dimensional correlation tensor; the high-dimensional correlation tensor is then concatenated to form a fault feature vector.

6. The distributed power distribution management method for the distribution cabinet as described in claim 5, characterized in that: The specific steps for generating the optimal power distribution adjustment scheme are as follows: In the PyTorch framework, spatial and temporal convolutional layers are invoked, and residual connections are used to stack the spatial and temporal convolutional layers across multiple scales to construct a spatiotemporal graph convolutional model. The fault feature vector and the distribution network topology are input into the spatiotemporal graph convolutional model. The spatial convolutional layer uses the graph attention mechanism to capture the neighborhood influence of the fault node and generate the neighborhood propagation feature matrix. The temporal convolutional layer predicts the state evolution after load transfer through causal dilated convolution and obtains the temporal evolution trend vector; The neighborhood propagation feature matrix and the temporal evolution trend vector are concatenated through cross-modal feature channels to output the optimal power distribution adjustment scheme.

7. The distributed power distribution management method for the distribution cabinet as described in claim 1, characterized in that: The specific steps for optimizing the power distribution network topology are as follows: The optimal power distribution adjustment scheme is backpropagated through an automatic differential engine to obtain the switch adjustment differential parameters, which are then mapped into control commands through an IEC protocol converter. Based on control commands, the circuit breakers in the power distribution cabinet are driven to perform sectional switching and opening and closing, and dynamic power distribution is carried out through a dynamic load balancing algorithm. Simultaneously, RTU equipment is used to collect real-time operation feedback data. Real-time operational feedback data is input into the power distribution network topology diagram, and node voltage similarity analysis is performed through dynamic time warping to obtain impedance correction coefficients; Based on the impedance correction coefficient, the voltage sensitivity weights of the nodes are iteratively updated, and the impedance parameters of the edges are optimized.

8. A distributed power distribution management system for a power distribution cabinet, based on the distributed power distribution management method for the power distribution cabinet according to any one of claims 1 to 7, characterized in that: It includes a topology mapping module, a fault tracing module, a solution generation module, and a closed-loop control module. The topology mapping module is used to collect heterogeneous operating status parameter sets, map the three-dimensional coordinates of the power distribution cabinets to nodes of the topology structure, map the electrical connection relationships of the power distribution cabinets to edges of the topology structure, and generate a power distribution network topology diagram. The fault tracing module is used to search for abnormal propagation patterns based on the power distribution network topology using an improved multi-scale graph diffusion algorithm, obtain fault propagation paths, extract spatiotemporal propagation features from the fault propagation paths and fuse them to form a fault feature vector. The scheme generation module is used to input the fault feature vector and the distribution network topology into the spatiotemporal graph convolutional model. The spatial convolutional layer captures the neighborhood influence of the fault node, and the temporal convolutional layer predicts the state evolution after load transfer, generating the optimal distribution adjustment scheme. The closed-loop control module is used to convert the optimal power distribution adjustment scheme into control commands through the automatic differential engine, drive the power distribution cabinet to perform segmented switch opening and closing power distribution, synchronously collect real-time operation feedback data, and optimize the power distribution network topology.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the distributed power distribution management method for the power distribution cabinet according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the distributed power distribution management method for the power distribution cabinet according to any one of claims 1 to 7.

Images