High-proportion new energy power distribution network complex dynamic topology identification method and system based on improved GCN

Abstract

The invention discloses a high-proportion new energy power distribution network complex dynamic topology identification method based on an improved GCN. The method comprises the steps of obtaining time sequence node feature data and an adjacent matrix of a to-be-detected high-proportion new energy power distribution network; processing the time sequence node feature data by using a time convolutional network to generate a dynamic node feature matrix; inputting the dynamic node feature matrix and the adjacent matrix into a trained multilayer graph convolutional network containing a residual connection mechanism, learning a long-distance dependency relationship between the nodes, and dynamically calculating a similarity weight between the nodes by adopting a strategy based on an adaptive basis function to capture short-distance dynamic dependency between the nodes; and according to a final node embedding vector output by the multilayer graph convolutional network, calculating a connection probability between the node pairs, and completing identification of the to-be-detected high-proportion new energy power distribution network complex dynamic topological structure. According to the method, the problems of insufficient adaptability, low calculation efficiency, poor expandability and the like during complex dynamic topology identification of the high-proportion new energy power distribution network can be solved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a method and system for identifying complex dynamic topology in high-proportion renewable energy distribution networks based on an improved GCN. Background Technology

[0002] The power output of new energy sources is highly volatile and intermittent, leading to frequent reconfigurations of the distribution network topology and even the formation of numerous unknown or temporary network connections. This highly uncertain, time-varying, and complex dynamic topology poses a severe challenge to core functions of the distribution network, such as state estimation, power flow calculation, security and stability assessment, fault location, and demand-side response. It has become a critical technical challenge that urgently needs to be addressed for the safe and efficient operation of current high-proportion new energy distribution networks.

[0003] Existing methods for distribution network topology identification mostly rely on building multi-classification models with fixed topology label sets, making it difficult to adapt to the dynamic evolution characteristics of topology structures under high penetration rates of renewable energy. In-depth research into existing complex topology network identification techniques and their application to distribution network topology identification can better address the growing demands. Complex topology network identification can effectively analyze the association patterns between nodes and reveal their interaction mechanisms. With ongoing research, studies on the problem of complex topology network identification can be summarized as driven response methods and deep learning methods.

[0004] Drive-response methods, which actively apply external stimuli (such as pulse control or adaptive feedback) and infer network connectivity based on system response, have achieved some success in identifying the topology of passive complex dynamic networks. Some researchers have used multi-timescale pulse control to identify passive networks; others have successfully identified passive networks with arbitrary inter-layer connections based on adaptive feedback controllers; still others have inferred unknown topologies by constructing drive-response synchronization mechanisms. However, high-proportion renewable energy distribution networks are essentially active networks composed of multiple coupled active units, whose nodes possess autonomous adjustment and energy injection capabilities. Therefore, the aforementioned drive-response methods for passive systems are difficult to directly apply. Furthermore, these methods typically have high computational complexity, high implementation costs, and face scalability bottlenecks when the node scale increases.

[0005] Deep learning methods can identify not only passive topologies but also active dynamic networks, relying solely on node state measurement data. Some researchers have inferred the topology of active networks based on variational autoencoders and graph neural networks (GNNs). However, each node in a GNN layer has its own weights, which can lead to parameter explosion when dealing with large-scale, high-proportion renewable energy distribution networks, affecting identification accuracy and speed. Other researchers have used graph convolutional neural networks (GCNs) to identify topologies in cases where some nodes are observable. Still others have applied GCNs to identify the topological states of microgrids. GCNs offer advantages such as all nodes sharing the same weights within the same layer, easier parameter training and optimization, better feature extraction capabilities, faster inference speed, and greater scalability, making them suitable for learning topological spatial dependencies in complex dynamic networks with high proportions of renewable energy distribution. In addition, variants such as Graph Attention Network (GAT) and Graph Isomorphic Network (GIN) have also been used to improve the ability to model node relationships. For example, some researchers have used GAT to extract key features and combined it with graph pooling to achieve topology discrimination, while others have enhanced GIN's ability to perceive complex connection patterns by using a random node discarding strategy.

[0006] In summary, although the theory of complex network topology identification is relatively mature, the unique active, dynamic, and partially observable characteristics of distribution networks under the background of high-proportion renewable energy integration place higher demands on existing methods. Furthermore, considering the strong fluctuations and intermittent nature of renewable energy output, there is an urgent need for a method for identifying complex dynamic topologies in distribution networks with high proportions of renewable energy, to address the shortcomings of existing technologies in dealing with the insufficient adaptability, low computational efficiency, and poor scalability of complex dynamic topology identification in such networks. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method and system for identifying complex dynamic topology of high-proportion renewable energy distribution networks based on an improved GCN, which addresses the shortcomings of the existing technology. This method can solve the problems of insufficient adaptability, low computational efficiency and poor scalability in the identification of complex dynamic topology of high-proportion renewable energy distribution networks.

[0008] To achieve the above objectives, according to one aspect of the present invention, a method for identifying complex dynamic topology in high-proportion renewable energy distribution networks based on an improved GCN is provided, comprising: Obtain the time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested; The temporal node feature data is processed using a Temporal Convolutional Network (TCN) to enhance the temporal dependency expression of node features and generate a dynamic node feature matrix. The dynamic node feature matrix and the adjacency matrix are input into a trained multi-layer graph convolutional network containing a residual connection mechanism. Multi-hop neighbor information is aggregated through stacked graph convolutional layers to learn the long-distance dependencies between nodes. In each layer of the multi-layer graph convolutional network, a strategy based on adaptive basis functions is used to dynamically calculate the similarity weights between nodes to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes. Based on the final node embedding vector output by the multi-layer graph convolutional network, the connection probability between node pairs is calculated to identify the complex dynamic topology of the high-proportion new energy distribution network under test.

[0009] The above scheme further includes: training a multi-layer graph convolutional network containing a residual connection mechanism by acquiring historical data of time-series node features and adjacency matrix of a high-proportion new energy distribution network.

[0010] In the above scheme, the processing of the temporal node feature data using a temporal convolutional network specifically includes: using a temporal convolutional network with causal convolution and dilated convolutional structures to capture the long-term and short-term dependency features of node states with different dilation rates.

[0011] In the above scheme, the characteristic is that the obtained adjacency matrix needs to undergo symmetric normalization processing to avoid the influence of node degree differences on feature propagation.

[0012] In the above scheme, the multilayer graph convolutional network including the residual connection mechanism includes: The output of each layer is added to the initial feature matrix through residual connections to prevent gradient vanishing and network degradation.

[0013] In the above scheme, the multi-layer graph convolutional network containing the residual connection mechanism preferably has 2 layers.

[0014] In the above scheme, the adaptive basis function-based strategy is specifically as follows: by calculating the cosine similarity between node embedding vectors and mapping it through an activation function, dynamic attention weights are obtained, which are used to weighted aggregate the features of neighboring nodes.

[0015] In the above scheme, the activation function is ReLU.

[0016] In the above scheme, the connection probability between the calculated node pairs is obtained by performing a dot product operation on the node embedding vectors and then mapping it through a probability mapping function.

[0017] In the above scheme, the probability mapping function is the Sigmoid function.

[0018] According to another aspect of the present invention, a complex dynamic topology identification system for high-proportion renewable energy distribution networks based on an improved GCN is provided, comprising: The data acquisition module is used to acquire time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested; The node feature enhancement module is used to process the temporal node feature data using a temporal convolutional network, enhance the temporal dependency expression of node features, and generate a dynamic node feature matrix. The processing module is used to input the dynamic node feature matrix and the adjacency matrix into a trained multi-layer graph convolutional network containing a residual connection mechanism, and to aggregate multi-hop neighbor information through stacked graph convolutional layers to learn the long-distance dependencies between nodes; in each layer of the multi-layer graph convolutional network, a strategy based on adaptive basis functions is used to dynamically calculate the similarity weights between nodes to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes; The topology identification output module is used to calculate the connection probability between node pairs based on the final node embedding vector output by the multi-layer graph convolutional network, thereby completing the identification of the complex dynamic topology of the high-proportion new energy distribution network under test.

[0019] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: (1) This invention provides a method for identifying complex dynamic topology of high-proportion new energy distribution networks based on improved GCN. This method can solve the problems of insufficient adaptability, low computational efficiency and poor scalability in the identification of complex dynamic topology of high-proportion new energy distribution networks.

[0020] (2) The present invention proposes a node feature enhancement strategy based on Temporal Convolutional Network (TCN), which uses TCN with different expansion rates and network layer stacking to extract the long-term dependence and short-term volatility of node states; proposes a Close-range Dynamic Dependency Capture (CDDC) strategy based on adaptive basis functions, which dynamically adjusts the information propagation weights between nodes by adjusting the basis functions of the nodes, handles the dynamic changes in the relationship between adjacent nodes, and captures the close-range dynamic dependence of high-proportion new energy distribution networks; proposes a long-range dependency learning strategy based on residual mechanism, which extracts long-range dependence by deep stacking of GCN layers, considering the mutual influence between multi-hop nodes, and improves the training efficiency of deep networks and prevents gradient vanishing and network degradation through residual connections. Attached Figure Description

[0021] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference figures denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating a method for identifying complex dynamic topology of a high-proportion renewable energy distribution network based on an improved GCN, as described in Embodiment 1 of the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the working principle of the improved GCN in Embodiment 1 of the present invention.

[0023] Figure 3 This is a schematic diagram of a complex dynamic topology identification algorithm for a high-proportion new energy distribution network based on an improved GCN, according to Embodiment 1 of the present invention.

[0024] Figure 4 This is a schematic diagram of causal convolution in Embodiment 1 of the present invention.

[0025] Figure 5 This is a schematic diagram of dilated convolution in Embodiment 1 of the present invention.

[0026] Figure 6 This is a schematic diagram of the long-distance dependency learning strategy based on residual connections in Embodiment 1 of the present invention.

[0027] Figure 7 This is a schematic diagram of the near-range dynamic dependency capture strategy based on adaptive basis functions in Embodiment 1 of the present invention.

[0028] Figure 8 This is a schematic diagram showing the performance comparison results of various algorithms for complex topology identification in Embodiment 1 of the present invention.

[0029] Figure 9 This is a schematic diagram of the identification process of the complex topology identification model of a high-proportion new energy distribution network in Embodiment 1 of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0031] It should be understood that the sequence number of each step in the embodiment does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0032] Example 1 This application provides a method for identifying complex dynamic topology in high-proportion renewable energy distribution networks based on an improved GCN. Please refer to [link / reference]. Figure 1 ,include: S1: Obtain the time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested.

[0033] In this embodiment, it can be understood that the core idea of ​​GCN (Graph Convolutional Networks) is that the representation of each node is not only determined by its own features, but also dynamically updated through information interaction with its neighboring nodes. This method can effectively capture the structural information of the graph and the complex relationships between nodes, thereby learning richer node representations. Its working principle is as follows: Figure 2 As shown. Therefore, this embodiment can use GCN for complex topology identification of high-proportion renewable energy distribution networks.

[0034] In practical applications, given a graph G The input data for GCN consists of two parts: the first is a node feature matrix, which contains the initial information of each node in the graph; the second is an adjacency matrix, which describes the connection relationships between nodes in the graph. Therefore, this embodiment needs to obtain the time-series node feature data and adjacency matrix of the high-proportion renewable energy distribution network to be tested. After inputting the data, the core calculation formula of GCN is as follows: (1) In the formula, For the first l The node feature matrix of the layer, when l When =0, The input node feature matrix; The adjacency matrix is ​​a symmetric standardized matrix; For the first l The learnable weight matrix of the layer; in this embodiment Choose the ReLU activation function.

[0035] The adjacency matrix needs to be symmetrically normalized to avoid the impact of node degree differences on feature propagation. The formula is as follows: (2) (3) In the formula, D It is a node degree matrix; I It is an identity matrix used to add self-loops to each node. This is the new node degree matrix. The standardized adjacency matrix. It can balance the feature propagation intensity of nodes with different degrees.

[0036] Based on this, the graph convolution process can be represented as: (4) GCN uses graph convolutional layers to perform information transfer and feature learning on nodes, generating output features for each node. For link prediction tasks, GCN captures potential connections between nodes by generating an embedding vector for each node. The connection probability between each pair of nodes is calculated by performing a dot product on the embedding vectors of those pairs. i and nodes j The probability that there is a connection can be expressed as (5) In the formula, and These are nodes i and nodes j The embedding vector. In this embodiment, the probability value is mapped using the Sigmoid function; the larger the output probability value, the higher the probability that there is an edge between the two nodes.

[0037] Based on the characteristics of distribution network node graph data, the GCN model has the following three limitations when applied to the complex topology identification task of high-proportion renewable energy distribution networks: 1) The node characteristics in the distribution network change dynamically with factors such as load changes, while GCN assumes that the node characteristics are fixed in each calculation, and cannot effectively capture the temporal dependence of node characteristics; 2) The connections between distribution network nodes are not limited to directly adjacent nodes, but GCN's graph convolution operation can only propagate single-layer information and cannot utilize extensive structural information; 3) The relationships and importance between different nodes in the distribution network are not fixed. GCN's graph convolution operation propagates information based on the adjacency matrix, but the weights in this process are fixed, ignoring the differences in the contributions of different adjacent nodes to the target node.

[0038] To address the problems encountered when applying Graph Convolutional Networks (GCNs) to complex topology identification in high-proportion renewable energy distribution networks, this embodiment proposes corresponding improvement strategies. To address the issue that GCNs cannot effectively capture the temporal dependencies of node features, this embodiment proposes a node feature enhancement strategy based on temporal convolutional networks. To address the issue that GCNs cannot utilize extensive structural information, this embodiment proposes a long-range dependency learning strategy based on residual mechanisms. To address the issue that GCN's graph convolution operations ignore the differences in contributions from different neighboring nodes to the target node, this embodiment proposes a short-range dynamic dependency capture strategy based on adaptive basis functions. Based on these three improvement strategies, a method for complex topology identification in high-proportion renewable energy distribution networks based on improved GCNs is proposed, as follows: Figure 3 As shown in the figure. The following sections of this embodiment will provide a detailed description of the above strategies.

[0039] S2 utilizes a temporal convolutional network to process temporal node feature data, enhancing the temporal dependency expression of node features and generating a dynamic node feature matrix.

[0040] Specifically, in this embodiment, a node feature enhancement strategy based on temporal convolutional networks is proposed: In complex topology network identification tasks, the quality of node features directly affects the predictive performance of the model. This embodiment considers the temporal and dynamic nature of distribution network node features, which traditional GCN models may struggle to effectively model. To enhance the temporal expressiveness of node features, this embodiment introduces a node feature enhancement strategy based on Temporal Convolutional Network (TCN).

[0041] The input to TCN is a time-series node feature matrix. Let N be the number of nodes, T be the number of time steps, and F be the feature dimension of each node. Assume that at time t, the node feature matrix is... Then the entire time series feature matrix can be represented as .

[0042] In TCN, the output of each time step of a causal convolution operation depends only on the current and past inputs, without accessing future information. The design of causal convolution guarantees the causality of time series data, making it suitable for time-series data. In causal convolution, the output... It is determined by the current and past parts of the input sequence, as shown in the following formula. (6) In the formula, These are the weights of the convolution kernel; It is a time step tk Input features; K This represents the kernel size. A schematic diagram of causal convolution is shown below. Figure 4 As shown.

[0043] Causal convolutions can meet the needs of time series modeling, but when the input time series has a long stride, directly using causal convolutions to extract time information leads to a sharp increase in the number of network layers. As the number of network layers increases, the training difficulty of the network also increases. In order to ensure that the network can handle long time series while keeping the number of neural network layers controllable and avoiding excessive network depth, this embodiment of TCN uses dilated convolutions to achieve this goal.

[0044] Dilated convolution expands the receptive field by inserting gaps between kernel elements, thereby capturing dependencies over longer periods with fewer layers. The specific convolution operation is as follows: (7) In the formula, The interval between each convolutional kernel element is... r ; r This represents the dilation rate, i.e., the spacing between each kernel element. By dilating the convolutions, TCNs can capture long-term dependencies in fewer convolutional layers, as illustrated in the diagram below. Figure 5 As shown. Finally, the processed node feature matrix can be represented as (8) In summary, this embodiment employs a node feature enhancement strategy, dynamically adjusting the feature representation of each node. The new node feature matrix enhances the temporal information within it, helping subsequent models better capture temporal dependencies and improve their predictive ability regarding relationships between nodes.

[0045] S3 inputs the dynamic node feature matrix and adjacency matrix into a pre-trained multi-layer graph convolutional network containing a residual connection mechanism. Multi-hop neighbor information is aggregated through stacked graph convolutional layers to learn the long-distance dependencies between nodes. In each layer of the multi-layer graph convolutional network, a strategy based on adaptive basis functions is used to dynamically calculate the similarity weights between nodes to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes.

[0046] It is understood that in this embodiment, the multi-layer graph convolutional network containing residual connection mechanism is trained by acquiring the historical data of time-series node features and adjacency matrix of high-proportion new energy distribution network, so as to obtain the trained multi-layer graph convolutional network containing residual connection mechanism.

[0047] Specifically, in this embodiment, a long-distance dependency learning strategy based on the residual mechanism is proposed: Power transmission is a complex process involving multiple nodes, not just directly connected nodes, but also upstream and downstream nodes. Therefore, topology identification tasks require full utilization of high-order neighbor information. However, traditional single-layer GCNs can only capture local information and struggle to cover global topological relationships and long-distance dependencies. Therefore, this embodiment proposes a long-distance dependency learning strategy based on a residual mechanism.

[0048] By stacking graph convolutional networks, the convolutional operations in each layer not only propagate information about the current node's direct neighbors, but also accumulate information about higher-order neighbors layer by layer. Specifically, assuming... A It is an adjacency matrix. X Let be the node feature matrix, then the th The feature update formula for layer +1 is: (9) (10) As the number of layers increases, the power of the adjacency matrix increases, expanding the information propagation range and enabling the extraction of features from more distant nodes. However, with a large number of layers, GCNs may encounter problems such as vanishing gradients or excessive feature smoothing. To address this issue, a residual connection mechanism is introduced, consisting of multiple stacked graph convolutional networks and residual connections, such as... Figure 6 As shown.

[0049] Graph convolutional layers need to directly learn the mapping function from input to output, and deep networks often face training difficulties. The core idea of ​​residual connections is "learning residuals," meaning that the learning objective of each layer is the difference between the input and output, reducing the learning difficulty. The computation process after L layers of feature learning is as follows: (11) In the formula, This is the initial characteristic matrix.

[0050] In summary, the long-distance dependency learning strategy based on residual mechanism in this embodiment captures the indirect relationships between nodes, constructing a more accurate and reliable topology. Especially in complex topologies or in the presence of noise, it can improve identification accuracy and effectively suppress gradient vanishing.

[0051] Specifically, in this embodiment, a long-distance dependency learning strategy based on the residual mechanism is also proposed: Node representations in GCNs are typically linear combinations of fixed weight matrices and neighboring node features. However, this approach may not fully capture the nonlinear relationships and potential complex dependencies between nodes in complex network structures, especially graph structures with intricate topologies like distribution networks. To overcome this limitation, this embodiment introduces a near-range dynamic dependency capture strategy based on adaptive basis functions, such as... Figure 7 As shown, the propagation weights are dynamically adjusted by calculating the relationships between nodes.

[0052] In GCN, node representations are updated through a fixed graph convolution operation. In the k The feature representation of layer +1 is as follows: (12) In the formula, Represents a node Feature representation at the k-th layer; It is a node The set of adjacent nodes; and These are nodes and nodes The degree; and These are the weight matrix and the bias matrix, respectively.

[0053] The adaptive basis function strategy dynamically updates node representations by adjusting the basis functions of each node. First, it computes the basis functions of each pair of nodes. and Similarity between The formula is as follows: (13) Then, the ReLU activation function is used to map the calculated similarity to an appropriate weight value. The formula is (14) For the input data, a weighted sum is performed based on the calculated weights to achieve effective feature aggregation, and the output feature of the current node is... It can be represented as (15) The key to the proximity dynamic dependency capture strategy lies in dynamically adjusting the information propagation weights between nodes to optimize the feature propagation process. By calculating the similarity between nodes, the model can capture the nonlinear relationships and potential dependencies between different types of nodes.

[0054] S4. Based on the final node embedding vector output by the multi-layer graph convolutional network, calculate the connection probability between node pairs to complete the identification of the complex dynamic topology of the high-proportion new energy distribution network under test.

[0055] In this embodiment, the improved multi-layer graph convolutional network outputs final node embedding vectors to capture possible connections between nodes. Specifically, by performing a dot product on the embedding vectors between node pairs, the connection probability between each pair of nodes is calculated. This probability is then mapped using the sigmoid function. The higher the output probability value, the greater the likelihood that an edge exists between the two nodes, thus completing the identification of the complex dynamic topology of the high-proportion renewable energy distribution network under test.

[0056] Specifically, in order to verify the improvement effects of the node feature enhancement strategy based on temporal convolutional networks, the long-range dependency learning strategy based on residual mechanisms, and the short-range dynamic dependency capture strategy based on adaptive basis functions, this embodiment uses the GCN model as the benchmark model to verify the effectiveness of each improvement strategy, and then uses ablation experiments to test the compatibility between the strategies.

[0057] (1) Node feature enhancement strategy based on temporal convolutional network The evaluation metrics of the models before and after introducing the node feature enhancement strategy based on temporal convolutional networks are compared, and the results are shown in Table 1.

[0058] Table 1 Experimental results of node feature enhancement strategies

[0059] As shown in Table 1, with the introduction of temporal convolutional networks, the improved method's precision, recall, and F1 score improved by 6.61%, 8.68%, and 7.67%, respectively, resulting in an overall improved recognition performance. This indicates that after the introduction of this strategy, temporal convolutional networks can effectively capture the temporal dependencies between nodes by modeling node features, and generate more dynamic and accurate feature representations for each node. By enhancing the temporal nature of node features, the model can better understand and predict the relationships between node pairs.

[0060] (2) Long-distance dependency learning strategy based on residual mechanism To verify the importance of high-order neighbor node information, the number of graph convolutional layers in GCN was adjusted and improved, affecting the model's feature learning ability. Experimental results under different numbers of graph convolutional layers were compared using evaluation metrics. Specific graph convolutional layer settings and link prediction results are shown in Table 2.

[0061] Table 2. Experimental evaluation metrics for different numbers of graph convolutional layers.

[0062] In the link prediction task, experiments verified the effects of different numbers of graph convolutional layers. A layer count of 2 resulted in the best overall performance, demonstrating its ability to effectively capture complex relationships between nodes and improve prediction capabilities. In contrast, a layer count of 1 performed weakly, mainly due to information propagation being limited to directly adjacent nodes. Performance decreased with more than 2 layers, primarily due to excessive smoothing caused by too much neighbor node information, resulting in a loss of sufficient node discriminative power. The introduction of a long-distance dependency learning strategy can utilize global information about the topology to improve the ability to identify complex topologies.

[0063] (3) Near-range dynamic dependency capture strategy based on adaptive basis function The results of comparing the evaluation indices for identifying complex topologies of high-proportion renewable energy distribution networks before and after the introduction of the near-field dynamic dependency capture (CDDC) strategy based on adaptive basis functions are shown in Table 3.

[0064] Table 3 Results of the Near-Range Dynamic Dependency Capture Strategy

[0065] As shown in Table 3, the topology identification capability of the GCN model is improved after introducing the near-range dynamic dependency capture strategy. The evaluation metrics Precision, Recall, and F1 Score increased by 2.21%, 8.11%, and 5.15%, respectively. This indicates that after the introduction of this strategy, the feature aggregation and update based on node correlation can be completed through adaptive basis functions, thereby fully mining the topology information of the distribution network and improving the link prediction performance of the model.

[0066] The different strategies yield varying degrees of improvement because they focus on different data features. Among them, the node feature enhancement strategy based on temporal convolutional networks achieves the most balanced performance improvement by capturing the temporal information of node data.

[0067] (4) Improved strategy ablation experiment Based on the GCN model, a node feature enhancement strategy, a long-distance dependency learning strategy, and a short-distance dynamic dependency capture strategy were introduced to obtain the experimental results of complex topology identification and ablation of high-proportion new energy distribution networks, as shown in Table 4.

[0068] Table 4 Ablation Experiment Results of Complex Topology Identification Model

[0069] Table 4 yields the following conclusions: 1) A comparison of Experiments 1 to 4 shows that, compared to the basic topology identification method, all three improved strategies introduced can improve Precision, Recall, and F1 Score. Among them, the node feature enhancement strategy has the most significant effect, as it can more effectively model the temporal dependency of node features; 2) A comparison of Experiments 1 to 7 shows that the improved methods based on any two strategies are better than the single-strategy method, indicating that the improved strategies can be well compatible and complementary. The combination of the node feature enhancement strategy and the near-field dynamic dependency capture strategy shows the best dual-strategy fusion effect. These two strategies can effectively extract and integrate the temporal information and topological features of node data simultaneously; 3) By comprehensively comparing all experiments, the method combining the three strategies achieved 0.8874, 0.9189 and 0.9029 in Precision, Recall and F1 Score, respectively. Compared with the best-performing dual-strategy fusion experiment (Experiment 7), the indicators were improved by 6.90%, 1.22% and 4.12%, respectively. This result is because the improved strategies have high compatibility and can fully explore the multi-dimensional features of distribution network data when working together, effectively improving the accuracy of topology identification.

[0070] For the task of complex topology identification in high-proportion renewable energy distribution networks, this embodiment compares the topology identification performance of the proposed improved method with the current mainstream GNN, GAT and GIN methods. The comparison results of the evaluation indicators are shown in Table 5.

[0071] Table 5 Comparative Experiments

[0072] As shown in Table 5, the method in this embodiment exhibits the best overall performance in the complex topology identification task of high-proportion renewable energy distribution networks, with Precision, Recall, and F1 Score of 0.8874, 0.9189, and 0.9029, respectively, outperforming other models. Compared with GIN, the improved method improves Precision, Recall, and F1 Score by 3.21%, 18.67%, and 10.81%, respectively. In high-proportion renewable energy distribution networks, power flow direction and voltage levels change due to renewable energy access, and the mean aggregation of GIN cannot fully represent the complex electrical connections between different nodes. The method in this embodiment can flexibly handle the importance of different node characteristics, enhancing the feature extraction capability of the topology structure. Compared with GAT, the method in this embodiment improves Precision, Recall, and F1 Score by 5.88%, 28.95%, and 17.21%, respectively. In new energy distribution network scenarios, the attention mechanism of GAT is easily affected by local perturbations. This embodiment enhances the perception of global topological features through a long-distance dependency learning strategy, utilizing global topological information to more accurately capture node connectivity relationships. GNN's performance is relatively poor, mainly because it struggles to effectively capture complex topological dependencies in node data. This demonstrates that the method in this embodiment has higher identification capabilities in topology identification tasks, can predict possible node connectivity relationships, and more accurately identify complex distribution network topologies. The topology identification performance of this embodiment compared to the mainstream GNN, GAT, and GIN methods is visualized as follows: Figure 8 As shown.

[0073] To verify the application effect of the method in a high-proportion renewable energy distribution network, this embodiment conducts a complex topology identification experiment based on an improved IEEE 33-node distribution network simulation model. The inputs are normalized distribution network node feature data and an adjacency matrix. The initial adjacency matrix represents the node connection relationships in the original distribution network topology. The output is the predicted topology structure composed of the node connection relationships. The identification process of the method in this embodiment is as follows: Figure 9 As shown.

[0074] The method in this embodiment can effectively utilize the spatiotemporal characteristics of node data, flexibly handle the relationships between different nodes, and improve the ability to perceive complex topologies. Through node feature enhancement strategies, long-distance dependency learning strategies, and short-distance dynamic dependency capture strategies, this method accurately identifies potential connections between nodes and effectively addresses topology changes caused by a high proportion of new energy access.

[0075] In summary, this embodiment provides a method for identifying complex dynamic topology of high-proportion renewable energy distribution networks based on an improved GCN. This method can solve the problems of insufficient adaptability, low computational efficiency, and poor scalability in the identification of complex dynamic topology of high-proportion renewable energy distribution networks.

[0076] Example 2 This application provides a complex dynamic topology identification system for high-proportion renewable energy distribution networks based on an improved GCN, including: The data acquisition module is used to acquire time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested; The node feature enhancement module is used to process temporal node feature data using a temporal convolutional network, enhance the temporal dependency expression of node features, and generate a dynamic node feature matrix. The processing module is used to input the dynamic node feature matrix and adjacency matrix into a pre-trained multi-layer graph convolutional network containing a residual connection mechanism. It aggregates multi-hop neighbor information through stacked graph convolutional layers to learn the long-distance dependencies between nodes. In each layer of the multi-layer graph convolutional network, it dynamically calculates the similarity weights between nodes using a strategy based on adaptive basis functions to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes. The topology identification output module is used to calculate the connection probability between node pairs based on the final node embedding vector output by the multi-layer graph convolutional network, thereby completing the identification of the complex dynamic topology of the high-proportion new energy distribution network under test.

[0077] It should be noted that, depending on the implementation needs, the various steps described in this application can be broken down into more steps, or two or more steps or parts of the steps can be combined into new steps to achieve the purpose of this invention.

[0078] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying complex dynamic topology in high-proportion renewable energy distribution networks based on an improved GCN, characterized in that, include: Obtain the time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested; The temporal node feature data is processed using a temporal convolutional network to enhance the temporal dependency expression of node features and generate a dynamic node feature matrix. The dynamic node feature matrix and the adjacency matrix are input into a trained multi-layer graph convolutional network containing a residual connection mechanism. Multi-hop neighbor information is aggregated through stacked graph convolutional layers to learn the long-distance dependencies between nodes. In each layer of the multi-layer graph convolutional network, a strategy based on adaptive basis functions is used to dynamically calculate the similarity weights between nodes to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes. Based on the final node embedding vector output by the multi-layer graph convolutional network, the connection probability between node pairs is calculated to identify the complex dynamic topology of the high-proportion new energy distribution network under test.

2. The method for identifying complex dynamic topology of a high-proportion renewable energy distribution network based on an improved GCN, as described in claim 1, further includes: The multi-layer graph convolutional network with residual connection mechanism was trained by acquiring historical data of time-series node characteristics and adjacency matrix of high-proportion renewable energy distribution networks.

3. The method for identifying complex dynamic topology of a high-proportion renewable energy distribution network based on an improved GCN according to claim 1, wherein the multilayer graph convolutional network including a residual connection mechanism comprises: The output of each layer is added to the initial feature matrix through residual connections to prevent gradient vanishing and network degradation.

4. The method for identifying complex dynamic topology of a high-proportion new energy distribution network based on an improved GCN according to claim 3, wherein the multi-layer graph convolutional network including the residual connection mechanism preferably has 2 layers.

5. The method for identifying complex dynamic topology of high-proportion new energy distribution networks based on improved GCN according to claim 1, wherein the strategy based on adaptive basis function is specifically as follows: by calculating the cosine similarity between node embedding vectors and mapping it through activation function, dynamic attention weights are obtained, which are used to weighted aggregate neighbor node features.

6. The method for identifying complex dynamic topology of high-proportion renewable energy distribution networks based on improved GCN according to claim 1, wherein the processing of the time-series node feature data using a temporal convolutional network specifically includes: A temporal convolutional network with causal convolution and dilated convolution structures is used to capture long-term and short-term dependency features of node states with different dilation rates.

7. The method for identifying complex dynamic topology of high-proportion renewable energy distribution networks based on improved GCN as described in claim 1, characterized in that, The obtained adjacency matrix needs to be symmetrically normalized to avoid the impact of node degree differences on feature propagation.

8. The method for identifying complex dynamic topology of a high-proportion renewable energy distribution network based on an improved GCN as described in claim 1, wherein the activation function is ReLU.

9. The method for identifying complex dynamic topology of a high-proportion new energy distribution network based on an improved GCN as described in claim 1, wherein the connection probability between node pairs is obtained by performing a dot product operation on the node embedding vectors and mapping them through a probability mapping function; The probability mapping function is the Sigmoid function.

10. A complex dynamic topology identification system for high-proportion renewable energy distribution networks based on an improved GCN, characterized in that, include: The data acquisition module is used to acquire time-series node characteristic data and adjacency matrix of the high-proportion renewable energy distribution network to be tested; The node feature enhancement module is used to process the temporal node feature data using a temporal convolutional network, enhance the temporal dependency expression of node features, and generate a dynamic node feature matrix. The processing module is used to input the dynamic node feature matrix and the adjacency matrix into a trained multi-layer graph convolutional network containing a residual connection mechanism, and to aggregate multi-hop neighbor information through stacked graph convolutional layers to learn the long-distance dependencies between nodes; in each layer of the multi-layer graph convolutional network, a strategy based on adaptive basis functions is used to dynamically calculate the similarity weights between nodes to adjust the intensity of neighbor node information propagation and capture the short-distance dynamic dependencies between nodes; The topology identification output module is used to calculate the connection probability between node pairs based on the final node embedding vector output by the multi-layer graph convolutional network, thereby completing the identification of the complex dynamic topology of the high-proportion new energy distribution network under test.

Images