Environmental element information embedded heterogeneous network key node mining method

Abstract

The invention discloses an environment element information embedded heterogeneous network key node mining method, which can accurately model the cross-layer relationship of a multi-layer power network according to cross-layer rotation embedding, multi-mode environment element embedding and reinforcement learning optimization strategies, improve the accuracy of an edge connection prediction task, and improve the accuracy of the edge connection prediction task. And meanwhile, high-risk key nodes can be identified in combination with reinforcement learning, a scientific basis is provided for disaster response, risk assessment and safety scheduling of the smart power grid, and the safety, toughness and intelligent level of a power grid system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power networks, and in particular, to a method for mining key nodes in a heterogeneous network with environmental factor information embedded therein. Background Art

[0002] Current research on power energy networks mostly focuses on the topological structure analysis of single-layer networks, while ignoring the complex interaction characteristics of multi-layer network systems. In multi-layer power networks (such as power generation networks, power transmission networks, and power distribution networks), the connection relationships between layers are not only affected by the internal topological structure of the network, but also by dynamic interferences from external environmental factors (such as typhoons, fires, earthquakes, etc.). However, existing edge prediction methods mainly rely on static modeling, only considering the fixed structural characteristics of single-layer networks, making it difficult to effectively model cross-layer relationships in multi-layer networks, and having low prediction accuracy under natural disaster conditions, thus being difficult to meet the security requirements of complex power systems. Summary of the Invention

[0003] To solve the limitations and defects of the existing technology, the present invention provides a method for mining key nodes in a heterogeneous network with environmental factor information embedded therein, including:

[0004] Obtaining environmental factor information and power network status information;

[0005] Embedding the environmental factor information and the power network status information as node attributes into a multi-layer network model according to a dynamic embedding mechanism, where the multi-layer network model adaptively adjusts the embedding representation according to a dynamic neural network and the environmental factor information;

[0006] Mapping cross-layer edges to rotation operations in the complex space to capture various patterns of cross-layer relationships and enhance the node association modeling ability between different hierarchical networks, where the various patterns include symmetry, anti-symmetry, reversibility, and combinability;

[0007] Evaluating the cross-layer relationship modeling ability using an edge prediction task, constructing a key node mining strategy according to a reinforcement learning framework, and identifying key nodes whose importance value for power network stability under natural disasters is greater than or equal to a preset value.

[0008] Optionally, the step of constructing a key node mining strategy according to a reinforcement learning framework and identifying key nodes whose importance value for power network stability under natural disasters is greater than or equal to a preset value includes:

[0009] Constructing a node importance evaluation mechanism based on deep reinforcement learning;

[0010] Evaluating the importance value of nodes for power network stability under natural disasters according to the node importance evaluation mechanism;

[0011] Nodes with the important value greater than or equal to the preset value are identified as key nodes.

[0012] Optionally, it further includes:

[0013] Use a graph convolutional neural network to extract features from the topological structure of the single-layer network and learn the node connection patterns within each layer of the network;

[0014] According to the environmental state information and the node connection patterns, enhance the dynamic perception ability of the multi-layer network model for the intra-layer structure, and provide feature representations for subsequent key node mining.

[0015] Optionally, the environmental factor information includes wind speed, temperature, humidity, precipitation, air pressure, earthquake magnitude, tidal amplitude, and thermal radiation intensity, and the power grid state information includes voltage and power.

[0016] Optionally, the step of adaptively adjusting the embedding representation includes:

[0017] Introduce a dynamic embedding mechanism for environmental perception, and adaptively adjust the node embedding representation according to changes in the external environment. The expression of the adaptive adjustment is as follows:

[0018] z′ i =z i +β·g(R i )

[0019] Where z i is the initial embedding of node i; g(R i ) is a non-linear mapping function of R i for adjusting the embedding vector; R i is the natural disaster risk score of node i; β is the adjustment coefficient.

[0020] Optionally, the step of mapping the cross-layer edges to rotation operations in the complex space includes:

[0021] Use rotation operations in the complex domain to model the embeddings of entities and relationships. The expression of the rotation operation is: Where h u is the embedding of node u, r is the embedding vector of the relationship, used to model the cross-layer relationship between cross-layer nodes, represents the element-wise product in the complex domain.

[0022] The present invention has the following beneficial effects:

[0023] The present invention provides a method for mining key nodes in a heterogeneous network with environmental factor information embedding. According to cross-layer rotation embedding, multi-modal environmental factor embedding, and a reinforcement learning optimization strategy, the present invention can not only accurately model the cross-layer relationships of a multi-layer power network, improve the accuracy of the link prediction task, but also identify high-risk key nodes in combination with reinforcement learning, providing a scientific basis for disaster response, risk assessment, and safety dispatching of smart grids, and improving the safety, resilience, and intelligence level of the power grid system. Description of the Drawings

[0024] Figure 1 It is an architecture diagram of the method for mining key nodes in a heterogeneous network with environmental factor information embedding provided in the first embodiment of the present invention.

[0025] Figure 2 It is a schematic diagram of the framework structure of the network without stratification and its abstract stratification structure provided in the first embodiment of the present invention.

[0026] Figure 3 It is an architecture diagram of the attribute environmental characteristics provided in the first embodiment of the present invention.

[0027] Figure 4 It is a schematic diagram of in-layer relationship modeling provided in the first embodiment of the present invention.

[0028] Figure 5 It is a schematic diagram of cross-layer relationship modeling provided in the first embodiment of the present invention.

[0029] Figure 6 It is a schematic diagram of the method for mining key nodes provided in the first embodiment of the present invention. Detailed Embodiments

[0030] To enable those skilled in the art to better understand the technical solutions of the present invention, the method for mining key nodes in a heterogeneous network with environmental factor information embedding provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0031] Embodiment 1

[0032] This embodiment provides a method for mining key nodes in a multi-modal multi-layer power grid based on cross-layer rotation embedding, including: constructing a multi-layer power grid model to analyze the interaction characteristics between the power generation, transmission, and distribution networks; mapping cross-layer connections to rotation operations in the complex space to capture various patterns of cross-layer relationships and enhance the ability to model node associations; introducing a dynamic embedding mechanism for environmental factors to embed real-time environmental states such as wind speed, humidity, and temperature as node attributes into the network, and implementing model adaptive adjustment through a dynamic neural network. To verify the effectiveness of the model, first, the cross-layer relationship modeling ability is evaluated using the link prediction task, and then, combined with the reinforcement learning method, the key node mining strategy is optimized to identify the key nodes that affect the stability of the power grid under severe weather or sudden disaster conditions. This method provides efficient decision-making support for disaster response and fault repair in smart grids, improving the safety and risk management capabilities of power grids.

[0033] This embodiment proposes a multi-modal key node mining method based on cross-layer rotation embedding. Based on the modeling of a multi-layer power grid, link prediction is combined as a validation of effectiveness, and further, reinforcement learning is introduced to optimize the key node mining strategy. Specifically, in the relationship modeling of the multi-layer network, cross-layer connections are mapped to rotation operations in the complex space, thereby capturing various patterns such as symmetry, anti-symmetry, reversibility, and combinability of cross-layer relationships and enhancing the ability to model node associations between different hierarchical networks.

[0034] To adapt to a complex dynamic environment, this embodiment further introduces a dynamic embedding mechanism for environmental factors, embeds real-time environmental states such as wind speed, humidity, and temperature as node attributes into the network model, and combines a dynamic neural network to enable the model to adaptively adjust the embedding representation and inference strategy according to changes in the external environment. For example, when a typhoon approaches, the model can identify the potential impact of wind speed on transmission lines and dynamically optimize the prediction strategy; in the event of an emergency such as a fire or earthquake, the model can focus on the connectivity of key nodes in the affected area, thereby enhancing the environmental adaptability of the model.

[0035] To further optimize the resilience analysis and disaster response capabilities of smart grids, this embodiment adopts a reinforcement learning framework to optimize the key node mining strategy based on the validation results of the link prediction task. By constructing a node importance evaluation mechanism based on deep reinforcement learning, the model can identify the nodes crucial for grid stability under natural disasters and provide targeted control suggestions. For example, the model can identify the key nodes that have the most significant impact on power supply under typhoon conditions or during an earthquake, thereby providing efficient decision-making support for the proactive defense and post-disaster recovery of the power grid.

[0036] In this embodiment, a method for mining key nodes of a power network in a complex and changeable environment is proposed by combining cross-layer rotation embedding, multi-modal environmental factor embedding, and reinforcement learning optimization strategy. This method can not only accurately model the cross-layer relationships of a multi-layer power network and improve the accuracy of the edge prediction task, but also identify high-risk key nodes by combining reinforcement learning, providing a scientific basis for disaster response, risk assessment, and safety dispatching of smart grids, and improving the safety, resilience, and intelligence level of the power grid system.

[0037] Based on dynamic embedding technology, this embodiment introduces environmental factor information such as typhoons and fires and power network state information such as voltage and power into a multi-modal graph convolutional neural network embedding framework. On the basis of realizing accurate and robust cross-layer and intra-layer edge prediction of the power network, it further combines reinforcement learning to optimize the key node mining strategy. This method includes three main parts:

[0038] Attribute environmental feature processing module: Dynamically embed environmental factors (such as wind speed, humidity, temperature) and power network state information (such as voltage, power), and input them as node attributes into a multi-layer network model to ensure that the model can adapt to complex dynamic environments, and optimize the embedding representation through an adaptive mechanism to improve the robustness of the prediction task.

[0039] Intra-layer relationship modeling module: Use a graph convolutional neural network (GCNN) to extract features from the topological structure of a single-layer network, learn the node connection patterns within each layer of the network, and combine environmental state information to enhance the model's dynamic perception ability of the intra-layer structure, providing high-quality feature representations for subsequent key node mining.

[0040] Cross-layer relationship modeling and key node mining module: Model the cross-layer edge relationships in a multi-layer power network, map the cross-layer interactions to rotation operations in the complex number space to capture various patterns of cross-layer relationships, and combine reinforcement learning to optimize the key node mining strategy. On the basis of validating the effectiveness of the edge prediction task, use a reinforcement learning framework to construct a key node selection strategy, identify the key nodes that have the greatest impact on the stability of the power grid under natural disasters, and generate an optimized control scheme to improve the risk management ability and disaster response efficiency of the power grid.

[0041] Figure 1 This is the architecture diagram of the heterogeneous network key node mining method for embedding environmental factor information provided in the first embodiment of the present invention. By integrating cross-layer rotation embedding, multi-modal environmental feature embedding, and reinforcement learning optimization strategy, this embodiment identifies the key nodes that affect the stability of the power grid under natural disasters, thereby improving the disaster resistance ability and fault recovery efficiency of the power grid.

[0042] In terms of data: A multi-layer heterogeneous energy network in a real scenario is constructed, and at the same time, the collection and preprocessing of its node feature data are involved.

[0043] Figure 2 It is a schematic diagram of the unstratified framework structure and its abstract stratified structure provided for the first embodiment of the present invention.

[0044] I. Flow architecture diagram of the attribute environment feature processing module

[0045] Figure 3 It is the attribute environment feature architecture diagram provided for the first embodiment of the present invention. In the process of modeling a multi-layer power network, the data collection, preprocessing, and fusion of physical attributes and environmental attributes are key links in model construction, and their quality directly affects the generalization ability and computational efficiency of the model. To enhance the effectiveness of multi-modal feature representation and improve the performance of edge prediction and key node mining in combination with dynamic embedding technology, this embodiment proposes a systematic data processing process. This process covers the standardized conversion of physical attributes, the structured expression of spatial information, and the dynamic modeling of environmental attributes to ensure that the model can achieve high-precision inference in a complex power network environment. The following standardized data processing process is proposed in this embodiment:

[0046] 1. Physical attribute processing

[0047] 1.1 Discrete attribute one-hot encoding

[0048] Discrete attributes in the power network (such as device type, voltage level, etc.) have a limited but discontinuous value range. Traditional numerical assignment methods may introduce meaningless numerical relationships. Therefore, one-hot encoding is used to perform numerical conversion on them. Specifically, assuming that there are n categories of device types, for a device belonging to category

[0049] i, its one-hot encoding is expressed as:

[0050] OneHot(i) = [0,…,1,…,0]

[0051] Among them, only the i-th dimensional component takes the value of 1, and the remaining dimensions are all 0. This encoding method avoids unreasonable ordinal relationships between categorical variables and at the same time enhances the model's ability to distinguish different categories, enabling it to effectively learn the role differences of power devices in the multi-layer network structure.

[0052] 1.2 Continuous attribute normalization

[0053] For continuous physical attributes such as transformer capacity and population density in the service area, the Min-Max Normalization method is adopted to eliminate the dimensional differences between different attributes and enable the model to have better convergence during the optimization process. The normalization transformation formula is as follows:

[0054]

[0055] Map the numerical features to the interval [0,1] to ensure that different attributes participate in model calculations on the same scale, and at the same time avoid affecting the stability of model weight learning due to uneven numerical ranges of attributes.

[0056] 1.3 Conversion of longitude and latitude to distance

[0057] The spatial distribution characteristics of the energy network are crucial for network edge prediction and key node mining tasks. To avoid the impact of the non-Euclidean characteristics of the original longitude and latitude data on model calculations, the Great-circle Distance calculation method is adopted. Taking the central substation or designated anchor point of the power network as the reference, the geographical location information of each node is converted into the spherical distance from the anchor point. The calculation formula is as follows: The great-circle distance formula is used for calculation:

[0058]

[0059] where r is the radius of the earth (taking about 6371 km); φ and λ are the latitude and longitude respectively (in radians). Subsequently, to ensure that the distance feature has the same numerical scale as other physical attributes, the calculated distance between nodes is normalized:

[0060]

[0061] 2. Environmental attribute processing

[0062] To adapt to the impact of various disasters (such as typhoons, earthquakes, fires, floods, etc.) on the power network, a general disaster warning model is designed to calculate the disaster risk score R of the node based on multi-dimensional environmental attributes i , which is used to dynamically adjust the embedded representation of the model.

[0063] In a complex power network system, natural disasters (such as fires, floods, etc.) have a profound impact on the dynamic changes of the network topology. Therefore, this embodiment proposes an environment-aware dynamic embedding modeling method, which calculates the disaster risk score (Risk Score, R i ) of the node based on multi-dimensional environmental attributes, and dynamically adjusts the embedded representation of the node accordingly, so as to enhance the adaptability of the model to external disturbances.

[0064] 2.1 Dynamic Modeling of Environmental Attributes

[0065] Assume that the environmental attributes include n influencing factors: wind speed (v), humidity (h), temperature (T), seismic intensity (e), etc. The environmental attribute vector of node i is expressed as:

[0066] E i = [v i , h i , T i , e i , p f,i , …]

[0067] where p f,i represents the probability or intensity of node i being affected by specific disasters such as typhoons, fires, etc., and this information can be obtained through historical statistics or real-time monitoring data.

[0068] 2.2 Disaster Risk Scoring Function

[0069] Combining environmental characteristics and node characteristics, define the disaster risk score of the node:

[0070]

[0071] where: E i,j is the eigenvalue of the environmental attribute j of node i; f j (E i,j ) is the normalization or characteristic function of attribute j; α j is the weight parameter, indicating the contribution of different attributes to the risk score. Examples of common characteristic functions: Wind speed: Humidity: Seismic intensity:

[0072] 2.3 Dynamic Embedding Adjustment

[0073] Combined with the disaster risk score R i , further introduce a dynamic embedding mechanism for environmental perception, so that the node embedding representation can be adaptively adjusted according to changes in the external environment. The adjustment formula is as follows:

[0074] z′ i = z i + β · g(R i )

[0075] where z i is the initial embedding of the node; g(R i ) is the non-linear mapping function of the risk score R i for adjusting the embedding vector; β is the adjustment coefficient. This mechanism ensures that the model can update the node representation in real time under natural disasters and improve its prediction accuracy under disaster conditions.

[0076] II. Intra-layer relationship modeling

[0077] Figure 4 This is a schematic diagram of the intra-layer relationship modeling provided in the first embodiment of the present invention. A graph neural network (GNN) can learn node representations by aggregating information from neighbors, thereby capturing potential connection patterns, which is beneficial for link prediction tasks. Multi-layer networks provide valuable information for enhancing link prediction. In view of this, our method involves using GNN in multi-layer networks, which simultaneously learns cross-network embeddings and integrates intra-layer and inter-layer structural features.

[0078] Various GNNs for node representation have been developed in general GNN layers. Generally speaking, a typical GNN layer is as follows:

[0079]

[0080]

[0081] where The output vector (representation) of node u at the k-th layer; The representation of node u in the previous layer (k - 1); Weight matrices that respectively adjust the contributions of the node's own representation and the neighbor node representation to the update; The neighbors of node u; a uv : The importance weight (attention) of the edge connecting u and v; θ: Activation function, used to introduce non-linear transformation (such as ReLU), indicating that the node representation at the k-th layer is calculated by combining the weighted sum of its own representation and the neighbor node representation.

[0082] III. Cross-layer relationship modeling

[0083] Figure 5 This is a schematic diagram of the cross-layer relationship modeling provided in the first embodiment of the present invention. Aggregating messages from other layers will help supplement node information, thereby enhancing the understanding of connection patterns. To enhance the representation learning of nodes in cross-layer networks, we calculate cross-layer anchor embeddings based on a rotation-based modeling method, while supporting the aggregation of multiple anchors. By utilizing the rotation embedding mechanism, we can effectively capture the cross-layer relationships between nodes and combine the weight mechanism to perform weighted fusion on the information of multiple anchors. On the basis of maintaining the original intra-layer neighbor structure information, the integration of cross-layer information is further optimized, making the embedded representation more accurate and semantic.

[0084] RotatE is a knowledge graph embedding method that uses rotation operations in the complex number domain to model the embeddings of entities and relationships, and its formula is: where h uis the embedding of node u, and r is the embedding vector of the relationship, modeling the cross-layer relationship between cross-layer nodes. denotes the element-wise product over the complex number field. Among them, the embedding of nodes is usually represented as complex numbers, and the rotation relationship is modeled by complex numbers with modulus 1 (ensuring that rotation does not change the modulus length).

[0085] There are multiple anchor points corresponding to other layers. Assume that each node u corresponds to an anchor point set Each anchor point has its embedding and the corresponding relationship r uu′ , and the cross-GNN corresponding to integrating the information between layers is as follows:

[0086]

[0087] which involves: multi-anchor point aggregation, rotation embedding update, and maintaining the original structural information (within the layer). For all anchor points perform a weighted sum, with the weight being w uu′ used to measure the importance of node u and its cross-layer anchor point u'. x u represents the original feature vector of node u, usually the fixed input feature, is the linear transformation matrix used to combine the node feature information. Aggregate through ; Re(·) is to take the real part of the complex number to ensure that the embedding result is in the real number field. The original formula is based on the real number field, while the rotation operation is based on the complex number field. Through the real part operation Re(·), the result in the complex number field can be mapped back to the real number field to be compatible with the original node representation. Use the rotation formula to calculate the embedding of the cross-layer anchor point, where r uu′ is the rotation vector of the relationship between u and u'. The first term and the second term are used to retain the in-layer structure and neighbor information of the node.

[0088] The objective function includes the in-layer loss and the inter-layer loss. Among them, the in-layer loss mainly retains the in-layer structural features, and the inter-layer loss retains the inter-layer structural features. Then jointly optimize the total loss to make multiple networks better unified. If two nodes often appear simultaneously in a fixed-length random walk, it indicates a higher probability of a link between them. Therefore, random walk captures the high-order proximity and provides insights into the connection patterns and potential links within the network. Therefore, in order to learn the embedding z i , apply the random walk-based objective function in the unsupervised setting:

[0089]

[0090] where j is the node co - occurring with i in the random walk sequence window, σ is the sigmoid function, P n (v) is the negative sampling probability distribution, and Q defines the number of negative samples. Proximal nodes are encouraged to have similar embeddings, while discrete nodes are distinct in the embedding space. The intra - layer loss is the sum of the intra - layer losses based on random walks for each layer: where n is the number of network layers. By minimizing the intra - layer loss, the intra - layer structural features of the network can be retained.

[0091] When designing the cross - layer loss function, a rotation embedding mechanism is introduced into cross - layer modeling to more accurately capture the cross - layer relationships between nodes. Specifically, based on the complex rotation operation of the node's rotation embedding and the relationship vector, the updated loss function calculates the cosine similarity of the cross - layer node embeddings to improve the accuracy of cross - layer alignment. In the positive sample loss, the rotated embedding alignment method is combined to ensure the semantic consistency between cross - layer anchor points; while in the negative sample part, by screening and optimizing hard negative samples, the problem of sample imbalance is effectively alleviated. The formula is as follows:

[0092]

[0093] For each node and its set of cross - layer anchor points there is The negative sample S - is selected from the set of neighbors, ensuring that the sampling process is consistent with the distribution of anchor points to improve the representativeness of hard negative samples. denotes that after the embedding of node is rotated by the relationship vector r ij , the real part is taken for calculating the similarity with . For each pair of negative samples the cosine similarity based on the rotated embedding is also calculated. Hard negative samples with similarities exceeding a certain threshold ∈ are screened by max(cos - ∈, 0). This cross - layer loss function design fully combines the characteristics of cross - layer networks and the advantages of rotation embeddings, providing more robust theoretical support and practical guidance for the node alignment task of multi - layer networks.

[0094] The total objective function is defined as a linear combination of intra - layer and inter - layer losses, so we can retain the intra - layer and inter - layer structural information through joint optimization:

[0095] L = α·L intra +(1 - α)·L inter

[0096] Among them, α is the weight parameter that weighs the two components of the objective function, and the Adam optimizer is used to update the parameters of the GNN layer. The node embedding update of each layer depends on the embedding of the previous layer, the embeddings of neighboring nodes, and the information across layers. Specifically, the cross-layer embedding models the relationship between cross-layer nodes through a rotation operation, making the cross-layer nodes more accurate in the embedding space. To perform edge prediction, it is first necessary to calculate the edge embedding for each pair of nodes (x, y). The edge embedding is jointly constructed from the embedding vectors of nodes x and y. Logistic Regression is used to predict the existence probability of edges, which is used to learn the relationship between the edge embedding features and the existence of edges: By learning the relationship between the input edge embedding and the label (the existence of the edge), P(x~y) is predicted, that is, given the embeddings of node pair (x, y), the probability of having an edge between them is calculated.

[0097] IV. Mining of Network Key Nodes Based on Graph Convolutional Neural Network

[0098] Figure 6 This is a schematic diagram of the key node mining method provided in Embodiment 1 of the present invention. By combining a graph convolutional neural network and reinforcement learning, and using an end-to-end optimization algorithm, the key node selection problem is converted into a decision-making process of an intelligent agent. First, the network is encoded into a low-dimensional space using the graph convolutional neural network algorithm. The size of the largest connected component of the network is defined as the reinforcement learning reward. Then, at the decoding end, a reinforcement learning DQN architecture is adopted, and the node state vector is cross-multiplied with a predefined action vector, and a multi-layer perceptron is used to generate a Q-value function. The Q-value represents the importance of the possible edges between node pairs to the network robustness. Considering the interaction between the intelligent agent and the environment, taking actions and obtaining feedback rewards, the graph neural network is used to update the node vector representation as the next state. Finally, based on the GMM multi-layer network generation model, a large number of small-scale multi-layer networks are generated to train the reinforcement learning model to finally accurately find the key nodes.

[0099] From the impact of virtual simulation natural disasters on network nodes, study the early warning signals of network collapse and quantify the change of the node's impact on network resilience. First, use the percolation model of the network to define the network nodes ΔN that need to be disintegrated for network collapse, and use the importance of node i predicted by the reinforcement learning agent, v i represents the degree of damage to the network robustness caused by removing this node, v i is calculated by the reinforcement learning agent trained in the previous stage, represents the network collapse critical value, S0 represents the set of nodes actually removed from the multi-layer network. When a set of nodes S is removed, the estimated value Ω of the actual damage degree of the network s = ∑ n∈s pn Finally, the network warning value Ω is expressed as the following formula, 0≤Ω m ≤1 is used to quantify the amount of damage the network can tolerate before collapsing.

[0100]

[0101] In terms of network topology optimization, three-dimensional tensors are used to store constraint relationships based on the types of constraints and the types of interactions with the environment. For example, the geographical accessibility between different nodes is represented by a binary vector, and the construction cost of the network line is represented by a normalized value. Subsequently, the vector representation of the relationship between nodes is obtained by bitwise multiplication of the node representation vectors. The change in the connected group after removing the node from the network and disconnecting the corresponding edge is used as the reward function of reinforcement learning. The reinforcement learning DQN architecture consistent with the above key nodes is adopted, and a large number of small-scale multi-layer networks instantiated by the GMM multi-layer network generation model are used as training data sets. Using the model and value evaluation function that represent the reinforcement learning model well, the edge importance score Q between any node pair is calculated. By adjusting the weights of the hyperparameters, the model is combined with the actual situation of the network system, the weighted resilience contribution value of the node pair is calculated, and the node pairs are sorted from high to low based on this value. The node with the largest contribution value to the weighted resilience of the network is the key node mined in the network.

[0102] This embodiment proposes a method for multi-modal multi-layer power network edge prediction and key node mining based on cross-layer rotation embedding. This method aims to address the shortcomings of traditional network edge prediction and node mining methods, which only focus on the single-layer network topology structure and ignore the interactive characteristics of multi-layer networks and the influence of external environmental factors. It constructs a multi-layer power network model and deeply analyzes the interactive characteristics between power generation, transmission and distribution networks. The cross-layer edges are mapped to rotation operations in complex space, which effectively captures various modes of cross-layer relationships and significantly enhances the node association modeling capabilities. At the same time, a dynamic embedding mechanism for environmental factors is introduced, and real-time environmental states such as wind speed, humidity, and temperature are embedded in the network as node attributes. The dynamic neural network is used to realize adaptive adjustment of the model, so that the model can comprehensively consider attribute factors to realize key node mining under severe weather or sudden disaster conditions.

[0103] At the technical implementation level, this method consists of an attribute environment feature processing module, an intra-layer relationship modeling module, and a cross-layer relationship modeling and key node mining module. The attribute environment feature processing module preprocesses the physical attributes (such as voltage level, transformer capacity) and environmental attributes (such as wind speed, humidity, temperature) of the power grid using methods such as one-hot encoding and normalization, and constructs a disaster warning model based on multi-dimensional environmental factors to calculate the disaster risk score of nodes, so as to dynamically adjust the node embedding representation. The intra-layer relationship modeling module uses a graph convolutional neural network to learn node representations by aggregating neighbor information, thereby capturing intra-layer topological features and potential connection patterns, and improving the accuracy of local edge prediction. The cross-layer relationship modeling and key node mining module calculates cross-layer anchor embeddings based on the rotation embedding method, models cross-layer edge relationships using rotation transformation, and optimizes multi-anchor information fusion through a weight mechanism to improve the cross-layer information integration ability. In addition, an intra-layer - inter-layer joint optimization objective function is constructed, and through random walk, cosine similarity, and the rotation embedding strategy, the structural consistency and predictability of node representations are optimized. Finally, a reinforcement learning framework is introduced, and based on the validity verification of edge prediction, the key node mining strategy is optimized to identify the key nodes affecting the stability of the power grid under natural disasters, thereby improving the disaster resistance ability and fault recovery efficiency of the power grid.

[0104] It can be understood that the above embodiments are merely exemplary embodiments adopted to illustrate the principles of the present invention, and the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also regarded as the protection scope of the present invention.

Claims

1. A method for mining key nodes in a heterogeneous network with environmental factor information embedded, characterized in that Including: Obtaining environmental factor information and power network status information; Embedding the environmental factor information and the power network status information as node attributes into a multi-layer network model according to a dynamic embedding mechanism, where the multi-layer network model adaptively adjusts the embedding representation according to a dynamic neural network and the environmental factor information; Mapping cross-layer edges to rotation operations in the complex space to capture various patterns of cross-layer relationships and enhance the node association modeling ability between different hierarchical networks, where the various patterns include symmetry, antisymmetry, reversibility, and combinability; Evaluating the cross-layer relationship modeling ability using an edge prediction task, constructing a key node mining strategy according to a reinforcement learning framework, and identifying key nodes whose importance values for the stability of the power network under natural disasters are greater than or equal to a preset value.

2. The method for mining key nodes of a heterogeneous network with environmental factor information embedded according to claim 1, wherein The step of constructing a key node mining strategy according to a reinforcement learning framework and identifying key nodes whose importance values for the stability of the power network under natural disasters are greater than or equal to a preset value includes: Constructing a node importance evaluation mechanism based on deep reinforcement learning; Evaluating the importance value of a node for the stability of the power network under natural disasters according to the node importance evaluation mechanism; Identifying nodes with importance values greater than or equal to the preset value as key nodes.

3. The method for mining key nodes of a heterogeneous network with environmental element information embedded according to claim 2, wherein It also includes: Using a graph convolutional neural network to extract features from the topological structure of a single-layer network and learn the node connection patterns within each layer of the network; Enhancing the dynamic perception ability of the multi-layer network model for the intra-layer structure according to the environmental state information and the node connection patterns, and providing a feature representation for subsequent key node mining.

4. The method for mining key nodes of a heterogeneous network with environmental element information embedded according to claim 3, characterized in that The environmental factor information includes wind speed, temperature, humidity, precipitation, air pressure, earthquake magnitude, tidal amplitude, and thermal radiation intensity, and the power network status information includes voltage and power.

5. The method for mining key nodes of a heterogeneous network with environmental element information embedded according to claim 4, wherein The step of adaptively adjusting the embedding representation includes: Introducing an environment-aware dynamic embedding mechanism to adaptively adjust the node embedding representation according to changes in the external environment, and the expression of the adaptive adjustment is as follows: z′ i = z i + β·g(R i ) Among them, z i is the initial embedding of node i; g(R i ) is a non-linear mapping function of R i for adjusting the embedding vector; R i is the natural disaster risk score of node i; β is an adjustment coefficient.

6. The method for mining key nodes of a heterogeneous network with environmental factor information embedded according to claim 5, wherein The step of mapping cross-layer edges to rotation operations in the complex space includes: Modeling the embeddings of entities and relationships using rotation operations in the complex domain, the expression of the rotation operation is: where h u is the embedding of node u, r is the embedding vector of the relationship, used to model the cross-layer relationship between cross-layer nodes, and ° represents the element-wise product in the complex domain.

Images