Electric power CPS false data injection attack detection method based on graph neural network
The grid graph topology model is constructed through graph neural networks, combined with GAT and HAN, and the problem of traditional methods being difficult to detect complex grid topology changes is solved, and efficient detection of false data injection attacks is achieved, which improves the security of the power system and data redundancy. sex.
Patent Information
- Application Number
- CN202411088381.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional data-driven methods are difficult to adapt to complex and dynamically changing grid topology, and cannot effectively detect false data injection attacks, resulting in insufficient security of the power system.
By constructing a graph topology model of the power grid, using graph attention network (GAT) and heterogeneous graph neural network (HAN) combined with multi-task learning and fuzzy inference systems, the relationship between grid nodes and lines is captured, and the relationship between grid nodes and lines is carried out. Detection and analysis of fake data injection attacks.
Improve the accuracy and efficiency of detecting abnormal states of the power system, enhance the safety protection capabilities of the power grid, and generate high-precision additional data without adding actual measurement equipment, enhance data redundancy, and improve potential Attack sensitivity.
Smart Images

Figure CN119071031A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of power protection systems, and in particular to a method for detecting false data injection attacks on power CPS based on graph neural networks. Background Art
[0002] In recent years, traditional power systems have produced problems such as low energy utilization and low efficiency. In order to deal with these problems, the State Grid Corporation of China proposed the construction of a power information-physical fusion system in 2020 as the core content of the smart grid strategy. Information-physical systems (CPS) can help achieve efficient allocation, real-time monitoring and scientific decision-making of power resources, and represent the development trend and direction of future power grids. However, while improving the efficiency of power resource allocation, real-time analysis and decision-making capabilities, security vulnerabilities in communication networks and information equipment also pose potential threats to smart grids. Hackers can use a variety of methods to attack smart grids, the most notable of which is to deliberately tamper with the data of smart devices in order to achieve the purpose of destroying the power system. This method is called false data injection attack (FDIA). FDIA destroys the integrity of power grid information by tampering with measurement and control data, and is highly permeable, concealed and disruptive. This attack method can interfere with the analysis and decision-making of the upper control center, leading to serious consequences, and is one of the most threatening attack methods to the power system. In addition to FDIA, hackers may also use other technical means, such as physical destruction, electromagnetic interference, and malware implantation, to attack smart grids. However, in comparison, FDIA is particularly dangerous due to its concealment and precision.
[0003] However, with the increasing complexity of power grid topology and the diversification of components, traditional data-driven methods face challenges. These methods usually only focus on the relationship between system measurement information and state quantities, ignoring the structural information between power grid nodes and lines. This neglect makes traditional methods insufficient in dealing with topological changes in the actual operation of power systems. Therefore, traditional data-driven methods are difficult to adapt to the complexity and dynamic changes of modern power systems, and new methods are urgently needed to make up for this shortcoming. In order to overcome these limitations, graph neural networks are introduced into the detection of false data injection attacks in smart grids. Graph neural networks are a deep learning method based on graph structured data. They can directly process power grid topological information and fully explore the relationship between nodes and edges, thereby improving the detection accuracy of abnormal states of power systems. Graph neural networks can automatically capture changes in power grid topology, making the model more flexible and robust in actual operation.
[0004] In summary, applying graph neural network technology to the detection of false data injection attacks in smart grids can help cope with complex and changeable grid topologies and operating environments, thereby improving the security protection level of the power system. Summary of the invention
[0005] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for detecting false data injection attacks on power CPS based on graph neural network.
[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0007] The present invention provides a method for detecting false data injection attacks of electric power CPS based on graph neural network, comprising the following steps:
[0008] 1) Construct the corresponding graph topology according to the distribution network model, considering the busbars in the distribution network as nodes and the connecting lines as edges;
[0009] 2) Incorporate the output of the pseudo-measurement model, the values of the measurement equipment, and the properties of the line into the corresponding area of the graph topology as input information for the graph neural network;
[0010] 3) Use the Graph Attention Network (GAT) model to model and analyze the power grid system, capture local structural information by introducing the attention mechanism, and assign learnable weight parameters to the association between each node;
[0011] 4) Perform anomaly detection on the output of the GAT model and set a threshold to detect abnormal fluctuations. When the residual exceeds the predetermined threshold, the system issues an alarm, indicating that there may be a false data injection attack.
[0012] As a preferred technical solution of the present invention, the mapping of global information is added to the GAT model, specifically: the global information is integrated into the graph neural network by using pooling operations, fully connected layers or customized global mapping layers; the dot product attention mechanism is used to perform feature fusion near important nodes. The dot product attention mechanism measures the similarity between nodes by calculating the dot product between them, thereby assigning weights.
[0013] As a preferred technical solution of the present invention, the dot product attention mechanism is combined with the original attention mechanism of GAT, and the global shared parameter attention mechanism of GAT is used in non-critical nodes or large-scale networks, and the dot product attention mechanism is preferentially used near important nodes.
[0014] As a preferred technical solution of the present invention, a multi-task learning framework is adopted to simultaneously perform dynamic state estimation and attack probability of distribution network nodes and share feature extraction modules, which can improve model performance.
[0015] As a preferred technical solution of the present invention, transfer learning technology is introduced to improve the generalization ability of the GAT model. Through transfer learning, a pre-trained model and parameters are used as a starting point, and then a new graph structure is used for fine-tuning on this basis.
[0016] As a preferred technical solution of the present invention, a heterogeneous graph neural network (HAN) is used to perform FDIA detection on the heterogeneous graph composed of the information physical layer in the power CPS. HAN effectively processes the heterogeneous information in the power CPS by introducing an attention mechanism, and achieves a balance between local structure and global characteristics.
[0017] As a preferred technical solution of the present invention, the false data injection attack detection of power CPS based on HAN includes the following steps: training HAN with data collected from the information layer and physical layer of the smart grid; obtaining the probability of false data injection attack on all nodes in the information network and the dynamic state estimation value of the node through HAN and multi-task learning; and performing a comprehensive analysis by comparing the state changes of the nodes and combining the probability of false data input attack on the nodes to improve the accuracy of false data injection attack detection in the information-physical fusion network.
[0018] As a preferred technical solution of the present invention, the comprehensive analysis adopts the Sugeno fuzzy inference system (TSK) in the fuzzy inference system (FIS), and uses the output node state difference of the graph neural network and the false data input attack probability as the input variables of the fuzzy inference system. The specific steps include:
[0019] 1) Design membership function: define three fuzzy subsets of low, medium and high for probability, divide the value range of state difference into two fuzzy subsets of low and high, and define the corresponding membership function;
[0020] 2) Design fuzzy rules: There are 6 fuzzy rules, which are determined based on practical problems and expert experience;
[0021] 3) Calculate the weight for each rule: the weight is equal to the product of the fuzzy subset membership corresponding to the input variable;
[0022] 4) Obtain the final output: Multiply the weight of each rule by the corresponding output value, sum them up, and then divide the result by the sum of the weights to obtain the normalized final output.
[0023] As a preferred technical solution of the present invention, the step of establishing a pseudo measurement model is also included, specifically:
[0024] Use CNN-BiLSTM-Attention neural network to predict node data. CNN-BiLSTM-Attention neural network combines convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM) and attention mechanism.
[0025] By collecting the equipment measurement data and generator output data in the distribution network as the input of the neural network, its output is the data of the important nodes or the nodes near the important nodes that are not equipped with the observers such as the Supervisory Control and Data Acquisition (SCADA) system and the Phase Measurement Unit (PMU);
[0026] The CNN-BiLSTM-Attention pseudo-measurement model performs pseudo-measurement output on the node voltage amplitude and voltage phase angle.
[0027] Compared with the prior art, the present invention has the following beneficial effects:
[0028] 1. This paper establishes a pseudo-measurement model, analyzes the distribution of existing high-precision measurement equipment in the power grid, and studies how to use the data obtained by existing equipment and the power grid structure information to build a pseudo-measurement model. This model will help to generate additional data with a certain degree of accuracy without adding actual measurement equipment, thereby improving the redundancy of power grid data, which will help to more effectively detect false data injection attacks on power CPS based on GNN;
[0029] 2. The present invention provides a detection method for false data injection attacks on power CPS based on graph neural networks. Graph neural networks can better mine and utilize grid structure information, thereby hopefully improving the accuracy and efficiency of FDIA detection;
[0030] 3. The present invention can realize multi-task learning and share feature representation in the same model, thereby improving the generalization ability of the model and making the detection results more reliable and stable;
[0031] 4. By making full use of data information at all levels of the smart grid, we can increase sensitivity to potential attacks and further strengthen grid security protection. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0033] Figure 1 It is a block diagram of the smart grid false data injection attack of the present invention;
[0034] Figure 2 It is the IEEE33 node distribution network topology diagram;
[0035] Figure 3 It is the flow chart of CNN-BiLSTM-Attention in the present invention;
[0036] Figure 4 It is the attention coefficient calculation graph;
[0037] Figure 5 This is the schematic diagram of false data detection. DETAILED DESCRIPTION
[0038] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0039] Example 1
[0040] The present invention provides a method for detecting false data injection attacks on power CPS based on graph neural network, and a method for detecting false data injection attacks on power CPS based on GAT.
[0041] Data preparation:
[0042] Collect relevant data of power CPS, including structural information of distribution network, data of measuring equipment, line attributes, etc.
[0043] Preprocess the data to ensure its accuracy and completeness.
[0044] Graph topology construction:
[0045] According to the distribution network model, the busbars are taken as nodes and the connecting lines as edges to construct an undirected graph topology.
[0046] Determine the properties of the vertices and edges in the graph, such as nodes representing buses and edges representing transmission lines.
[0047] GAT model training:
[0048] The preprocessed data is fed into the GAT model and the attention mechanism is applied.
[0049] Through the connection between spatial nodes, the information of neighboring nodes is aggregated.
[0050] A multi-head attention mechanism is used to stabilize the learning process, and Dropout and EarlyStopping are used to prevent overfitting.
[0051] False data injection attack detection:
[0052] Data preprocessing: Preprocess the data in the power CPS, convert equipment, lines, etc. into a graph structure, and provide labels for some nodes.
[0053] GAT model construction: Construct a GAT model to capture the topological relationships and high-order patterns in power CPS.
[0054] Model training and optimization: Use the training set to train the GAT model and adjust the hyperparameters to obtain the best detection performance.
[0055] Attack detection: The dynamic state estimation output by GAT is compared with the static estimation value to detect false data injection attacks through threshold detection and probability output.
[0056] Example 2: Detection of false data injection attack on the physical layer of power CPS based on improved GAT
[0057] Improved concentration:
[0058] The dot product attention mechanism is introduced into the GAT model. When calculating the attention coefficient, the features of the starting node and the target node are mapped using a matrix to enhance the expressiveness of the attention coefficient.
[0059] A global attention mechanism is added to perform an inner product operation on the features of each node and the global features to obtain a global attention weight term to make up for the deficiency of GAT in capturing long-distance dependencies.
[0060] Multi-task learning:
[0061] Using the multi-task learning framework, the neural network outputs the node status value and the node attack rate after the feature input.
[0062] The downstream network of the neural network consists of MLP layers with different parameters, which perform classification tasks and dynamic state estimation tasks respectively.
[0063] In the classification task, the fully connected layer and sigmoid function calculation model are used to give the membership of the node under attack, and the improved cross entropy loss function FocalLoss is used as the loss.
[0064] In the state estimation process, the mean square error (MSE) is used as the loss function of the estimation task.
[0065] The total loss function is the sum of the classification task and state estimation task loss functions.
[0066] Improved model generalization ability:
[0067] Based on the improved model, transfer learning training is added to enable the model to have good generalization capabilities for power grid diagrams with different topological structures.
[0068] Example 3: FDIA detection of CPS fusion information physical layer based on HAN
[0069] CPS cyber-physical network modeling:
[0070] The topology of each single-layer network is extracted from the power CPS to construct an electrical information-physical interdependent network.
[0071] Physical elements such as generators, substations and distributed generation are connected through power lines to build a power grid, and network devices are connected and organized into a network through communication channels.
[0072] Power CPS information physical feature extraction:
[0073] Collect the status of the smart grid industrial control network, including network information flow, data packets, communication rate, etc.
[0074] The collected traffic data is label encoded and one-hot encoded to form binary features.
[0075] The data of heterogeneous graphs on both sides of the cyber-physical network are synchronized in time to form an event sequence.
[0076] The correlation between the states of information network and physical network nodes is calculated, highly correlated features are deleted, and attribute dimension reduction is performed on the state vector.
[0077] The data features of the network layer and the data features of the physical layer are used as feature vectors of heterogeneous graph nodes to establish HAN.
[0078] Heterogeneous Graph Neural Network Modeling:
[0079] Specific types of transformation matrices are set for different nodes in the smart grid graph to map different types of node features.
[0080] The importance of nodes is calculated based on the meta-path, and the structural information is injected into the model through Masked attention to obtain the importance weight coefficients between node pairs.
[0081] Nodes are embedded based on the original path, aggregated through neighbor projection features with corresponding coefficients, and a multi-head attention mechanism is used to improve the stability of training.
[0083] The problem of meta-path selection and semantic fusion in heterogeneous graphs is solved through the semantic-level attention mechanism, and the final embedding is finally obtained by fusing specific embeddings of different semantics.
[0084] HAN-based detection of false data injection attacks on power CPS:
[0085] The HAN is trained by collecting data from the information layer and physical layer of the smart grid.
[0086] Through HAN and multi-task learning, the probability of false data injection attack on all nodes in the information network and the dynamic state estimation value of the nodes are obtained.
[0087] The Sugeno fuzzy inference system (TSK) in the fuzzy inference system (FIS) is adopted, and the output node state difference of the graph neural network and the false data input attack probability are taken as input variables.
[0088] Membership functions and fuzzy rules are designed, weights are calculated for each rule, and the final output is obtained, so as to comprehensively analyze and determine whether the node is attacked by false data injection.
[0089] Example 4: Establishment and application of pseudo-measurement model
[0090] Pseudo measurement model establishment:
[0091] The CNN-BiLSTM-Attention neural network is used for node data prediction, which combines the convolutional neural network (CNN), the bidirectional long short-term memory network (BiLSTM) and the attention mechanism.
[0092] CNN can extract features while retaining spatial information, BiLSTM can capture long-term dependencies in time series data, and the attention mechanism can assign different weights to different input data.
[0093] Data input and output:
[0094] The equipment measurement data, generator output data and other factors in the distribution network are collected as inputs to the neural network.
[0095] The output of the neural network is the data of important nodes or nodes adjacent to important nodes that are not equipped with a supervisory control and data acquisition (SCADA) system, a phasor measurement unit (PMU) and other observers.
[0096] Pseudo measurement output and verification:
[0097] The CNN-BiLSTM-Attention pseudo-measurement model performs pseudo-measurement output on the node voltage amplitude and voltage phase angle.
[0098] For example, nodes 14, 19, 18, 22, 25, and 33 are used to perform pseudo-measurement output on the voltage amplitude and phase angle at node 2, and compared with the true values to verify the accuracy of the pseudo-measurement model.
[0099] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for detecting false data injection attacks in power CPS based on graph neural network, It is characterized in that The following steps are involved: 1) Construct the corresponding graph topology according to the distribution network model, taking the busbars in the distribution network as nodes and the connecting lines as edges; 2) Incorporate the output of the pseudo-measurement model, the values of the measurement equipment, and the properties of the line into the corresponding area of the graph topology as input information for the graph neural network; 3) Use the graph attention network model to model and analyze the power grid system, capture local structural information by introducing the attention mechanism, and assign learnable weight parameters to the association between each node; 4) Perform anomaly detection on the output of the GAT model and set a threshold to detect abnormal fluctuations. When the residual exceeds the predetermined threshold, the system issues an alarm, indicating that there may be a false data injection attack.
2. According to the graph neural network-based power CPS false data injection attack detection method of claim 1, It is characterized in that The GAT model adds a mapping of global information, specifically: Use pooling operations, fully connected layers, or custom global mapping layers to integrate global information into graph neural networks; The dot product attention mechanism is used for feature fusion near important nodes. The dot product attention mechanism measures the similarity between nodes by calculating the dot product between them, thereby assigning weights.
3. According to the graph neural network-based power CPS false data injection attack detection method of claim 2, It is characterized in that The dot product attention mechanism is combined with the original attention mechanism of GAT, and the global shared parameter attention mechanism of GAT is used in non-critical nodes or large-scale networks, and the dot product attention mechanism is preferentially adopted near important nodes.
4. According to the graph neural network-based power CPS false data injection attack detection method of claim 1, It is characterized in that A multi-task learning framework is adopted to simultaneously estimate the dynamic state of distribution network nodes and the attack probability, and share the feature extraction model to improve model stability and recognition accuracy.
5. According to the graph neural network-based power CPS false data injection attack detection method of claim 4, It is characterized in that The transfer learning technology is introduced to improve the generalization ability of the GAT model. Through transfer learning, the pre-trained model and parameters are used as the starting point, and then the new graph structure is used for fine-tuning on this basis.
6. According to the graph neural network-based power CPS false data injection attack detection method of claim 1, It is characterized in that A heterogeneous graph neural network is used to perform FDIA detection on the heterogeneous graph composed of the information physical layer in the power CPS. HAN effectively processes the heterogeneous information in the power CPS by introducing the attention mechanism and achieves a balance between local structure and global characteristics.
7. According to claim 6, the method for detecting false data injection attacks on power CPS based on graph neural network, It is characterized in that The HAN-based detection of false data injection attacks on power CPS includes the following steps: The HAN is trained on the data collected from the information layer and physical layer of the smart grid; Through HAN and multi-task learning, the probability of false data injection attack on all nodes in the information network and the dynamic state estimation value of the node are obtained; By comparing the state changes of nodes and combining the probability of false data input attacks on nodes, a comprehensive analysis is performed to improve the accuracy of false data injection attack detection in cyber-physical fusion networks.
8. According to the graph neural network-based power CPS false data injection attack detection method of claim 7, It is characterized in that The comprehensive analysis adopts the Sugeno fuzzy inference system in the fuzzy inference system, and takes the output node state difference of the graph neural network and the false data input attack probability as the input variables of the fuzzy inference system. The specific steps include: 1) Design membership function: define three fuzzy subsets of low, medium and high for probability, divide the value range of state difference into two fuzzy subsets of low and high, and define the corresponding membership function; 2) Design fuzzy rules: There are 6 fuzzy rules, which are determined based on practical problems and expert experience; 3) Calculate the weight for each rule: the weight is equal to the product of the fuzzy subset membership corresponding to the input variable; 4) Obtain the final output: Multiply the weight of each rule by the corresponding output value, sum them up, and then divide the result by the sum of the weights to obtain the normalized final output.
9. The method for detecting false data injection attacks on power CPS based on graph neural network according to any one of claims 1 to 8, It is characterized in that It also includes the steps of establishing a pseudo measurement model, specifically: Use CNN-BiLSTM-Attention neural network to predict node data. CNN-BiLSTM-Attention neural network combines convolutional neural network, bidirectional long short-term memory network and attention mechanism. By collecting the equipment measurement data and generator output data in the distribution network as the input of the neural network, its output is the data of the important nodes or the nodes adjacent to the important nodes that are not equipped with the monitoring and data acquisition system, the phase measurement unit and the observer; The CNN-BiLSTM-Attention pseudo-measurement model performs pseudo-measurement output on the node voltage amplitude and voltage phase angle.
Citation Information
Patent Citations
False data injection attack recognition method based on line topology analysis and power flow characteristics
CN113904786A
Power distribution network fault diagnosis and positioning method based on super-resolution and graph neural network
CN114779015A
False data injection attack detection method, detection terminal and storage medium
CN116032553A
Power distribution network prediction auxiliary interval state estimation method for network attack detection
CN117039890A
Power system state estimation method and system under false data injection attack
CN117175550A
Cited By
Node injection attack method and device based on target subgraph division
CN119323234A
False data injection attack and defense method based on graph neural network parameter estimation priority
CN122247754A
A spoofing attack and defense method based on graph neural network parameter estimation priority
CN122247754B
