False data attack detection method and device for energy storage system of power distribution network

Abstract

The invention provides a false data attack detection method and device for a power distribution network energy storage system, and relates to the technical field of power system information security. The method comprises the following steps: acquiring spatial-temporal characteristic data to be measured according to a power distribution network topological structure and multi-source data; inputting spatio-temporal characteristic data to be detected into the trained enhanced physical information neural network model, outputting an abnormal score value, comparing the abnormal score value with a dynamic threshold value, and judging whether false data attack occurs or not; the training process of the enhanced physical information neural network model comprises the following steps: acquiring training spatio-temporal feature data; respectively carrying out spatial feature extraction and time feature extraction on the training spatial-temporal feature data through a data network and a physical network; and simultaneously training the data network and the physical network by minimizing the loss function until the loss function converges. The method can solve the problems that a traditional detection method based on state estimation is easy to bypass, and an existing physical information neural network is poor in constraint rigidity and a pure data driving method is poor in interpretability.

Description

Technical Field

[0001] This invention relates to the field of power system information security technology, and in particular to a method and device for detecting fake data attacks in distribution network energy storage systems. Background Technology

[0002] As new power systems evolve towards higher proportions of renewable energy and bidirectional interactive models, distributed energy storage systems have become critical nodes. However, the deep coupling between the information and physical layers also exposes them to new cyber-physical attacks, particularly spoofed data injection attacks targeting energy storage systems. These attacks manipulate measured values ​​(such as voltage, current, and state of charge) with meticulously crafted malicious data, bypassing traditional state estimation-based residual detection while inducing the energy management system to make incorrect decisions, leading to frequency instability, equipment overload, and even cascading failures.

[0003] Existing detection methods have the following main shortcomings: Purely data-driven methods, such as recurrent neural networks and graph neural networks, can capture complex spatiotemporal patterns, but they rely heavily on massive amounts of labeled data. They have poor generalization ability to "zero-day" attacks that have not appeared in the training set, and due to their "black box" nature, their decision-making process lacks physical interpretability.

[0004] Methods based on pure physical models, such as extended Kalman filtering based on residual detection, have clear physical meaning, but are sensitive to model accuracy and noise statistics. They are prone to failure when faced with carefully designed cooperative attacks that satisfy some physical constraints.

[0005] Traditional physical information neural networks: They initially attempt to integrate physical laws with neural networks, but their physical constraints are usually added to the loss function in the form of fixed weights and hard boundaries. This makes it difficult to balance the uncertainty of physical models, measurement noise and real attack signals in complex scenarios, resulting in insufficient sensitivity and accuracy in detecting covert attacks such as slow injection and multi-node collaboration.

[0006] Therefore, there is an urgent need for a detection scheme that can deeply integrate the physical mechanism of the power grid with data intelligence and has adaptive learning and reasoning capabilities, so as to accurately identify various complex attacks against energy storage systems. Summary of the Invention

[0007] This invention provides a method and device for detecting fake data attacks in distribution network energy storage systems. It solves the problems of traditional state estimation-based detection methods being easily bypassed, the rigidity of existing physical information neural network constraints, and the poor interpretability of pure data-driven methods. It achieves a balance between high detection rate, low false alarm rate, and strong interpretability, significantly improving the security defense capability of distribution network energy storage systems against new fake data attacks.

[0008] In a first aspect, embodiments of the present invention provide a method for detecting spoofing attacks on a power distribution network energy storage system, comprising: Collect multi-source data from the distribution network energy storage system, and obtain the spatiotemporal characteristic data to be measured based on the distribution network topology and the multi-source data. The spatiotemporal feature data to be tested is input into a trained augmented physical information neural network model, and anomaly scores are output. The anomaly scores are compared with dynamic thresholds to determine whether a false data attack has occurred. The training process for the augmented physical information neural network model is as follows: Training spatiotemporal feature data is obtained based on multi-source training data and the distribution network topology of the distribution network energy storage system. Spatial and temporal features are extracted from the training spatiotemporal feature data through a first graph attention network and a first gated recurrent unit in the data network. Spatial and temporal features are also extracted from the training spatiotemporal feature data through a second graph attention network and a second gated recurrent unit embedded with physical constraints in the physical network. The data network and physical network are trained simultaneously by minimizing the loss function until the loss function converges. The loss function is a weighted sum of the physical network loss, the data network loss, and the regularization loss, and the weight of the physical network loss is an adaptive relaxation factor.

[0009] In one possible implementation, the step of extracting spatial and temporal features from the training spatiotemporal feature data through a first graph attention network and a first gated recurrent unit in the data network includes: Linear transformations are applied to the training spatiotemporal feature data to obtain augmented data. The enhanced data is input into a first graph attention network to extract spatial features and output the first spatial features. The first spatial feature is input into the first gated loop unit to extract the temporal feature and output the first spatiotemporal feature.

[0010] In one possible implementation, spatial and temporal feature extraction is performed on the training spatiotemporal feature data through a second graph attention network with embedded physical constraints and a second gated recurrent unit within the physical network, including: Linear transformations are applied to the training spatiotemporal feature data to obtain augmented data. The enhanced data is input into a second graph attention network embedded with physical constraints to extract spatial features and output second spatial features. The second spatial feature is input into the second gated loop unit to extract the temporal feature and output the second spatiotemporal feature.

[0011] In one possible implementation, the loss function is: ; in, It is a loss function. It is a loss of data network. It is a physical network loss. It is the regularization loss. for t Time-adaptive relaxation factor It is a temperature parameter; ; in, It is the basic relaxation factor value. It is the adjustment range of the relaxation factor value. yes t Estimated noise level of time-matter data. It is the noise threshold. It is the adjustment coefficient.

[0012] In one possible implementation, after extracting spatial and temporal features from the training spatiotemporal feature data through the first graph attention network and the first gated loop unit in the data network, the method further includes: reconstructing the extracted first spatiotemporal features through a reconstruction function to obtain the prediction sample corresponding to the training spatiotemporal feature data. The data network loss is: ; The physical network loss is: ; The regularization loss is: ; in, This represents the number of samples in the multi-source training data. It is the first output of the data network. i Predicted samples corresponding to multiple source training data It is the first i Multiple training data sources It is the first weighting coefficient. It is the second weighting coefficient. It is the third weighting coefficient. It is the fourth weighting coefficient. It is the battery equivalent circuit constraint loss. Represents the network topology constraint loss. This represents the energy conservation constraint loss. This represents the thermal dynamic constraint loss. This represents the set of all trainable weight parameters in a neural network model that enhances physical information. It is the first regularization coefficient. It is the second regularization coefficient. It is an L1 norm. It is an L2 norm.

[0013] In one possible implementation, the spatiotemporal feature data to be tested is input into a trained augmented physical information neural network model, which outputs anomaly scores. The anomaly scores are then compared with a dynamic threshold to determine whether a spoofing attack has occurred, including: The spatial and temporal features of the spatiotemporal feature data to be tested are extracted by the first graph attention network and the first gated recurrent unit in the data network of the trained augmented physical information neural network model, and the data network output residual is obtained. Spatial and temporal features of the spatiotemporal feature data to be tested are extracted by a second graph attention network and a second gated recurrent unit embedded in the physical network to obtain the output residual of the physical network. The residual output from the data network and the residual output from the physical network are combined to output an anomaly score.

[0014] In one possible implementation, the data network output residual and the physical network output residual are fused to output anomaly scores, including: according to The residual output from the data network and the residual output from the physical network are combined to output an anomaly score. in, These are abnormal score values. It is the physical network output residual. It is the data network output residual. It is the physical network normalization factor. It is the data network normalization factor. It is the physical network fusion weight. It is the weight of data network fusion; The physical network fusion weights and the data network fusion weights are adaptively assigned through confidence level evaluation: ; ; in, It is the confidence level of the physical network. It is the confidence level of the data network. It is a temperature parameter; ; ; in, This represents the number of samples in the multi-source training data. It is a preset threshold for the physical network. It is a preset threshold for the data network. It is a norm.

[0015] In one possible implementation, the dynamic threshold is obtained based on the detection baseline threshold and the residual fluctuation of the physical network output.

[0016] In one possible implementation, the dynamic threshold is: ; in, It is the aforementioned basic detection threshold. It is the residual fluctuation of the physical network output. yes t A measure of the volatility of the residual output by the physical network at any given time. k It is the sensitivity adjustment coefficient.

[0017] Secondly, embodiments of the present invention provide a false data attack detection device for a power distribution network energy storage system, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect or any possible implementation of the first aspect.

[0018] In this invention, addressing the core shortcomings of traditional state estimation-based fake data attack detection methods—such as their susceptibility to bypassing and insufficient detection capability against novel collaborative attacks—a dual-driven physical-data verification mechanism is constructed. This fundamentally undermines the attacker's ability to construct perfect attack vectors using the null space of a linear system. An adaptive relaxation factor is introduced as a weight for the physical network loss, along with a dynamic threshold, enabling the model to intelligently balance physical priors and data uncertainties. This significantly improves detection accuracy and generalization ability while ensuring strong interpretability. This invention not only effectively compensates for the inherent vulnerabilities of traditional methods but also provides crucial technical support for building a new generation of smart grid protection systems with inherent security capabilities. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the training process of the enhanced physical information neural network model provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the implementation of the method for detecting fake data attacks in a power distribution network energy storage system provided in this embodiment of the invention. Figure 3 This is a schematic diagram of the topology architecture of the IEEE 14-node smart grid energy storage system provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating the implementation of a fake data attack detection method based on an IEEE 14-node smart grid energy storage system provided in this embodiment of the invention. Figure 5 This is a schematic diagram of the structure of the fake data attack detection device for the power distribution network energy storage system provided in an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] See Figure 1 The flowchart illustrating the training implementation of the enhanced physical information neural network model provided in this embodiment of the invention is described in detail below: In step 101, training spatiotemporal feature data are obtained based on multi-source training data of the distribution network energy storage system and the distribution network topology.

[0022] For example, multi-source data from the distribution network energy storage system are collected, including electrical measurement data from the monitoring system, synchronization phasor data, and status data from the energy storage battery management system. The distribution network topology is modeled as a graph structure. , where the set of nodes Represents the parent node and the set of edges. Representing electrical connection relationships, node attribute matrix It includes real-time measurement data. A spatiotemporal feature tensor (i.e., training spatiotemporal feature data) is constructed, incorporating the spatial topological relationships and time-series dependencies of the distribution network energy storage system. ,in For the number of nodes, The length of the time window. For feature dimensions.

[0023] In step 102, spatial and temporal features are extracted from the training spatiotemporal feature data through the first graph attention network and the first gated recurrent unit in the data network.

[0024] In this embodiment, the training spatiotemporal feature data is linearly transformed to obtain augmented data; the augmented data is input into the first graph attention network to extract spatial features and output the first spatial feature; the first spatial feature is input into the first gated recurrent unit to extract temporal features and output the first spatiotemporal feature.

[0025] For example, for Spatiotemporal feature data of node training Perform shared linear transformations to enhance feature representation capabilities: ; ; in, It is a node Enhanced input feature data, It is a node The augmented input feature data is then called the augmented data. This is a learnable weight matrix.

[0026] Define the graph attention coefficient as: ; in, For learnable weight matrix, For attention vectors, This represents vector concatenation. It is the graph attention coefficient. i and j All of them are nodes in the distribution network energy storage system, among which, This indicates a node All neighbors Summation; LeakyReLU is the activation function.

[0027] The spatial features output by the attention network in the first graph are as follows: ; in, It is a non-linear activation function.

[0028] Gated loop units are used to handle timing dependencies, and their calculation formula is as follows: ; in, It is a sigmoid activation function. It resets the learnable weights of the gates. It resets the learnable bias of the gate. yes t Reset the gate output at all times. It is the hyperbolic tangent activation function. These are the learnable weights of the candidate states. It is a learnable bias for the candidate state. It updates the learnable weights of the gates. It is an update gate that can learn the bias. This indicates element-wise multiplication. yes Temporarily hide your status at all times. yes t Update the gate output constantly. yes Time Node i The input state of the gated loop unit (also) (Moment graph network output features) yes Time Node The output of the gated cyclic unit model is the first spatiotemporal feature.

[0029] In step 103, spatial and temporal features are extracted from the training spatiotemporal feature data through a second graph attention network and a second gated recurrent unit that embed physical constraints in the physical network.

[0030] In this embodiment, the training spatiotemporal feature data is linearly transformed to obtain augmented data; the augmented data is input into a second graph attention network embedded with physical constraints to extract spatial features and output a second spatial feature; the second spatial feature is input into a second gated recurrent unit to extract temporal features and output a second spatiotemporal feature.

[0031] For example, after embedding physical constraints in the second graph attention network, step 102 is repeated to output the second spatiotemporal features.

[0032] Among them, physical constraints include battery equivalent circuit constraints, network topology constraints, energy conservation constraints, and thermal dynamic constraints.

[0033] In step 104, the data network and the physical network are trained simultaneously by minimizing the loss function until the loss function converges.

[0034] After extracting spatial and temporal features from the training spatiotemporal feature data through the first graph attention network and the first gated recurrent unit in the data network, the method further includes: reconstructing the extracted first spatiotemporal features through a reconstruction function to obtain the prediction samples corresponding to the training spatiotemporal feature data.

[0035] For example, the prediction sample corresponding to the training spatiotemporal feature data is: ; in, It is the first output of the data network. i Predicted samples corresponding to multiple source training data It is a refactoring function.

[0036] The loss function is: ; in, It is a loss function. It is a loss of data network. It is a physical network loss. It is the regularization loss. for t Time-adaptive relaxation factor It is a temperature parameter; ; in, It is the basic relaxation factor value. It is the adjustment range of the relaxation factor value. yes tEstimated noise level of time-matter data. It is the noise threshold. It is the adjustment coefficient.

[0037] For example, based on experience, set , , , .

[0038] The data network loss is: ; The physical network loss is: ; The regularization loss is: ; in, This represents the number of samples in the multi-source training data. It is the first i The predicted data output by each model. It is the first i Multiple training data sources It is the first weighting coefficient. It is the second weighting coefficient. It is the third weighting coefficient. It is the fourth weighting coefficient. It is the battery equivalent circuit constraint loss. Represents the network topology constraint loss. This represents the energy conservation constraint loss. This represents the thermal dynamic constraint loss. This represents the set of all trainable weight parameters in a neural network model that enhances physical information. It is the first regularization coefficient. It is the second regularization coefficient. It is an L1 norm. It is an L2 norm.

[0039] For example, , .

[0040] If the loss function converges to satisfy the following constraint equation, the training process is terminated, and the trained augmented physical information neural network model is obtained: ; in, For example, the convergence threshold is set to... .

[0041] Based on the above, such as Figure 2The flowchart illustrating the implementation of the spoofing attack detection method for distribution network energy storage systems provided in this embodiment of the invention is detailed below: In step 201, multi-source data of the distribution network energy storage system are collected, and spatiotemporal characteristic data to be measured are obtained based on the distribution network topology and multi-source data.

[0042] In step 202, the spatiotemporal feature data to be tested is input into the trained augmented physical information neural network model, and the abnormal score is output. The abnormal score is compared with the dynamic threshold to determine whether a false data attack has occurred.

[0043] In this embodiment, spatial and temporal features are extracted from the spatiotemporal feature data to be tested using the first graph attention network and the first gated recurrent unit in the data network of the trained augmented physical information neural network model, to obtain the data network output residual; spatial and temporal features are extracted from the spatiotemporal feature data to be tested using the second graph attention network and the second gated recurrent unit embedded with physical constraints in the physical network, to obtain the physical network output residual; the data network output residual and the physical network output residual are fused to output the anomaly score.

[0044] according to It integrates the residuals from the data network output and the residuals from the physical network output to output anomaly scores.

[0045] in, These are abnormal score values. It is the physical network output residual. The problem is a residual in the data network. It is the physical network normalization factor. It is the data network normalization factor. It is the physical network fusion weight. It is the weight of data network fusion.

[0046] For example, the physical path output residual is: ; The data path output residual is: ; ; in, Indicates matrix concatenation. The output residual is constrained by the battery equivalent circuit. It is the network topology constraint output residual. It is the energy conservation constraint output residual. It is the thermal dynamic constraint physical output residual. It is a data network node i Output residual.

[0047] Physical network fusion weights and data network fusion weights are adaptively assigned using confidence level evaluation: ; ; in, It is the confidence level of the physical network. It is the confidence level of the data network. It is a temperature parameter.

[0048] ; ; in, This represents the number of samples in the multi-source training data. It is a preset threshold for the physical network. It is a preset threshold for the data network. It is a norm.

[0049] In this embodiment, the dynamic threshold is obtained based on the detection baseline threshold and the residual fluctuation of the physical network output.

[0050] The dynamic threshold is: ; in, It is the basic detection threshold. It is the residual fluctuation of the physical network output. yes t A measure of the volatility of the residual output by the physical network at any given time. k It is the sensitivity adjustment coefficient.

[0051] For example, It is determined by the statistical distribution of normal data (such as the 99.7th percentile); .

[0052] At each detection time, if It was determined that no fake data attack had occurred.

[0053] At each detection time, if It can detect when a fake data attack has occurred and issue an attack alert.

[0054] It should be understood that the sequence number of each step in the above embodiments 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 the present invention.

[0055] The following specific embodiments illustrate the method for detecting fake data attacks in a power distribution network energy storage system provided by the present invention: like Figure 3 As shown, the IEEE 14-node smart grid energy storage system topology is configured as follows: The network comprises 14 bus nodes and corresponding radial / mesh branches; one lithium-ion battery energy storage system is deployed, connected to key node 5, with a total energy storage capacity of 1.2MWh and a rated total power of 500kW. The measurement system provides comprehensive coverage, including synchronous phasor measurement units deployed on all nodes, fully covered smart meters, and battery management systems for each energy storage unit.

[0056] like Figure 4 As shown, the method for detecting fake data attacks based on the IEEE 14-node smart grid energy storage system includes the following steps: S1. Multi-source heterogeneous data fusion and spatiotemporal feature construction.

[0057] Collect multi-source data from the IEEE 14-node smart grid energy storage system, including electrical measurement data from the monitoring system, synchronization phasor data, and status data from the energy storage battery management system. Model the distribution network topology as a graph structure. , where the set of nodes Represents the parent node and the set of edges. Representing electrical connection relationships, node attribute matrix It includes real-time measurement data. A spatiotemporal feature tensor (i.e., training spatiotemporal feature data) is constructed, incorporating the spatial topological relationships and time-series dependencies of the distribution network energy storage system. ,in For the number of nodes, The length of the time window. For feature dimensions.

[0058] S2. Construct an enhanced physical information neural network model.

[0059] S21 extracts spatial and temporal features from the training spatiotemporal feature data through the first graph attention network and the first gated recurrent unit in the data network.

[0060] Linear transformations are applied to the training spatiotemporal feature data to obtain augmented data.

[0061] right Spatiotemporal feature data of node training Perform shared linear transformations to enhance feature representation capabilities: ; ; in, It is a node Enhanced input feature data, It is a node The augmented input feature data is then called the augmented data. This is a learnable weight matrix.

[0062] The enhanced data is input into the first graph attention network to extract spatial features and output the first spatial features.

[0063] Define the graph attention coefficient as: ; in, For learnable weight matrix, For attention vectors, This represents vector concatenation. It is the graph attention coefficient. i and j All of them are nodes in the distribution network energy storage system, among which, This indicates a node All neighbors Summation; LeakyReLU is the activation function.

[0064] The spatial features output by the attention network in the first graph are as follows: ; in, It is a non-linear activation function.

[0065] The first spatial feature is input into the first gated loop unit, the temporal feature is extracted, and the first spatiotemporal feature is output.

[0066] Gated loop units are used to handle timing dependencies, and their calculation formula is as follows: ; in, It is a sigmoid activation function. It resets the learnable weights of the gates. It resets the learnable bias of the gate. yes t Reset the gate output at all times. It is the hyperbolic tangent activation function. These are the learnable weights of the candidate states. It is a learnable bias for the candidate state. It updates the learnable weights of the gates. It is an update gate that can learn the bias. This indicates element-wise multiplication. yes Temporarily hide your status at all times. yes t Update the gate output constantly. yes Time Node i The input state of the gated loop unit (also) (Moment graph network output features) yes Time Node The output of the gated cyclic unit model is the first spatiotemporal feature.

[0067] S22 uses a second graph attention network and a second gated recurrent unit embedded in the physical network to extract spatial and temporal features from the training spatiotemporal feature data.

[0068] Linear transformation is applied to the training spatiotemporal feature data to obtain augmented data; the augmented data is input into a second graph attention network embedded with physical constraints to extract spatial features and output second spatial features; the second spatial features are input into a second gated recurrent unit to extract temporal features and output second spatiotemporal features.

[0069] After embedding physical constraints in the second graph attention network, repeat S21 to output the second spatiotemporal feature.

[0070] Among them, physical constraints include battery equivalent circuit constraints, network topology constraints, energy conservation constraints, and thermal dynamic constraints.

[0071] This invention embeds battery equivalent circuit constraints, network topology constraints, energy conservation constraints, and thermal dynamic constraints into a graph attention network, fundamentally undermining the basis for attackers to construct perfect attack vectors using the null space of a linear system.

[0072] S23 merges the residuals from the data network and the physical network outputs, and outputs the abnormal score values.

[0073] according to It integrates the residuals from the data network output and the residuals from the physical network output to output anomaly scores.

[0074] in, These are abnormal score values. It is the physical network output residual. The problem is a residual in the data network. It is the physical network normalization factor. It is the data network normalization factor. It is the physical network fusion weight. It is the weight of data network fusion.

[0075] The physical path output residual is: ; The data path output residual is: ; ; ; in, Indicates matrix concatenation. The output residual is constrained by the battery equivalent circuit. It is the network topology constraint output residual. It is the energy conservation constraint output residual. It is the thermal dynamic constraint physical output residual. It is a data network node i Output residual, It is the first output of the data network. i Predicted samples corresponding to multiple source training data It is a refactoring function. It is the first i Multiple training data sources.

[0076] Physical network fusion weights and data network fusion weights are adaptively assigned using confidence level evaluation: ; ; in, It is the confidence level of the physical network. It is the confidence level of the data network. It is a temperature parameter.

[0077] ; ; in, This represents the number of samples in the multi-source training data. It is a preset threshold for the physical network. It is a preset threshold for the data network. It is a norm.

[0078] S3. Construct the loss function.

[0079] The loss function is: ; in, It is a loss function. It is a loss of data network. It is a physical network loss. It is the regularization loss. for t Time-adaptive relaxation factor It is a temperature parameter.

[0080] ; in, It is the basic relaxation factor value. It is the adjustment range of the relaxation factor value. yest Estimated noise level of time-matter data. It is the noise threshold. It is the adjustment coefficient.

[0081] Based on experience , , , .

[0082] The data network loss is: ; The physical network loss is: ; The regularization loss is: ; in, This represents the number of samples in the multi-source training data. It is the first i The predicted data output by each model. It is the first i Multiple training data sources It is the first weighting coefficient. It is the second weighting coefficient. It is the third weighting coefficient. It is the fourth weighting coefficient. It is the battery equivalent circuit constraint loss. Represents the network topology constraint loss. This represents the energy conservation constraint loss. This represents the thermal dynamic constraint loss. This represents the set of all trainable weight parameters in a neural network model that enhances physical information. It is the first regularization coefficient. It is the second regularization coefficient. , ; It is an L1 norm. It is an L2 norm.

[0083] S4. Train a neural network model for enhanced physical information using a loss function.

[0084] If the loss function converges to satisfy the following constraint equation, the training process is terminated, and the trained augmented physical information neural network model is obtained: ; in, To be the convergence threshold, take .

[0085] S5, Attack Detection.

[0086] Attack detection threshold It is not a fixed value, but is dynamically adjusted according to the current operating status of the system in order to reduce the false alarm rate when the system is under significant disturbance.

[0087] The dynamic threshold is: ; in, It is the basic detection threshold. It is the residual fluctuation of the physical network output. yes t A measure of the volatility of the residual output by the physical network at any given time. k This is the sensitivity adjustment coefficient. Among them, It is determined by the statistical distribution of normal data (such as the 99.7th percentile); .

[0088] At each detection time, if It was determined that no fake data attack had occurred.

[0089] At each detection time, if It can detect when a fake data attack has occurred and issue an attack alert.

[0090] S6. Test Results and Comparison.

[0091] To verify the superiority of this invention, it was compared with standard physical information neural networks and traditional chi-square detection methods. The comparison results are shown in Table 1.

[0092] Table 1

[0093] Results Analysis: The enhanced physical information neural network detection method proposed in this invention shows a significant improvement in detection performance compared to the traditional chi-square detection method and the standard physical information neural network. Experimental results show that the detection rate of this method reaches 98.7%, which is 9.4 percentage points higher than the traditional chi-square detection method, and the false alarm rate is reduced to 0.9%, a reduction of more than 57%. Furthermore, compared to the standard physical information neural network, it further improves the detection rate by 3.7 percentage points and reduces the false alarm rate by 0.6 percentage points. This performance gain stems from the adaptive physical constraint mechanism and the confidence fusion strategy of the data network and physical network introduced in this invention. This overcomes the inherent defects of traditional methods that rely on linearized models, fixed thresholds, and static physical constraints, and can effectively identify covert attacks that satisfy residual statistical characteristics, reducing false alarms caused by model uncertainty and changes in operating conditions. Therefore, this invention achieves synergistic optimization of high detection rate and low false alarm rate while maintaining physical interpretability, verifying its effectiveness and technological advancement in resisting new types of false data attacks in power systems.

[0094] This invention also provides a fake data attack detection device 5 for a power distribution network energy storage system, including a memory 51 and a processor 50. The memory 51 stores a computer program 52, and the processor 50 executes the computer program 52 to implement the method described in the above method embodiment.

[0095] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0096] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting fake data attacks in a power distribution network energy storage system, characterized in that, include: Collect multi-source data from the distribution network energy storage system, and obtain the spatiotemporal characteristic data to be measured based on the distribution network topology and the multi-source data. The spatiotemporal feature data to be tested is input into a trained augmented physical information neural network model, and anomaly scores are output. The anomaly scores are compared with dynamic thresholds to determine whether a false data attack has occurred. The training process for the augmented physical information neural network model is as follows: Training spatiotemporal feature data is obtained based on multi-source training data and the distribution network topology of the distribution network energy storage system. Spatial and temporal features are extracted from the training spatiotemporal feature data through a first graph attention network and a first gated recurrent unit in the data network. Spatial and temporal features are also extracted from the training spatiotemporal feature data through a second graph attention network and a second gated recurrent unit embedded with physical constraints in the physical network. The data network and physical network are trained simultaneously by minimizing the loss function until the loss function converges. The loss function is a weighted sum of the physical network loss, the data network loss, and the regularization loss, and the weight of the physical network loss is an adaptive relaxation factor.

2. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 1, characterized in that, The step of extracting spatial and temporal features from the training spatiotemporal feature data through a first graph attention network and a first gated recurrent unit in the data network includes: Linear transformations are applied to the training spatiotemporal feature data to obtain augmented data. The enhanced data is input into a first graph attention network to extract spatial features and output the first spatial features. The first spatial feature is input into the first gated loop unit to extract the temporal feature and output the first spatiotemporal feature.

3. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 1, characterized in that, Spatial and temporal features of the training spatiotemporal feature data are extracted using a second graph attention network with embedded physical constraints and a second gated recurrent unit within the physical network, including: Linear transformations are applied to the training spatiotemporal feature data to obtain augmented data. The enhanced data is input into a second graph attention network embedded with physical constraints to extract spatial features and output second spatial features. The second spatial feature is input into the second gated loop unit to extract the temporal feature and output the second spatiotemporal feature.

4. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 1, characterized in that, The loss function is: ； in, It is a loss function. It is a loss of data network. It is a physical network loss. It is the regularization loss. for t Time-adaptive relaxation factor It is a temperature parameter; ； in, It is the basic relaxation factor value. It is the adjustment range of the relaxation factor value. yes t Estimated noise level of time-matter data. It is the noise threshold. It is the adjustment coefficient.

5. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 4, characterized in that, After extracting spatial and temporal features from the training spatiotemporal feature data through the first graph attention network and the first gated loop unit in the data network, the method further includes: reconstructing the extracted first spatiotemporal features through a reconstruction function to obtain the prediction sample corresponding to the training spatiotemporal feature data. The data network loss is: ； The physical network loss is: ； The regularization loss is: ； in, This represents the number of samples in the multi-source training data. It is the first output of the data network. i Predicted samples corresponding to multiple source training data It is the first i Multiple training data sources It is the first weighting coefficient. It is the second weighting coefficient. It is the third weighting coefficient. It is the fourth weighting coefficient. It is the battery equivalent circuit constraint loss. Represents the network topology constraint loss. This represents the energy conservation constraint loss. Indicates thermal dynamic constraint loss, This represents the set of all trainable weight parameters in a neural network model that enhances physical information. It is the first regularization coefficient. It is the second regularization coefficient. It is an L1 norm. It is an L2 norm.

6. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 1, characterized in that, The spatiotemporal feature data to be tested is input into a trained augmented physical information neural network model, which outputs anomaly scores. These anomaly scores are then compared with a dynamic threshold to determine whether a spoofing attack has occurred, including: The spatial and temporal features of the spatiotemporal feature data to be tested are extracted by the first graph attention network and the first gated recurrent unit in the data network of the trained augmented physical information neural network model, and the data network output residual is obtained. Spatial and temporal features of the spatiotemporal feature data to be tested are extracted by a second graph attention network and a second gated recurrent unit embedded in the physical network to obtain the output residual of the physical network. The residual output from the data network and the residual output from the physical network are combined to output an anomaly score.

7. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 6, characterized in that, By fusing the data network output residual and the physical network output residual, an anomaly score is output, including: according to The residual output from the data network and the residual output from the physical network are combined to output an anomaly score. in, These are abnormal score values. It is the physical network output residual. It is the data network output residual. It is the physical network normalization factor. It is the data network normalization factor. It is the physical network fusion weight. It is the weight of data network fusion; The physical network fusion weights and the data network fusion weights are adaptively assigned through confidence level evaluation: ； ； in, It is the confidence level of the physical network. It is the confidence level of the data network. It is a temperature parameter; ； ； in, This represents the number of samples in the multi-source training data. It is a preset threshold for the physical network. It is a preset threshold for the data network. It is a norm.

8. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 1, characterized in that, The dynamic threshold is obtained based on the detection baseline threshold and the residual fluctuation of the physical network output.

9. The method for detecting fake data attacks in a power distribution network energy storage system according to claim 8, characterized in that, The dynamic threshold is: ； in, It is the aforementioned basic detection threshold. It is the residual fluctuation of the physical network output. yes t A measure of the volatility of the residual output by the physical network at any given time. k It is the sensitivity adjustment coefficient.

10. A method and device for detecting fake data attacks in a power distribution network energy storage system, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 9.

Images