A method, application method, equipment and medium for detecting power grid load redistribution attacks.

By constructing a power grid load redistribution attack detection method, and utilizing differential values ​​and multi-level decomposition techniques, combined with projection distance and minimum alignment distance, the low efficiency of existing detection methods under abnormal data from individual sensors is solved, achieving efficient and accurate attack identification.

CN123122857APending Publication Date: 2026-11-17GUIZHOU POWER GRID CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610945656.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing load redistribution attack detection methods are inefficient when faced with carefully constructed anomalous data from individual sensors, and require a large number of labeled attack samples for supervised training, which limits their adaptability and practicality.

Method used

By acquiring real-time measurement data sequences from a single sensor, calculating the difference value to construct a hysteresis vector and forming a measurement matrix, performing multi-level decomposition to extract feature vectors, constructing a projection matrix, calculating the projection distance and minimum alignment distance, constructing a comprehensive anomaly index, and combining it with a dynamic time warping algorithm for detection.

Benefits of technology

It improves the detection efficiency and accuracy of various types of load redistribution attacks, reduces false alarm and false negative rates, effectively identifies anomalies caused by data amplitude and sequence tampering, and adapts to different types of attacks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122857A_ABST
    Figure CN123122857A_ABST
Patent Text Reader

Abstract

This application relates to the field of network security technology and provides a method, application method, device, and medium for detecting power grid load redistribution attacks. The method includes: focusing on time-series data from a single sensor, converting real-time measurements into a differential sequence reflecting local dynamics and constructing a measurement matrix; using a multi-level decomposition based on elimination trees to extract feature vectors layer by layer from the measurement matrix and constructing a projection matrix; using the projection distance between the current lag vector and the historical predicted vector in the projection subspace to capture abnormal deviations in the data structure; constructing a multi-dimensional change spectrum composed of the projection distance and the rate of change of feature vectors at each level; calculating the minimum alignment distance between the spectrum and the historical normal change spectrum using a dynamic time warping algorithm to detect abnormal changes in the time-series pattern; and weightedly fusing the projection distance and the alignment distance to form a comprehensive anomaly index, effectively improving the detection efficiency and accuracy of various types of load redistribution attacks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of network security technology, and in particular to a method, application method, device and medium for detecting power grid load redistribution attacks. Background Technology

[0002] Load redistribution attacks are a typical type of spoofed data injection attack. Attackers tamper with sensor measurement data deployed at the user end, circumventing traditional bad data detection mechanisms and inducing the power grid control center to make incorrect scheduling decisions, resulting in significant economic losses. Existing load redistribution attack detection methods are typically based on system-level global data, utilizing data-driven techniques such as state estimation residual analysis, principal component analysis, or density-based clustering algorithms to extract overall abnormal patterns from the measurements of all sensors. Then, they compare residuals, reconstruction errors, or anomaly scores with preset thresholds to determine whether the system has been attacked.

[0003] When attackers meticulously construct measurements from individual sensors, their anomalies are easily masked by macroscopic statistical features at the system level. This results in existing methods being inefficient at detecting sophisticated, enhanced load redistribution attacks that are highly stealthy and aimed at increasing power generation costs. Furthermore, existing data-driven detectors typically require supervised training with labeled attack samples, while advanced attack samples are extremely difficult to obtain in real-world systems, further limiting their adaptability and practicality. Summary of the Invention

[0004] This application provides a method, application method, device and medium for detecting power grid load redistribution attacks, which solves the problem of low detection efficiency for multiple types of load redistribution attacks.

[0005] The first aspect of this application provides a method for detecting power grid load redistribution attacks, including: Acquire a real-time measurement data sequence from a single sensor, calculate the difference between the measurements at two adjacent time points in the sequence, construct a hysteresis vector of a specified length from multiple consecutive difference values, and form a measurement matrix; The measurement matrix is ​​decomposed into multiple levels, and a projection matrix is ​​constructed based on the feature vectors of each level obtained from the decomposition. The projection distance between the current lag vector and the historical prediction vector under the projection matrix is ​​calculated. Construct the current change spectrum, which is a multi-dimensional sequence composed of the projection distance and the rate of change of each layer feature vector; calculate the minimum alignment distance between the current change spectrum and the historical normal change spectrum pre-extracted from the normal data in the training phase using a dynamic time warping algorithm; The projection distance and the minimum alignment distance are weighted and summed to obtain a comprehensive anomaly index; The magnitude of the comprehensive anomaly index is compared with the preset threshold. If the magnitude exceeds the preset threshold, it is determined that the sensor has suffered a load redistribution attack.

[0006] Optionally, the step of performing multi-level decomposition on the measurement matrix and constructing a projection matrix based on the feature vectors obtained from each level of decomposition includes: Obtain the elimination tree corresponding to the measurement matrix. The elimination tree contains tree nodes at multiple levels, and each tree node corresponds to a sub-matrix in the measurement matrix. Singular value decomposition is performed on the submatrix corresponding to each tree node to extract the top features representing the main change patterns of the submatrix. A singular vector is used as the feature vector corresponding to the tree node; The feature vectors corresponding to each tree node at the same level are concatenated to obtain the level feature vector of that level. The feature vectors of each level are combined to construct the projection matrix.

[0007] Optionally, calculating the projection distance between the current lag vector and the historical prediction vector under the projection matrix includes: Select multiple historical lag vectors that are closest to the current time, and perform an exponentially weighted average of each historical lag vector according to its time proximity to obtain the historical prediction vector. Project the current lag vector and the historical prediction vector onto the subspace spanned by the column vectors of the projection matrix, respectively, to obtain the current projection vector and the predicted projection vector. Calculate the Euclidean distance between the current projection vector and the predicted projection vector, and use it as the projection distance.

[0008] Optionally, the step of calculating the minimum alignment distance between the current variation spectrum and the historical normal variation spectrum pre-extracted from the normal data in the training phase using a dynamic time warping algorithm includes: Based on the projection distance and the rate of change of each layer feature vector at adjacent time points, a multi-dimensional sequence is constructed to obtain the current change spectrum; Acquire the historical normal variation spectrum pre-extracted from normal data in the same manner during the training phase; The current variation spectrum and the historical normal variation spectrum are nonlinearly aligned on the time axis using a dynamic time warping algorithm, and the minimum cumulative distance between the two is calculated as the minimum alignment distance.

[0009] A second aspect of this application provides an application of the detection method described above in identifying multiple types of load redistribution attacks, including general load redistribution attacks, covert load redistribution attacks, and enhanced load redistribution attacks.

[0010] Optionally, the general load redistribution attack is established through the following steps: Obtain the net injected power vector of the power grid and construct a DC power flow model that characterizes the relationship between the net injected power vector and the power flow vector; Generate an injected load measurement vector, wherein the algebraic sum of each element in the injected load measurement vector is zero, and the absolute value of each element does not exceed a preset proportion of the corresponding normal load measurement value. The injected load measurement vector is superimposed on the net injected power vector, the injected power flow vector is calculated according to the DC power flow model, and the change in the power generation vector is set to zero.

[0011] Optionally, the stealthy load redistribution attack is established through the following steps: Multiple historical measurement vectors are obtained, and the first objective function is to minimize the sum of the Euclidean distances between the currently injected measurement vector and each of the historical measurement vectors. Under the constraints of the general load redistribution attack, the first objective function is optimized to obtain the injected load measurement vector and power flow vector; The injected load measurement vector and power flow vector obtained by the optimization solution also satisfy the following: the ratio of the power flow change of at least one transmission line to the maximum power flow capacity exceeds the preset power flow over-limit threshold.

[0012] Optionally, the enhanced load redistribution attack is established through the following steps: The goal is to obtain the power generation cost under normal conditions, and use the difference between the power generation cost calculated after injecting erroneous data and the normal power generation cost as the second objective function. Under the constraints of the general load redistribution attack or the covert load redistribution attack, the second objective function is maximized to obtain the injected load measurement vector and power flow vector, such that the power generation cost after injection is higher than the normal power generation cost.

[0013] A third aspect of this application provides a computer device, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for detecting power grid load redistribution attacks.

[0014] A fourth aspect of this application provides a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the aforementioned method for detecting power grid load redistribution attacks.

[0015] As can be seen from the above technical solutions, this application has the following advantages: This application enhances the sensitivity to microscopic changes in single-point data by focusing on time-series data from a single sensor, transforming real-time measurements into differential sequences reflecting local dynamics and constructing a measurement matrix. It employs a multi-level decomposition based on elimination trees to extract feature vectors layer by layer from the measurement matrix and construct a projection matrix, achieving a hierarchical characterization of the main data change patterns and improving the accuracy of feature extraction. By utilizing the projection distance between the current hysteresis vector and the historical prediction vector in the projection subspace, it captures abnormal deviations in the data structure, effectively identifying distortions in the measurement value structure caused by attacks. Simultaneously, it constructs a multi-dimensional change spectrum composed of the projection distance and the rate of change of feature vectors at each level. Using a dynamic time warping algorithm, it calculates the minimum alignment distance with the historical normal change spectrum to detect abnormal changes in the time-series pattern, enabling the discovery of covert actions by attackers to rearrange data order. Finally, it weights and fuses the projection distance and alignment distance to form a comprehensive anomaly index, achieving dual verification of structural and time-series features, effectively reducing false alarm and false negative rates. This allows it to distinguish between normal load fluctuations and malicious attacks, significantly improving the detection efficiency and accuracy for various types of load redistribution attacks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for detecting a power grid load redistribution attack according to this application. Figure 2 This is a schematic diagram of the construction process of the projection matrix in this application; Figure 3 This is a schematic diagram illustrating the calculation process of the projection distance in this application; Figure 4 This is a schematic diagram illustrating the process of constructing the variation spectrum and calculating the alignment distance in this application; Figure 5 This graph illustrates the increase in power generation costs caused by the enhanced load redistribution attack in this application. Figure 6 A comparison chart showing the injection errors of general load redistribution attacks and stealthy load redistribution attacks in this application; Figure 7 A comparison chart showing the injection errors of covert load redistribution attacks and enhanced load redistribution attacks in this application. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] This application provides a method for detecting power grid load redistribution attacks, applied to a power grid monitoring system. This system includes load measurement sensors deployed at the user end, data acquisition and communication devices, and a monitoring server located at the power grid control center. Specifically, the method executes the following detection process through the monitoring server: the data acquisition and communication devices acquire load measurement data from individual sensors in real time and transmit it to the monitoring server; the monitoring server calculates the difference between adjacent time points on the received real-time measurement data sequence, constructs a hysteresis vector, and forms a measurement matrix; the monitoring server's processor executes a multi-level decomposition algorithm based on a elimination tree to perform singular value decomposition on each level of the sub-matrix to extract feature vectors and generate a projection matrix; the monitoring server calls pre-stored historical normal data in its memory, calculates the projection distance and multi-dimensional change spectrum, and compares the current change spectrum with the historical normal change spectrum using a dynamic time warping algorithm to obtain the minimum alignment distance; the monitoring server's processor weightedly fuses the projection distance and alignment distance into a comprehensive anomaly index, compares it with a preset threshold, and outputs an attack alarm signal when the comprehensive anomaly index exceeds the threshold. This method is implemented based on the actual hardware equipment operating in the power grid monitoring system and has clear feasibility.

[0019] Please see Figure 1 This application provides a method for detecting power grid load redistribution attacks, comprising the following steps: S1. Obtain the real-time measurement data sequence of a single sensor, calculate the difference between the measurement values ​​at two adjacent time points in the sequence, construct a hysteresis vector of a specified length from multiple consecutive difference values, and form a measurement matrix; A single sensor refers to a physical sensor deployed at the user end of a power system to measure load power or current. It continuously acquires and outputs measured values ​​according to a preset sampling period. This step involves acquiring a real-time measurement data sequence from the sensor through a data acquisition and communication device. This data sequence consists of continuous measured values ​​arranged in chronological order, and can be represented as: in: Indicates in Sensor measurements acquired at all times For time indexing.

[0020] Since load redistribution attacks typically manifest as anomalous jumps in measured values ​​between adjacent time points, directly using the original measured values ​​is insufficient to effectively capture such local dynamic changes. Therefore, the aforementioned measurement data sequence is differentially processed, calculating the difference between measured values ​​at two adjacent time points to obtain a difference sequence. Each difference value can be expressed as: in: Indicates the first Each difference value reflects the sensor measurement value from... Time's up The magnitude of change over time. Through differential processing, the slow trend components in the original measurement data are suppressed, while the short-term rapid changes caused by the attack are highlighted, thereby enhancing the sensitivity to local anomalies.

[0021] After obtaining the difference sequence, lag vectors are constructed based on the difference values. This specified length refers to the number of difference values ​​contained in each lag vector, denoted as . Its value is a positive integer preset based on the sensor sampling frequency and attack duration characteristics. For example, when the sampling period is on the order of seconds and the attack duration is on the order of minutes... The typical value range is 10 to 50. A length of... The hysteresis vector is composed of continuous The difference values ​​are arranged sequentially, indexed by time. The lag vector starting from is represented as: in: For the first There are lag vectors, with dimensions of . It depicts the process from Start time, length is The local dynamic pattern of measurements within a time window.

[0022] Furthermore, a sliding window approach is used to sequentially select from the difference sequence. A series of continuous lag vectors form the measurement matrix. At the current time... Using the reference point, select the lag vector. , ,…, Arranged by column, the constituent dimensions are Measurement matrix : in: The number of selected lag vectors, whose values ​​are positive integers preset based on system computing resources and real-time requirements, for example... The typical value range for the above measurement matrix is ​​5 to 20. In the middle, the first column to the... The columns correspond to the lag vectors respectively. to With the first Taking the last column as an example, this column corresponds to the current time. Lag vector starting from Its elements are arranged from top to bottom as follows: From time Start continuous Each column represents a difference value; similarly, the other columns correspond to consecutive values ​​from their respective time starting points. Each difference value. Measurement matrix. Each column is a hysteresis vector, and each row is the difference value at the same relative position under different time windows, which fully represents the change pattern of sensor measurement data at the current moment and in the adjacent historical time period.

[0023] S2. Perform multi-level decomposition on the measurement matrix, construct a projection matrix based on the feature vectors of each level obtained from the decomposition, and calculate the projection distance between the current lag vector and the historical prediction vector under the projection matrix. In this embodiment, multi-level decomposition is based on the elimination tree corresponding to the measurement matrix, with the feature vectors of each level corresponding to tree nodes in the elimination tree. This step constructs an elimination tree for the measurement matrix, performing multi-granularity decomposition of the measurement data in a tree-like hierarchical structure to extract the main change patterns at different scales. Based on this, a projection matrix capable of characterizing the normal behavior contours of the data is constructed. Then, the projection distance is used to quantify the structural deviation between the current measurement data and historical normal patterns, providing a first-dimensional feature for anomaly detection. Please refer to [link to previous text]. Figure 2 The process of constructing the projection matrix is ​​as follows: S211. Obtain the elimination tree corresponding to the measurement matrix. The elimination tree contains tree nodes at multiple levels, and each tree node corresponds to a sub-matrix in the measurement matrix. S212. Perform singular value decomposition on the submatrix corresponding to each tree node, and extract the prefixes representing the main change patterns of the submatrix. A singular vector is used as the feature vector corresponding to the tree node; S213. Concatenate the feature vectors corresponding to each tree node at the same level to obtain the level feature vector of that level; S214. Combine the feature vectors of each level to construct the projection matrix.

[0024] Specifically, an elimination tree is a tree-like data structure generated based on the sparse or correlated structure of a measurement matrix. Its root node corresponds to the entire measurement matrix, leaf nodes correspond to indivisible sub-blocks within the matrix, and intermediate nodes correspond to sub-matrices of different granularities. The method for constructing an elimination tree includes: first, calculating the correlation coefficients between columns of the measurement matrix; merging columns with correlations higher than a set threshold into a supernode; and recursively merging columns until a hierarchical tree is formed. Alternatively, the elimination tree can be generated using a graph Laplacian matrix corresponding to the measurement matrix through nested partitioning. Each tree node in the elimination tree corresponds to a sub-matrix in the measurement matrix, which consists of all columns (or rows) belonging to that node and its subtrees. Different levels of tree nodes correspond to different scales of data partitioning; higher-level nodes correspond to larger-scale sub-matrices, reflecting global change patterns; lower-level nodes correspond to smaller-scale sub-matrices, capturing local detailed features.

[0025] For each node in the eliminated tree, let its corresponding submatrix be denoted as . , dimension .in This is the number of columns contained in that node. Perform singular value decomposition: in: for The left singular vector matrix, for The right singular vector matrix, for A diagonal matrix whose diagonal elements are singular values. The magnitude of the singular values ​​reflects the energy proportion of the pattern represented by the corresponding singular vector in the submatrix; the first few singular vectors can capture the main change patterns of the submatrix.

[0026] To preserve dominant features and filter out noise, before extraction The left singular vectors are used as the feature vectors of this tree node, where The value of can be determined based on the cumulative energy ratio, and a value that satisfies is selected. The smallest integer as The value of is taken as follows. Let the extracted feature vector matrix be . Its column vector is the feature vector corresponding to the tree node.

[0027] After extracting features from all tree nodes, for all nodes at the same level, their corresponding feature vector matrices are concatenated column-wise to form the level-specific feature vector matrix. ,in Indicates a hierarchical index. This represents the number of nodes at this level. Next, the feature vector matrices of each level are combined column-wise, and then either filtered by principal components or directly merged to construct the final projection matrix. Each column of the projection matrix represents a normal change pattern basis vector extracted from the measurement data, and the subspace spanned by the projection matrix is ​​a low-dimensional representation of the change pattern of sensor measurement data during normal operation of the power grid.

[0028] Please see Figure 3 The calculation process for the projected distance is as follows: S221. Select multiple historical lag vectors that are closest to the current time, and perform an exponentially weighted average of each historical lag vector according to its time proximity to obtain the historical prediction vector. S222. Project the current lag vector and the historical prediction vector onto the subspace spanned by the column vectors of the projection matrix, respectively, to obtain the current projection vector and the predicted projection vector; S223. Calculate the Euclidean distance between the current projection vector and the predicted projection vector, and use it as the projection distance.

[0029] Specifically, let's assume the current time is The current lag vector has been constructed. (obtained from S1) To obtain a prediction vector that represents the recent normal trend of change, the vector closest to the current time is selected. Historical lag vectors Since data closer to the current moment contributes more to the prediction, an exponentially weighted average is used to generate the historical prediction vector: in: This is the attenuation coefficient, which controls the rate at which the weights decay. The larger the value, the more emphasis is placed on data from the most recent moment. The denominator is a normalization factor to ensure that the sum of the weights is 1.

[0030] Then, calculate the current hysteresis vector. Compared with historical prediction vectors In the projection matrix Projection into Zhang's subspace. The projection operation is achieved by multiplying a vector by a projection matrix: in: and These are the current projection vector and the predicted projection vector, located in the... In the low-dimensional subspace spanned by the column vectors, this subspace captures the main directions of change in normal data.

[0031] Finally, the Euclidean distance between the two projection vectors is calculated as the projection distance. : in: The dimension of the projected subspace is equal to the projection matrix. The number of columns. Projection distance. The deviation of the current measurement data from the normal variation pattern subspace was quantified: when the sensor is subjected to a load redistribution attack, the injected erroneous data causes the structural profile of the current hysteresis vector to deviate from the normal pattern, resulting in a significant increase in the projection distance. This distance will serve as an important component of the subsequent comprehensive anomaly index.

[0032] S3. Construct the current change spectrum, which is a multi-dimensional sequence composed of the projection distance and the rate of change of the feature vectors at each level; calculate the minimum alignment distance between the current change spectrum and the historical normal change spectrum pre-extracted from the normal data in the training phase using the dynamic time warping algorithm; The current change spectrum refers to a multi-dimensional time series constructed based on the current time, integrating the projected distance and the rate of change of feature vectors at each level over adjacent time steps. It characterizes the temporal evolution of sensor measurement data in terms of both structural profile and hierarchical dynamics. In actual attacks, attackers may not only tamper with the magnitude of measured values ​​but also bypass anomaly detection based on single points or single distances by rearranging the order of data segments. Therefore, relying solely on projected distance is insufficient to effectively identify such temporal anomalies. Here, a dynamic time warping algorithm is used to non-linearly align the current change spectrum with historical normal change spectra to quantify the differences in their temporal morphology, thus providing a second dimension of feature for detecting temporal sequence anomalies. Please refer to [link to relevant documentation]. Figure 4 The process of constructing the variation spectrum and calculating the alignment distance is as follows: S31. Based on the projection distance and the rate of change of the feature vectors of each level at adjacent time intervals, construct a multi-dimensional sequence to obtain the current change spectrum; S32. Obtain the historical normal variation spectrum pre-extracted from normal data in the same manner during the training phase; S33. Using a dynamic time warping algorithm, nonlinear alignment is performed on the time axis between the current variation spectrum and the historical normal variation spectrum, and the minimum cumulative distance between the two is calculated as the minimum alignment distance.

[0033] Specifically, let the current time be denoted as The projected distance corresponding to the current hysteresis vector has been calculated in S2. And the feature vectors of each level. For each level Hierarchical feature vector matrix Take the former Given _n_ principal eigenvectors, calculate the rate of change of each eigenvector at adjacent time points. Let the _n_ principal eigenvector be _n_. The first level The projection coefficients of the eigenvectors at the current time are: The projection coefficients at the previous time step were Then the rate of change of the eigenvector can be defined as: Where: projection coefficient It is obtained by projecting the current hysteresis vector onto the corresponding feature vector.

[0034] By combining the projection distance with the rate of change of the feature vectors at each level, a multi-dimensional change vector at the current moment is constructed. : This change vector integrates two types of information: structural anomalies (projected distance) and hierarchical dynamic anomalies (rates of change of each eigenvector). To characterize the temporal evolution features over a period of time, the current time is used as the starting point. As the endpoint, take a length of The time window is used to arrange the change vectors at each moment within the window in chronological order, forming the current change spectrum. : Current change spectrum It is a dimension , in which The dimension of the change vector. The length of the spectrum is denoted by . This sequence records the continuous trajectory of sensor measurement data across structural profiles and dynamic characteristics at various levels.

[0035] During the training phase, a segment of normal sensor time-series data, verified manually or confirmed by system records to have not been attacked, is selected. Following the exact same method described above, a historical normal variation spectrum is pre-constructed and stored. This historical normal variation spectrum serves as a reference benchmark, characterizing the typical shape and fluctuation range of sensor data variation spectrum under normal operating conditions.

[0036] To compare the current change spectrum Compared with the historical normal change spectrum To determine the similarity, a dynamic time warping algorithm is used for non-linear alignment. A space of size is constructed. Cumulative distance matrix Its elements Indicates the current change spectrum number The first change vector and the historical normal change spectrum The minimum cumulative distance between the changing vectors. The recursive calculation process is as follows: in: The first in the current change spectrum The change vector and the in the historical normal change spectrum The Euclidean distance between the transformed vectors. Initialization conditions are... Boundary values and Calculated by accumulation.

[0037] After a complete recursive calculation, the minimum alignment distance is obtained. The smaller this distance value, the more similar the current variation spectrum is to the historical normal variation spectrum in terms of temporal pattern, meaning the lower the probability that the sensor's measurement data has been rearranged or inserted with abnormal segments. Conversely, if this distance increases significantly, it indicates that the current data's temporal evolution pattern has deviated from the normal trajectory, and it is highly likely to have suffered a load redistribution attack, especially those covert attacks that circumvent single-point detection by rearranging data order. This minimum alignment distance will serve as the second dimension of the subsequent comprehensive anomaly index.

[0038] S4. Sum the projected distance and the minimum alignment distance in a weighted manner to obtain the comprehensive anomaly index; S5. Compare the comprehensive abnormal indicators with the preset threshold. If the value exceeds the preset threshold, it is determined that the sensor has been subjected to a load redistribution attack.

[0039] The projection distance is calculated using S2 and S3 respectively. and minimum alignment distance Subsequently, these two indicators are weighted and fused to construct a comprehensive anomaly index that can comprehensively reflect the degree of data anomaly. Projection distance measures the degree of structural deviation of the current measurement data within a subspace constructed from normal data, focusing on detecting structural contour anomalies caused by data amplitude tampering. Minimum alignment distance measures the degree of difference in temporal morphology between the current variation spectrum and historical normal variation spectra, focusing on detecting temporal evolution anomalies caused by data order reversal. The two are complementary in their detection logic; the former is sensitive to numerical tampering, and the latter is sensitive to order reversal. By weighting and summing the two, dual verification of load redistribution attacks can be achieved, effectively reducing false positives or false negatives that might result from a single indicator.

[0040] Comprehensive abnormal indicators The calculation formula is: in: and These are weighting coefficients used to adjust the contribution ratio of projection distance and minimum alignment distance to the overall anomaly index. The value of the weighting coefficient can be configured based on the frequency and severity of the two types of attacks. In a typical configuration, when numerical manipulation attacks are the primary threat to the system, this can be set... When timing-swapping attacks are the primary threat, settings can be configured. By default, it is preferable to Both contribute equally. The weighting coefficients can also be determined by statistically analyzing historical normal data during the training phase, with the optimization objective being to minimize the fluctuation of the overall abnormal index of normal data.

[0041] The preset threshold set here This threshold is a judgment boundary determined based on historical normal data sequences during the training phase. The method for determining the threshold is as follows: During the training phase, select a segment of normal sensor time-series data that has been manually verified or confirmed by system records to have not suffered any attacks. According to steps S1 to S4, calculate the comprehensive anomaly index of this segment of normal data at each time point, and record the maximum value. Considering the seasonal or periodic fluctuations in load under normal operating conditions, and to allow for a reasonable margin, at the maximum value... Multiply by a margin factor The preset threshold is obtained: Among them, the margin coefficient The value is set according to the system's tolerance for false alarm rate and false negative rate, with a typical range of 1.1 to 1.5. The smaller the value, the closer the threshold is to the historical maximum value, the higher the sensitivity to anomalies, and the lower the false negative rate, but the probability of normal fluctuations being misjudged as attacks increases. A larger value results in a more lenient threshold and a lower false alarm rate, but it also reduces detection sensitivity and may cause some attacks to go undetected. In practical applications, the threshold can be flexibly adjusted according to the security level requirements of the power grid monitoring system.

[0042] During the real-time detection phase, a comprehensive anomaly index is calculated for each current moment. Then, compare it with a preset threshold. The data is compared. When the comprehensive anomaly index is less than or equal to the preset threshold, it indicates that the current measurement data is consistent with the normal pattern in both structural profile and temporal morphology, and its deviation is within the historical normal fluctuation range. Therefore, it is determined that the sensor has not been subjected to a load redistribution attack, and the data can be normally entered into the subsequent power grid state estimation and scheduling decision-making process. When the comprehensive anomaly index is greater than the preset threshold, it indicates that the current measurement data has an abnormal deviation from the historical normal fluctuation range in terms of structural profile or temporal morphology. It is highly likely that an attacker has injected false load measurement data, so it is determined that the sensor has been subjected to a load redistribution attack. At this time, the monitoring server can trigger an alarm signal to notify the power grid operation and maintenance personnel to take corresponding security response measures, such as marking the sensor's data as suspicious, switching to a backup measurement data source, or initiating further attack type identification and attack source location processes.

[0043] This application provides an application of the detection method described above in identifying multiple types of load redistribution attacks. Based on the detection method described in any of the preceding embodiments, this application performs real-time detection on sensor time-series data and outputs a comprehensive anomaly index. When the comprehensive anomaly index exceeds a preset threshold, it not only determines that the sensor has suffered a load redistribution attack, but also further identifies the attack type based on the mathematical characteristics of the attack. These multiple types of load redistribution attacks include general load redistribution attacks, stealthy load redistribution attacks, and enhanced load redistribution attacks. The three types of attacks increase sequentially in terms of construction complexity, stealth, and attack effectiveness. The above detection method has effective detection capabilities for all three types of attacks. The establishment process for each of the three types of attacks is described in detail below: Furthermore, a general load redistribution attack is established through the following steps: 1. Obtain the net injected power vector of the power grid and construct a DC power flow model that characterizes the relationship between the net injected power vector and the power flow vector; 2. Generate the injected load measurement vector. The algebraic sum of all elements in the injected load measurement vector is zero, and the absolute value of each element does not exceed the preset ratio of the corresponding normal load measurement value. 3. Superimpose the injected load measurement vector onto the net injected power vector, calculate the injected power flow vector according to the DC power flow model, and set the change in the power generation vector to zero.

[0044] Specifically, the general load redistribution attack is the most basic type of spoofing attack. Its goal is to inject specific erroneous values ​​into sensor measurement data, so that the injected measurement data can pass the bad data detection mechanism of the power grid control center based on the DC power flow model, thereby misleading the control center to make incorrect scheduling decisions.

[0045] Obtain the net injected power vector under the current operating state of the power grid ,in Let be the number of grid nodes, and let the net injected power vector represent the difference between the generating power and the load power at each node. A DC power flow model is constructed, which describes the net injected power vector. With transmission line power flow vector Linear relationship between them: in: It is a susceptance diagonal matrix, where the diagonal elements are the susceptance values ​​of each line; This is a node-branch correlation matrix, describing the connection relationships between each line and node; The nodal admittance matrix; For power flow vector, This refers to the number of power transmission lines.

[0046] Secondly, generate the injected load measurement vector. To ensure that the injected attack data passes bad data detection, two basic constraints must be met. Constraint 1: The algebraic sum of all elements in the injected load measurement vector is zero, i.e.: in: This is a row vector consisting entirely of 1s. This constraint ensures that the injected spurious load data cancel each other out in total, preventing significant changes in the balance between power generation and consumption, thus circumventing power balance-based detection. Constraint 2: The absolute value of each injected load measurement does not exceed a preset proportion of the corresponding normal load measurement, i.e.: in: For the first Load measurement values ​​injected by each node. This is the load measurement value under normal conditions for this node. This is the preset scaling parameter. Preset scaling Based on the range of injection amplitude that the attacker wishes to control: The larger the value, the greater the range of erroneous data that an attacker can inject, and the greater the impact on the system, but the risk of being detected also increases accordingly. The smaller the value, the more covert the attack, but the impact on the system may not be sufficient to cause substantial damage. In real-world attack scenarios, It is usually set to a value between 0.1 and 0.5.

[0047] Next, the generated injection load measurement vector will be... Superimposed on the original net injected power vector, the net injected power vector after the attack is obtained. Based on the DC power flow model, calculate the injected power flow vector: At the same time, to maintain the stealth of the attack, the change in the power generation vector is set to zero, that is... This makes the power generation vector after the attack... To maintain consistency with the original power generation, thereby avoiding anomalies in the power generation data.

[0048] The general load redistribution attack established through the above steps generates an attack vector containing spoofed load data. ,in This is the payload measurement vector after the attack. This attack vector can be injected into the sensor measurement data deployed at the user end to test and verify the effectiveness of the aforementioned detection method, and to verify whether the detection method can still detect anomalies through single-sensor time series analysis even when the attack meets the above constraints.

[0049] Furthermore, a stealthy load redistribution attack is established through the following steps: 1. Obtain multiple historical measurement vectors, and minimize the sum of the Euclidean distances between the currently injected measurement vector and each historical measurement vector as the first objective function; 2. Under the constraints of the general load redistribution attack, the first objective function is optimized to obtain the injected load measurement vector and power flow vector.

[0050] Specifically, the stealthy load redistribution attack is based on the general load redistribution attack, and further considers the stealth of the attack. Its goal is to make the measurement vector after injecting erroneous data as close as possible to multiple historical normal measurement vectors in the data space, thereby circumventing data-driven detection methods based on clustering or distance anomalies.

[0051] First, obtain the length as Historical measurement vector sequence These historical vectors are normal measurement data selected from the power grid historical database that are close to the current time. Each vector contains the load measurement value and power flow value at the corresponding time. A first objective function is constructed, which is the measurement vector after the current erroneous data injection. The sum of Euclidean distances between the vectors and each historical measurement vector: The physical meaning of this objective function is to make the post-attack measurement vector as close as possible to the central region of multiple historical normal data points in Euclidean space. When the attack vector is highly similar to the historical normal data, data-driven clustering anomaly detectors will have difficulty distinguishing it from the normal data.

[0052] The optimization solution is performed under the constraints of a general load redistribution attack. The optimization process needs to satisfy the following constraints simultaneously: the algebraic sum of all elements in the injected load measurement vector is zero; the absolute value of each injected value does not exceed a preset proportion of the normal load; the change in the power generation vector is zero; and the power flow vector after injection conforms to the DC power flow model.

[0053] In addition to satisfying the above constraints, the optimization solution also requires that: there exists at least one transmission line whose power flow change ratio exceeds the preset power flow over-limit threshold, i.e.: in: For the first The change in power flow of the transmission line This represents the maximum power flow capacity of the line. This is a preset threshold parameter. It should be noted that this preset power flow over-limit threshold is... This is a constraint parameter during the attack construction phase. Its function is to ensure that the erroneous data injected by the attack causes sufficiently significant power flow fluctuations on at least one line, thereby giving the attack the potential to affect the operation of the power grid. This parameter The preset threshold used in the detection method described in S5 differs from the one used in the above-mentioned method in terms of definition, purpose, and determination: It is the power flow threshold set by the attacker when constructing the attack. It is the boundary for determining anomalies, which is determined by the testing party based on statistical analysis of historical normal data.

[0054] Furthermore, the enhanced load redistribution attack is established through the following steps: 1. Obtain the power generation cost under normal grid conditions, and use the difference between the power generation cost calculated after injecting erroneous data and the normal power generation cost as the second objective function; 2. Under the constraints of general load redistribution attack or covert load redistribution attack, the second objective function is maximized to obtain the injected load measurement vector and power flow vector, so that the power generation cost after injection is higher than the normal power generation cost.

[0055] Specifically, the enhanced load redistribution attack, based on the two types of attacks mentioned above, further introduces an optimization objective for attack effectiveness. Its core purpose is to ensure that the attack can not only evade detection but also actually cause economic losses to the power grid and significantly increase the cost of power generation.

[0056] First, obtain the power generation cost of the power grid under normal conditions. The cost of electricity generation is usually expressed as the sum of quadratic functions of the power output of each generator: in: For the number of generators, For the first The power output of the generator, , , For the first Cost coefficient of a generator. Power generation cost under normal conditions. It is the minimum cost corresponding to the output of each generator obtained through economic dispatch optimization under the conditions of satisfying load demand and line capacity constraints: in: and These are the lower and upper limits of power generation, respectively. This is the vector representing the maximum power flow capacity of the line.

[0057] Secondly, a second objective function is constructed, which is the power generation cost calculated after injecting erroneous data. Compared with normal power generation costs Difference: in: The power generation cost obtained by solving the economic scheduling problem after the attack: The physical meaning of this objective function is that by injecting false load data, the power grid control center will perform economic dispatch based on erroneous data, resulting in a power generation cost that is significantly higher than the normal state, thus causing economic losses.

[0058] The optimization solution is performed under the constraints of general load redistribution attacks or stealth load redistribution attacks. When solving under the constraints of general attacks, the enhanced attack balances basic stealth and economic disruption; when solving under the constraints of stealth attacks, the enhanced attack simultaneously possesses data-level stealth, physical-level power flow impact, and economic disruption, making it the most challenging attack form for detection methods.

[0059] By maximizing the solution, the injected load measurement vector and power flow vector are obtained. After being injected into the sensor data, this attack vector will cause the grid control center to re-perform economic dispatch based on false load data. The dispatch result increases the actual power generation cost, thus causing economic losses to the grid operator. This enhanced load redistribution attack model is used to verify whether the aforementioned detection method can effectively identify the sensor data anomalies caused by attacks with the explicit goal of increasing power generation costs through a dual verification mechanism of projection distance and minimum alignment distance, thereby providing comprehensive technical support for grid security protection.

[0060] Based on the detection methods and the establishment process of the three types of load redistribution attacks described in detail in the foregoing embodiments, the effectiveness of the detection method of this application is verified through specific experiments below. The attack effects of general load redistribution attacks, stealthy load redistribution attacks, and enhanced load redistribution attacks are compared under different parameter conditions to verify the superiority of enhanced load redistribution attacks in causing economic losses.

[0061] A power grid model was constructed on a power system simulation platform for experiments. Load data was derived from actual measurements of a regional power grid. The experiments tested the real-world impact of general load redistribution attacks, covert load redistribution attacks, and enhanced load redistribution attacks on the system's generation costs. Please refer to [link / reference]. Figure 5 , Figure 6 and Figure 7 Experimental results show that, under the same preset proportional parameters, only enhanced load redistribution attacks can significantly increase power generation costs, with a single successful enhanced attack causing power generation costs to surge by nearly $60 per megawatt-hour. Further comparative results indicate that whether stealthy load redistribution attacks can increase power generation costs is not directly related to the magnitude of the injection error or the attack mode. These experiments collectively demonstrate that the key factor determining the effectiveness of an attack lies in whether the construction of the attack vector aims to increase power generation costs, thus validating the necessity of researching enhanced load redistribution attacks.

[0062] Furthermore, this application also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for detecting power grid load redistribution attacks. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0063] Furthermore, embodiments of this application also provide a computer-readable storage medium storing computer instructions that, when executed on a computer device, cause the computer device to perform the aforementioned detection method for power grid load redistribution attacks. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0064] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0065] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method of detecting a grid load redistribution attack, the method comprising: include: Acquire a real-time measurement data sequence from a single sensor, calculate the difference between the measurements at two adjacent time points in the sequence, construct a hysteresis vector of a specified length from multiple consecutive difference values, and form a measurement matrix; The measurement matrix is ​​decomposed into multiple levels, and a projection matrix is ​​constructed based on the feature vectors of each level obtained from the decomposition. The projection distance between the current lag vector and the historical prediction vector under the projection matrix is ​​calculated. Construct the current change spectrum, which is a multi-dimensional sequence composed of the projection distance and the rate of change of each layer feature vector; calculate the minimum alignment distance between the current change spectrum and the historical normal change spectrum pre-extracted from the normal data in the training phase using a dynamic time warping algorithm; The projection distance and the minimum alignment distance are weighted and summed to obtain a comprehensive anomaly index; The magnitude of the comprehensive anomaly index is compared with the preset threshold. If the magnitude exceeds the preset threshold, it is determined that the sensor has suffered a load redistribution attack.

2. The method of claim 1, wherein, The step of performing multi-level decomposition on the measurement matrix and constructing a projection matrix based on the feature vectors obtained from each level of decomposition includes: Obtain the elimination tree corresponding to the measurement matrix. The elimination tree contains tree nodes at multiple levels, and each tree node corresponds to a sub-matrix in the measurement matrix. performing singular value decomposition on each sub-matrix corresponding to the tree node to extract the first singular vectors representing the main variation mode of the sub-matrix as the feature vector corresponding to the tree node ; The feature vectors corresponding to each tree node at the same level are concatenated to obtain the level feature vector of that level. The feature vectors of each level are combined to construct the projection matrix.

3. The method of claim 2, wherein, The calculation of the projection distance between the current lag vector and the historical prediction vector under the projection matrix includes: Select multiple historical lag vectors that are closest to the current time, and perform an exponentially weighted average of each historical lag vector according to its time proximity to obtain the historical prediction vector. Project the current lag vector and the historical prediction vector onto the subspace spanned by the column vectors of the projection matrix, respectively, to obtain the current projection vector and the predicted projection vector. Calculate the Euclidean distance between the current projection vector and the predicted projection vector, and use it as the projection distance.

4. The method of claim 1, wherein, The step of calculating the minimum alignment distance between the current variation spectrum and the historical normal variation spectrum pre-extracted from the normal data in the training phase using a dynamic time warping algorithm includes: Based on the projection distance and the rate of change of each layer feature vector at adjacent time points, a multi-dimensional sequence is constructed to obtain the current change spectrum; Acquire the historical normal variation spectrum pre-extracted from normal data in the same manner during the training phase; The current variation spectrum and the historical normal variation spectrum are nonlinearly aligned on the time axis using a dynamic time warping algorithm, and the minimum cumulative distance between the two is calculated as the minimum alignment distance.

5. The use of the detection method of claim 1 in identifying multi-type load redistribution attacks. The various types of load redistribution attacks include general load redistribution attacks, stealthy load redistribution attacks, and enhanced load redistribution attacks.

6. Use of the detection method according to claim 5 in the identification of multi-type load redistribution attacks, characterized in that, The general load redistribution attack is established through the following steps: Obtain the net injected power vector of the power grid and construct a DC power flow model that characterizes the relationship between the net injected power vector and the power flow vector; Generate an injected load measurement vector, wherein the algebraic sum of each element in the injected load measurement vector is zero, and the absolute value of each element does not exceed a preset proportion of the corresponding normal load measurement value. The injected load measurement vector is superimposed on the net injected power vector, the injected power flow vector is calculated according to the DC power flow model, and the change in the power generation vector is set to zero.

7. Use of the detection method according to claim 5 in the identification of multi-type load redistribution attacks, characterized in that, The stealthy load redistribution attack is established through the following steps: Multiple historical measurement vectors are obtained, and the first objective function is to minimize the sum of the Euclidean distances between the currently injected measurement vector and each of the historical measurement vectors. Under the constraints of the general load redistribution attack, the first objective function is optimized to obtain the injected load measurement vector and power flow vector; The injected load measurement vector and power flow vector obtained by the optimization solution also satisfy the following: the ratio of the power flow change of at least one transmission line to the maximum power flow capacity exceeds the preset power flow over-limit threshold.

8. Use of the detection method according to claim 5 in the identification of multi-type load redistribution attacks, characterized in that, The enhanced load redistribution attack is established through the following steps: The goal is to obtain the power generation cost under normal conditions, and use the difference between the power generation cost calculated after injecting erroneous data and the normal power generation cost as the second objective function. Under the constraints of the general load redistribution attack or the covert load redistribution attack, the second objective function is maximized to obtain the injected load measurement vector and power flow vector, such that the power generation cost after injection is higher than the normal power generation cost.

9. A computer device, comprising: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for detecting power grid load redistribution attacks according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed on a computer device, cause the computer device to perform the method for detecting power grid load redistribution attacks according to any one of claims 1 to 8.