A method and system for real-time electricity market false data injection attack based on double-target generative adversarial network

Abstract

The application provides a real-time power market false data injection attack method and system based on a double-target generative adversarial network, and relates to the field of power market security. The method comprises the following steps: a double-target generative adversarial network combined with a real-time power market model is constructed, the network comprises a generator, a discriminator, and a fluctuation regularization term for improving concealment and an incentive regularization term for improving destructiveness; the network is trained by using real load data, so that the generator learns to generate false data that can pass the bad data detection and mislead the market scheduling to cause economic loss or line congestion; and finally, the trained generator is used to implement the attack. The generated attack does not require sensitive information of the power system, has high concealment and high destructiveness, and provides an effective basis for evaluating and enhancing the market defense system.

Description

Technical Field

[0001] This invention relates to the field of power market security protection, specifically a method and system for real-time power market fake data injection attacks based on a bi-target generative adversarial network. Background Technology

[0002] To ensure the availability and integrity of the electricity market, modern power grids are integrating massive amounts of real-time and offline data for monitoring, control, and protection. However, this new data-driven landscape makes the electricity market more vulnerable to cyberattacks and potentially more severe consequences than ever before. False Data Injection Attacks (FDIA) are a type of cyberattack that disrupts the integrity of grid information. FDIAs tamper with measurement data received by monitoring centers, causing them to misjudge system conditions and make incorrect dispatching decisions. This is one of the most threatening attack methods against the electricity market. Understanding the potential attack mechanisms of FDIAs is crucial for better protecting the operation of the electricity market.

[0003] Current research on FDIA attack methods targeting the power market is not sufficiently practical. The proposed FDIA methods lack real-world impact and are unsuitable as a basis for analyzing FDIA defense measures in the power market. Firstly, existing modeling-based FDIA research assumes attackers possess all or part of the sensitive information of the power system (e.g., line parameters, network topology, state estimation models), using this as a basis to launch FDIA attacks that disrupt the stable operation of the power market. However, in reality, the confidentiality level of such sensitive information is too high for attackers to access. Secondly, existing deep learning-based FDIA research learns the distribution of measurements received by the monitoring center, infers the monitoring center's state estimation model, and then generates false data from random noise and sends it to the monitoring center. However, the generated false data is all random data resembling the distribution of real load data, differing too much from the actual load values ​​in the power market. This makes it difficult to evade the widespread anomaly detection in the power market, resulting in insufficient stealth. Furthermore, because the distribution of false data is too close to real data, it cannot affect the state estimation results, thus failing to disrupt the safe and stable operation of the power market, resulting in insufficient destructive power. Therefore, designing an FDIA method that requires no sensitive information, is highly stealthy, and highly destructive requires further research. Summary of the Invention

[0004] The purpose of this invention is to solve the technical problems existing in the prior art and to provide a method and system for real-time power market fake data injection attacks based on generative adversarial networks, so as to provide a basis for the formulation of power market security protection measures.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for injecting fake data into the real-time electricity market based on a bi-target generative adversarial network includes the following steps:

[0007] Step 1: Based on the bad data detection model and optimal incremental scheduling model of the real-time electricity market, construct a fake data injection attack model for the real-time electricity market;

[0008] Step 2: Based on the fake data injection attack model constructed in Step 1, construct a dual-target generative adversarial network model. The dual-target generative adversarial network model includes a generator, a discriminator, a fluctuation regularization term to improve concealment, and an incentive regularization term to improve destructiveness.

[0009] Step 3: Train the bi-objective generative adversarial network model constructed in Step 2: First, input the historical real load data into the generator and discriminator; then, input the fake load data generated by the generator into the discriminator and the electricity market model; finally, use the output of the electricity market model as a regularization term, and combine it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges and the optimized generator is obtained.

[0010] Step 4: Use the optimized generator obtained in Step 3 to carry out a fake data injection attack: Input the real load data of the real-time power market into the optimized generator to generate fake load data that is covert and destructive, and inject the fake load data into the power system measurement data to mislead the power market dispatch or cause line congestion.

[0011] Furthermore, step one includes:

[0012] Constructing a defective data detection model: Defective data in actual production processes follows a Gaussian distribution. ,in It is the first Standard deviation of the group of measurements, sum of squared residuals of the measurements Follows chi-square distribution ,in For measured values, The number of independent variables in the measured values. The specific process for determining whether there is bad data in the estimated measurements is as follows:

[0013] Select confidence level ,according to Calculate the corresponding residual threshold ;

[0014] Calculate the sum of squared residuals between the actual and estimated measurements:

[0015] ;

[0016] like If the condition is met, then there is bad data in the measurement values. Discard the measurement values ​​or selectively remove bad data.

[0017] Constructing an optimal incremental scheduling model: The optimal incremental scheduling strategy for the real-time electricity market is as follows:

[0018] ;

[0019] ;

[0020] in, Each node Generator power upper and lower limits For all transmission lines, These are bus power generation and bus load, respectively. and line traffic ,in, These are the number of busbars with generators, the number of busbars with loads, and the number of transmission lines, respectively. For the power increment of the generators at each node, To increase the power flow of each transmission line, Incremental load for each node;

[0021] Then, the TSO sets a reference electricity price. This allows for the determination of real-time electricity prices for all buses in the system. The electricity price for each busbar is as follows:

[0022] ;

[0023] in, The first Jacobian matrix List, Let be the shadow price vector of the transmission line. The value of each element is determined by the congestion status of the corresponding transmission line; the more congested the transmission line, the lower the value. The larger the corresponding element value;

[0024] FDIA attackers inject spurious data into the real-time electricity market by tampering with measurement values. Misleading state value estimation results This allows them to profit from or disrupt the normal operation of the electricity market by detecting and exploiting bad data.

[0025] The following is a model for real-time electricity market fake data injection attack based on a bi-objective generative adversarial network:

[0026] ;

[0027] in For the attack vector, Measurements taken after a data injection attack. These represent the number of line faults caused by the FDIA and the profits gained by the attackers from the electricity market, respectively. To achieve a comprehensive attack effect, The weights are respectively for the profit target and the line failure target.

[0028] Furthermore, step two includes:

[0029] Based on the fake data injection attack model constructed in step one, a dual-objective generative adversarial network model is constructed to obtain the optimal attack vector in the fake data injection attack. The objective function of the dual-objective generative adversarial network model is as follows:

[0030] ;

[0031] in, To identify the network, To generate the network, For the actual load distribution, The distribution of the generator output, The generator outputs spurious measurements. for It is the probability of real data;

[0032] The fluctuation regularization term added to the generating network is as follows:

[0033] ;

[0034] in For fluctuation regularization, The learning threshold is used to control the similarity between fake data and real data; The weights of the fluctuation regularization term are given in two cases.

[0035] The objective function for generating the final bi-objective generative adversarial network model is as follows:

[0036] ;

[0037] Among them, incentive regularization term and These can be used to incentivize DOGAN to generate false measurements that cause large electricity price fluctuations or line faults. and These are electricity price and line fault incentive parameters, used to amplify the losses caused by the attack.

[0038] Furthermore, step three includes:

[0039] Training a bi-objective generative adversarial network model using real-world workload data:

[0040] First, pre-train the discriminator for two rounds. , The loss function is:

[0041] ;

[0042] Secondly, the historical real load data is input into the generator G and the discriminator D. The loss function of G is:

[0043] ;

[0044] Then, the spurious load data generated by the generator is input into the discriminator D and the electricity market model;

[0045] Finally, the discriminator's parameters are updated using the discriminator's output, and the generator's parameters are updated using the discriminator's output combined with the electricity market model's output as a bi-objective regularization term.

[0046] Furthermore, step four includes:

[0047] Obtain real-time load data from the real-time electricity market. The sample is then fed into the generator G, which has been trained and optimized in step three; generator G outputs false measurements. Using the false measurement value The original measurement values ​​received by the power system monitoring center were tampered with; wherein, the false measurement values... The constraint of the fluctuation regularization term is satisfied, ensuring that the estimated residual of the monitoring center is less than the threshold for bad data detection, thus passing the bad data detection; simultaneously, the false measurement value... It includes attack characteristics corresponding to incentive regularization terms, misleads the state estimation results of the power system, and induces the real-time power market to generate erroneous dispatch instructions that lead to line congestion or large fluctuations in electricity prices.

[0048] A real-time electricity market spoofing data injection attack system based on a bi-target generative adversarial network includes:

[0049] The attack model construction module is used to construct a fake data injection attack model for the real-time electricity market based on the bad data detection model and the optimal incremental scheduling model of the real-time electricity market.

[0050] A dual-target generative adversarial network model, constructed based on the aforementioned fake data injection attack model, includes:

[0051] A generator is used to receive real payload data and generate fake payload data;

[0052] A discriminator, connected to the generator, is used to determine whether the input data is real load data or fake load data generated by the generator.

[0053] A fluctuation regularization term calculation unit, connected to the generator, is used to calculate and apply a fluctuation regularization term based on the difference between the generator output and the real input, so as to improve the concealment of fake data.

[0054] An incentive regularization term calculation unit, connected to the generator and interacting with an external power market model, is used to calculate and apply incentive regularization terms based on the scheduling results or line status output by the power market model, in order to increase the destructiveness of spurious data.

[0055] The model training module is connected to the generator, discriminator, electricity market model, fluctuation regularization term calculation unit, and incentive regularization term calculation unit, respectively. It is used to input historical real load data into the generator and discriminator; then, it inputs the spurious load data generated by the generator into the discriminator and electricity market model; finally, it uses the output of the electricity market model as the regularization term, and combines it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges, thus obtaining the optimized generator.

[0056] The attack implementation module, connected to the trained generator, is used to input real load data from the real-time electricity market into the optimized generator to generate deceptive and destructive false load data, and inject the false load data into the power system measurement data to mislead electricity market scheduling or cause line congestion.

[0057] Furthermore, the attack model construction module is specifically used for:

[0058] Constructing a defective data detection model: Defective data in actual production processes follows a Gaussian distribution. ,in It is the first Standard deviation of the group of measurements, sum of squared residuals of the measurements Follows chi-square distribution ,in For measured values, The number of independent variables in the measured values. The specific process for determining whether there is bad data in the estimated measurements is as follows:

[0059] Select confidence level ,according to Calculate the corresponding residual threshold ;

[0060] Calculate the sum of squared residuals between the actual and estimated measurements:

[0061] ;

[0062] like If the condition is met, then there is bad data in the measurement values. Discard the measurement values ​​or selectively remove bad data.

[0063] Constructing an optimal incremental scheduling model: The optimal incremental scheduling strategy for the real-time electricity market is as follows:

[0064] ;

[0065] ;

[0066] in, Each node Generator power upper and lower limits For all transmission lines, These are bus power generation and bus load, respectively. and line traffic ,in, These are the number of busbars with generators, the number of busbars with loads, and the number of transmission lines, respectively. For the power increment of the generators at each node, To increase the power flow of each transmission line, Incremental load for each node;

[0067] Then, the TSO sets a reference electricity price. This allows for the determination of real-time electricity prices for all buses in the system. The electricity price for each busbar is as follows:

[0068] ;

[0069] in, The first Jacobian matrix List, Let be the shadow price vector of the transmission line. The value of each element is determined by the congestion status of the corresponding transmission line; the more congested the transmission line, the lower the value. The larger the corresponding element value;

[0070] FDIA attackers inject spurious data into the real-time electricity market by tampering with measurement values. Misleading state value estimation results This allows them to profit from or disrupt the normal operation of the electricity market by detecting and exploiting bad data.

[0071] The following is a model for real-time electricity market fake data injection attack based on a bi-objective generative adversarial network:

[0072] ;

[0073] in For the attack vector, Measurements taken after a data injection attack. These represent the number of line faults caused by the FDIA and the profits gained by the attackers from the electricity market, respectively. To achieve a comprehensive attack effect, The weights are respectively for the profit target and the line failure target.

[0074] Furthermore, the construction steps of the dual-objective generative adversarial network model include:

[0075] Based on the constructed fake data injection attack model, a dual-objective generative adversarial network (GAN) model is built to obtain the optimal attack vector in the fake data injection attack. The objective function of the dual-objective GAN model is as follows:

[0076] ;

[0077] in, To identify the network, To generate the network, For the actual load distribution, The distribution of the generator output, The generator outputs spurious measurements. for It is the probability of real data;

[0078] The fluctuation regularization term added to the generating network is as follows:

[0079] ;

[0080] in For fluctuation regularization, The learning threshold is used to control the similarity between fake data and real data; The weights of the fluctuation regularization term are given in two cases.

[0081] The objective function for generating the final bi-objective generative adversarial network model is as follows:

[0082] ;

[0083] Among them, incentive regularization term and These can be used to incentivize DOGAN to generate false measurements that cause large electricity price fluctuations or line faults. and These are electricity price and line fault incentive parameters, used to amplify the losses caused by the attack.

[0084] Furthermore, the model training module is specifically used for:

[0085] Training a bi-objective generative adversarial network model using real-world workload data:

[0086] First, pre-train the discriminator for two rounds. , The loss function is:

[0087] ;

[0088] Secondly, the historical real load data is input into the generator G and the discriminator D. The loss function of G is:

[0089] ;

[0090] Then, the spurious load data generated by the generator is input into the discriminator D and the electricity market model;

[0091] Finally, the discriminator's parameters are updated using the discriminator's output, and the generator's parameters are updated using the discriminator's output combined with the electricity market model's output as a bi-objective regularization term.

[0092] Furthermore, the attack implementation module is specifically used to: obtain the current real-time load data of the real-time electricity market. The sample is then fed into the generator G, which has been trained and optimized in step three; generator G outputs false measurements. Using the false measurement value The original measurement values ​​received by the power system monitoring center were tampered with; wherein, the false measurement values... The constraint of the fluctuation regularization term is satisfied, ensuring that the estimated residual of the monitoring center is less than the threshold for bad data detection, thus passing the bad data detection; simultaneously, the false measurement value... It includes attack characteristics corresponding to incentive regularization terms, misleads the state estimation results of the power system, and induces the real-time power market to generate erroneous dispatch instructions that lead to line congestion or large fluctuations in electricity prices.

[0093] The advantages of this invention are:

[0094] 1. Combining a real-time electricity market fake data injection attack model and a generative adversarial network model, the generative adversarial network is trained using real load data without requiring sensitive information about the power system;

[0095] 2. The generator input is replaced with real load data instead of random noise, so that the generated fake data is close to the actual load in the electricity market, thereby avoiding abnormal electricity consumption detection and improving the concealment of FDIA; a fluctuation regularization term that can dynamically adjust the fluctuation range of fake data is designed to further improve the concealment of FDIA.

[0096] 3. An incentive regularization term that can cause economic losses in the electricity market and line faults was designed. By interacting with the electricity market model for training, the destructiveness of FDIA was improved. Attached Figure Description

[0097] Figure 1 A flowchart illustrating a real-time electricity market spoofing data injection attack method based on a bi-target generative adversarial network, provided as an embodiment of the present invention;

[0098] Figure 2 This is a flowchart of the training process for the DOGAN bi-objective generative adversarial network model provided in an embodiment of the present invention.

[0099] Figure 3 A comparison chart showing the effectiveness of the dual-objective generative adversarial network model DOGAN provided in this embodiment of the invention in avoiding bad data detection compared with other methods;

[0100] Figure 4 This is a comparison chart showing the economic losses caused by the dual-objective generative adversarial network model DOGAN provided in this embodiment of the invention and other methods. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] The power system monitoring center acquires the operating status of the power system and performs state estimation (SE) through the Supervisory Control and Data Acquisition (SCADA) system. The SCADA system obtains power system operating status information through numerous sensors. Information such as voltage, current, and phase, which have high real-time requirements, can be sampled by synchronous phasor measurement units (PMUs). Status information such as node power and line transmission power, which have lower real-time requirements, can be measured and uploaded by remote terminal units (RTUs). The electricity market consists of a day-ahead market and a real-time market. In the real-time electricity market, the energy price for each participant is determined by the TSO (Treatment System Operator). The TSO receives the state estimation results uploaded by the power system monitoring center to perform optimal incremental scheduling and calculate the Locational Marginal Price (LMP) for each node in the system. Attackers can tamper with the measurements received by the monitoring center to mislead the state estimation results, thereby altering the TSO's optimal incremental scheduling strategy and potentially profiting from or disrupting the stable operation of the power system.

[0103] Please see Figure 1 This invention provides a method for injecting fake data into the real-time electricity market based on a dual-target generative adversarial network, comprising the following steps:

[0104] Step 1: Based on the bad data detection model and optimal incremental scheduling model of the real-time electricity market, construct a fake data injection attack model for the real-time electricity market;

[0105] Step 2: Based on the fake data injection attack model constructed in Step 1, construct a dual-target generative adversarial network model. The generative adversarial network model DOGAN includes a generator, a discriminator, a fluctuation regularization term to improve concealment, and an incentive regularization term to improve destructiveness.

[0106] Step 3: Train the Generative Adversarial Network (DOGAN) model constructed in Step 2: First, input historical real load data into the generator and discriminator; then, input the fake load data generated by the generator into the discriminator and the electricity market model; finally, use the output of the electricity market model as a regularization term, and combine it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges, and obtain the optimized generator.

[0107] Step 4: Use the optimized generator obtained in Step 3 to carry out a fake data injection attack: Input the real load data of the real-time power market into the optimized generator to generate fake load data that is covert and destructive, and inject the fake load data into the power system measurement data to mislead the power market dispatch or cause line congestion.

[0108] In this embodiment of the invention, step one includes:

[0109] The power system state estimation model includes a bad data detection module to detect and filter bad data.

[0110] Defective data detection module: Defective data in the measurements follows a Gaussian distribution. (in It is the first The standard deviation of the group of measurements; therefore, the mean square residual of the measurements. Follows chi-square distribution ,in This represents the number of independent variables in the measurement data. The process for determining whether there is any bad data in the estimated measurement values ​​is as follows:

[0111] 1) Select confidence level ,according to Calculate the corresponding residual threshold .

[0112] 2) Calculate the mean square error between the actual measured value and the estimated measured value. .

[0113] 3) If If the condition is met, then there is bad data in the measurement values, and the measurement values ​​in that group can be discarded or bad data can be selectively removed.

[0114] The optimal incremental dispatch strategy for the real-time electricity market is as follows:

[0115] ;

[0116] ;

[0117] in, Each busbar Generator power upper and lower limits This refers to the collection of all power transmission lines. These are bus power generation and bus load, respectively. and line traffic ,in, These are the number of busbars with generators, the number of busbars with loads, and the number of transmission lines, respectively. For the power increment of each bus generator, To increase the power flow of each transmission line, This represents the load increment for each bus.

[0118] Then, the TSO sets a reference electricity price. This allows for the determination of real-time electricity prices for all buses in the system. The electricity price for each bus is shown below.

[0119]

[0120] in, The Jacobian matrix is ​​the first... List, Let be the shadow price vector of the transmission line. The value of each element is determined by the congestion status of the corresponding transmission line. The more congested the transmission line, the higher the value. The larger the corresponding element value, the better.

[0121] FDIA attackers targeting the real-time electricity market manipulate raw measurement values. Misleading state value estimation results Or the estimated result of the measurement value This allows them to profit from or disrupt the normal operation of the electricity market by detecting bad data. The FDIA model for the real-time electricity market is shown below.

[0122]

[0123] in For the attack vector, Measurements taken after a data injection attack. These represent the number of line faults caused by the FDIA and the profits gained by the attackers from the electricity market, respectively. To achieve a comprehensive attack effect, The weights are respectively for the profit target and the line failure target.

[0124] In this embodiment of the invention, step two includes:

[0125] The optimal attack vector is obtained using GANs. A traditional GAN ​​consists of two network models: a generator G and a discriminator D. The generator G network consists of 4 fully connected layers, and except for the output which uses tanh, the remaining layers all use the ReLU activation function. The discriminator D consists of 3 fully connected layers, with the output layer using the Sigmoid activation function and the remaining layers using the Leaky ReLU activation function. The objective function of the GAN is shown below.

[0126]

[0127] in For discriminator, For generator, For the actual load distribution, The distribution of the generator output, The generator outputs spurious measurements. for It represents the probability of real data. During the training process, according to Update its own parameters to improve the accuracy of discrimination. Similarly, according to and It updates its own parameters to learn the distribution of measured values, thereby generating fake data that is closer to the real data.

[0128] In order to make The output is detected by malicious data detection; the method proposed in this invention generates tampered measurement values. Formula (5) must be true. The residual is ,in, This is the monitoring center's estimate of the tampered measurement values. Residual. The smaller the value, the more likely it is to be a tampered measurement. The higher the probability of passing the test, the more similar the actual measurement will be to the corresponding estimated measurement, resulting in a lower residual. To avoid the possibility of spurious data being too close to the actual data, a fluctuation regularization term is added to the generator of the GAN network, incorporating the idea of ​​a variational autoencoder. As shown below.

[0129]

[0130] in, The threshold for fake data being too similar to real data. These are the weights of the regularization term in two cases. They can be adjusted... The size of the generated fake data is adjusted so that it fluctuates within a certain range of the real data, preventing the model from becoming too close to the real data after learning the distribution of the real data.

[0131] To increase the destructiveness of FDIA, an activation regularization term is used. To incentivize DOGAN to generate false measurements that cause large electricity price fluctuations, among which This is a hyperparameter representing the weight of this incentive regularization term. Electricity prices in the real-time electricity market are primarily related to power flow and line congestion status. This regularization term incentivizes DOGAN to generate spurious measurements that lead to large-scale power flow changes. Then, to induce FDIA and cause even larger line faults, this invention adds incentive regularization to DOGAN. .in It is the weight of line faults. Large-scale line faults can lead to severe load imbalance, causing generators to trip or loads to disconnect.

[0132] In this embodiment of the invention, the training process for DOGAN is as follows: Figure 2 As shown, step three includes:

[0133] DOGAN is trained using real-world payload data. First, a discriminator D is pre-trained for two rounds, with the loss function of D being... Second, historical real load data is input into the generator G and the discriminator D. The loss function of G is... Then, the spurious load data generated by the generator is input into the discriminator D and the electricity market model. Finally, the discriminator's parameters are updated using the discriminator's output, and the generator's parameters are updated using the discriminator's output combined with the electricity market model's output as a bi-objective regularization term.

[0134] Step four includes:

[0135] Obtain real-time load data from the real-time electricity market. Input it into the generator after training and optimization in step three. generator Output false measurement values Using the false measurement values The original measurement values ​​received by the power system monitoring center were tampered with; wherein, the false measurement values... The constraint of the fluctuation regularization term is satisfied, ensuring that the estimated residual of the monitoring center is less than the threshold for bad data detection, thus passing the bad data detection; simultaneously, the false measurement value... It includes attack characteristics corresponding to incentive regularization terms, misleads the state estimation results of the power system, and induces the real-time power market to generate erroneous dispatch instructions that lead to line congestion or large fluctuations in electricity prices.

[0136] In this embodiment of the invention, the stealth and harmfulness of the fake data injection attack deployed using DOGAN are as follows: Figure 3 and Figure 4 As shown, Figure 3 (Comparison chart of DOGAN's effectiveness in avoiding bad data detection with other methods) and Figure 4 (The comparison chart of economic losses caused by DOGAN and other methods) visually demonstrates the core advantages of the proposed method (DOGAN) compared to existing technologies. The differences between this invention and existing technologies are mainly reflected in the following three aspects:

[0137] I. The fundamental difference between attack premises and information dependence

[0138] Existing technologies (based on modeling optimization): Their attack methods heavily rely on the mastery of sensitive internal information of the power system, such as accurate network topology, line parameters, and state estimation models (Jacobi matrix H). Their effectiveness depends heavily on the completeness and accuracy of the information, and their application is limited in practice due to the difficulty in obtaining sensitive information.

[0139] Existing techniques (based on deep learning): While not requiring a specific model, they typically generate spoof data from random noise. This results in a significant difference between the data distribution and the actual workload, such as... Figure 3 As shown, its success rate in detecting bad data may be low, and its concealment is poor.

[0140] This invention (DOGAN) innovatively uses real historical load data as input to the generator. This allows attacks to be constructed without any sensitive information about the power system, and can be carried out using only publicly available or observable measurement data, significantly lowering the attack threshold and making it more closely aligned with real-world threat scenarios.

[0141] II. Significantly enhanced attack stealth (ability to bypass detection)

[0142] Figure 3 The core comparison metric is the "success rate of detecting bad data".

[0143] Existing technologies (deep learning based on random noise): While the generated fake data possesses a certain degree of random realism, its correlation with the real load context at a specific moment is weak, and the residuals... It is easy to exceed the residual threshold. This leads to the system identifying it as "bad data" and filtering it, such as... Figure 3 Its success rate curve is low or unstable.

[0144] This invention (DOGAN): uses a specially designed fluctuation regularization term to dynamically constrain data generation. Compared to real input The mechanism ensures that spurious data is not randomly generated, but rather fluctuates reasonably around the real data, thus reducing the residuals. Maintaining the residual threshold Within. For example Figure 3 As shown, the DOGAN curve is significantly higher than other comparative methods, proving that it can continuously and effectively bypass bad data detection and has extremely strong concealment.

[0145] III. Targeted Enhancement of Offensive Destructiveness (Ability to Cause Damage)

[0146] Figure 4 The core comparative indicators are "the economic losses caused (such as the total amount of electricity price fluctuations and line congestion costs)" or "the number of line faults".

[0147] Existing modeling and optimization methods may achieve targeted destruction under ideal information, but the prerequisites are stringent.

[0148] Traditional deep learning methods primarily aim to improve stealth (mimicking real data) and lack mechanisms to guide attacks to produce specific destructive effects. Therefore, even if the generated fake data is successfully injected, its actual impact on market scheduling and system operation may be very weak, such as... Figure 4 The resulting loss curve is at a low level.

[0149] This invention (DOGAN): creatively introduces a dual-objective incentive regularization term ( , ).

[0150] Electricity price incentives ( ): This drives the generator to produce false data that can cause drastic fluctuations in node electricity prices.

[0151] Line fault excitation items ( ): This drives the generator to produce false data that can cause line power flow to exceed limits, congestion, or even failure.

[0152] Through interaction and feedback with the electricity market model during training, DOGAN can simultaneously learn "how to hide" and "how to disrupt". For example... Figure 4 As shown, the DOGAN curve is much higher than that of traditional deep learning methods, indicating that it can achieve significant economic or physical damage while ensuring stealth, thus achieving a balance between stealth and destructiveness.

[0153] In summary, through Figure 3 and Figure 4 The comparison clearly shows:

[0154] The DOGAN method proposed in this invention solves the attack prerequisite problem by using real data as input, the concealment problem by using a fluctuation regularization term, and the destructiveness problem by using an incentive regularization term. It no longer requires difficult-to-obtain sensitive information, and the generated fake data can bypass system detection with a high probability. Figure 3 ), and can also proactively and effectively cause serious market economic losses or systemic operational risks ( Figure 4 This creates a generational difference from existing technologies, providing a more effective analytical model and benchmark for assessing and defending against high-threat fake data injection attacks in the real world.

[0155] This invention also provides a corresponding real-time electricity market spoofing data injection attack system based on a dual-target generative adversarial network, comprising:

[0156] The attack model construction module is used to construct a fake data injection attack model for the real-time electricity market based on the bad data detection model and the optimal incremental scheduling model of the real-time electricity market.

[0157] A dual-target generative adversarial network model, constructed based on the aforementioned fake data injection attack model, includes:

[0158] A generator is used to receive real payload data and generate fake payload data;

[0159] A discriminator, connected to the generator, is used to determine whether the input data is real load data or fake load data generated by the generator.

[0160] A fluctuation regularization term calculation unit, connected to the generator, is used to calculate and apply a fluctuation regularization term based on the difference between the generator output and the real input, so as to improve the concealment of fake data.

[0161] An incentive regularization term calculation unit, connected to the generator and interacting with an external power market model, is used to calculate and apply incentive regularization terms based on the scheduling results or line status output by the power market model, in order to increase the destructiveness of spurious data.

[0162] The model training module is connected to the generator, discriminator, electricity market model, fluctuation regularization term calculation unit, and incentive regularization term calculation unit, respectively. It is used to input historical real load data into the generator and discriminator; then, it inputs the spurious load data generated by the generator into the discriminator and electricity market model; finally, it uses the output of the electricity market model as the regularization term, and combines it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges, thus obtaining the optimized generator.

[0163] The attack implementation module, connected to the trained generator, is used to input real load data from the real-time electricity market into the optimized generator to generate deceptive and destructive false load data, and inject the false load data into the power system measurement data to mislead electricity market scheduling or cause line congestion.

[0164] This invention has the following features and effects:

[0165] 1. Achieved efficient attack construction without sensitive information: By constructing a dual-target generative adversarial network that combines an electricity market attack model and training it with publicly available historical real load data, the attack model can learn and simulate system response without relying on highly confidential core sensitive information such as power grid topology and parameters.

[0166] 2. Effectively balances the stealth and destructiveness of the attack: By designing a dynamic fluctuation regularization term, the generated false data is constrained to fluctuate reasonably around the real data in order to pass the bad data detection, thus improving the stealth of the attack; at the same time, by designing an incentive regularization term, the generator is guided to produce false data that can cause significant electricity price fluctuations or line congestion, thus ensuring the destructiveness of the attack.

[0167] 3. Provides realistic attack simulation and evaluation benchmarks: The fake data generated by this method is more in line with actual load characteristics and market response, providing more concealed, more destructive, and more realistic threat scenarios for power market security protection research, which helps to develop and verify more effective defense measures.

[0168] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for injecting fake data into the real-time electricity market based on a bi-target generative adversarial network, characterized in that, Includes the following steps: Step 1: Based on the bad data detection model and optimal incremental scheduling model of the real-time electricity market, construct a fake data injection attack model for the real-time electricity market; Step 2: Based on the fake data injection attack model constructed in Step 1, construct a dual-target generative adversarial network model. The dual-target generative adversarial network model includes a generator, a discriminator, a fluctuation regularization term to improve concealment, and an incentive regularization term to improve destructiveness. Step 3: Train the bi-objective generative adversarial network model constructed in Step 2: First, input historical real load data into the generator and discriminator; then, input the fake load data generated by the generator into the discriminator and the electricity market model; finally, use the output of the electricity market model as a regularization term, and combine it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges and the optimized generator is obtained. Step 4: Use the optimized generator obtained in Step 3 to carry out a fake data injection attack: Input the real load data of the real-time power market into the optimized generator to generate fake load data that is covert and destructive, and inject the fake load data into the power system measurement data to mislead the power market dispatch or cause line congestion.

2. The method according to claim 1, characterized in that, Step one includes: Constructing a defective data detection model: Defective data in actual production processes follows a Gaussian distribution. ,in It is the first Standard deviation of the group of measurements, sum of squared residuals of the measurements Follows chi-square distribution ,in For measured values, The number of independent variables in the measured values. The specific process for determining whether there is bad data in the estimated measurements is as follows: Select confidence level ,according to Calculate the corresponding residual threshold ; Calculate the sum of squared residuals between the actual and estimated measurements: ； like If the condition is met, then there is bad data in the measurement values. Discard the measurement values ​​or selectively remove bad data. Constructing an optimal incremental scheduling model: The optimal incremental scheduling strategy for the real-time electricity market is as follows: ； ； in, Each node Generator power upper and lower limits For all transmission lines, These are bus power generation and bus load, respectively. and line traffic ,in, These are the number of busbars with generators, the number of busbars with loads, and the number of transmission lines, respectively. For the power increment of the generators at each node, To increase the power flow of each transmission line, Incremental load for each node; Then, the TSO sets a reference electricity price. This allows for the determination of real-time electricity prices for all buses in the system. The electricity price for each busbar is as follows: ； in, The first Jacobian matrix List, Let be the shadow price vector of the transmission line. The value of each element is determined by the congestion status of the corresponding transmission line; the more congested the transmission line, the lower the value. The larger the corresponding element value; FDIA attackers inject spurious data into the real-time electricity market by tampering with measurement values. Misleading state value estimation results This allows them to profit from or disrupt the normal operation of the electricity market by detecting and exploiting bad data. The following is a model for real-time electricity market fake data injection attack based on a bi-objective generative adversarial network: ； in For the attack vector, Measurements taken after a data injection attack. These represent the number of line faults caused by the FDIA and the profits gained by the attackers from the electricity market, respectively. To achieve a comprehensive attack effect, The weights are respectively for the profit target and the line failure target.

3. The method according to claim 1, characterized in that, Step two includes: Based on the fake data injection attack model constructed in step one, a dual-objective generative adversarial network model is constructed to obtain the optimal attack vector in the fake data injection attack. The objective function of the dual-objective generative adversarial network model is as follows: ； in, To identify the network, To generate the network, For the actual load distribution, The distribution of the generator output, The generator outputs spurious measurements. for It is the probability of real data; The fluctuation regularization term added to the generating network is as follows: ； in For fluctuation regularization, The learning threshold is used to control the similarity between fake data and real data; The weights of the fluctuation regularization term are given in two cases. The objective function for generating the final bi-objective generative adversarial network model is as follows: ； Among them, incentive regularization term and These can be used to incentivize DOGAN to generate false measurements that cause large electricity price fluctuations or line faults. and These are electricity price and line fault incentive parameters, used to amplify the losses caused by the attack.

4. The method according to claim 1, characterized in that, Step three includes: Training a bi-objective generative adversarial network model using real-world workload data: First, pre-train the discriminator for two rounds. , The loss function is: ； Secondly, the historical real load data is input into the generator G and the discriminator D. The loss function of G is: ； Then, the spurious load data generated by the generator is input into the discriminator D and the electricity market model; Finally, the discriminator's parameters are updated using the discriminator's output, and the generator's parameters are updated using the discriminator's output combined with the electricity market model's output as a bi-objective regularization term.

5. The method according to claim 1, characterized in that, Step four includes: Obtain real-time load data from the real-time electricity market. The sample is then fed into the generator G, which has been trained and optimized in step three; generator G outputs false measurements. Using the false measurement value The original measurement values ​​received by the power system monitoring center were tampered with; wherein, the false measurement values... The constraint of the fluctuation regularization term is satisfied, ensuring that the estimated residual of the monitoring center is less than the threshold for bad data detection, thus passing the bad data detection; simultaneously, the false measurement value... It includes attack characteristics corresponding to incentive regularization terms, misleads the state estimation results of the power system, and induces the real-time power market to generate erroneous dispatch instructions that lead to line congestion or large fluctuations in electricity prices.

6. A real-time electricity market spoofing data injection attack system based on a dual-target generative adversarial network, characterized in that, include: The attack model construction module is used to construct a fake data injection attack model for the real-time electricity market based on the bad data detection model and the optimal incremental scheduling model of the real-time electricity market. A dual-target generative adversarial network model, constructed based on the aforementioned fake data injection attack model, includes: A generator is used to receive real payload data and generate fake payload data; A discriminator, connected to the generator, is used to determine whether the input data is real load data or fake load data generated by the generator. A fluctuation regularization term calculation unit, connected to the generator, is used to calculate and apply a fluctuation regularization term based on the difference between the generator output and the real input, so as to improve the concealment of fake data. An incentive regularization term calculation unit, connected to the generator and interacting with an external power market model, is used to calculate and apply incentive regularization terms based on the scheduling results or line status output by the power market model, in order to increase the destructiveness of spurious data. The model training module is connected to the generator, discriminator, electricity market model, fluctuation regularization term calculation unit, and incentive regularization term calculation unit, respectively. It is used to input historical real load data into the generator and discriminator; then, it inputs the spurious load data generated by the generator into the discriminator and electricity market model; finally, it uses the output of the electricity market model as the regularization term, and combines it with the output of the discriminator to update the parameters of the generator and discriminator until the model converges, thus obtaining the optimized generator. The attack implementation module, connected to the trained generator, is used to input real load data from the real-time electricity market into the optimized generator to generate deceptive and destructive false load data, and inject the false load data into the power system measurement data to mislead electricity market scheduling or cause line congestion.

7. The system according to claim 6, characterized in that, The attack model construction module is specifically used for: Constructing a defective data detection model: Defective data in actual production processes follows a Gaussian distribution. ,in It is the first Standard deviation of the group of measurements, sum of squared residuals of the measurements Follows chi-square distribution ,in For measured values, The number of independent variables in the measured values. The specific process for determining whether there is bad data in the estimated measurements is as follows: Select confidence level ,according to Calculate the corresponding residual threshold ; Calculate the sum of squared residuals between the actual and estimated measurements: ； like If the condition is met, then there is bad data in the measurement values. Discard the measurement values ​​or selectively remove bad data. Constructing an optimal incremental scheduling model: The optimal incremental scheduling strategy for the real-time electricity market is as follows: ； ； in, Each node Generator power upper and lower limits For all transmission lines, These are bus power generation and bus load, respectively. and line traffic ,in, These are the number of busbars with generators, the number of busbars with loads, and the number of transmission lines, respectively. For the power increment of the generators at each node, To increase the power flow of each transmission line, Incremental load for each node; Then, the TSO sets a reference electricity price. This allows for the determination of real-time electricity prices for all buses in the system. The electricity price for each busbar is as follows: ； in, The first Jacobian matrix List, Let be the shadow price vector of the transmission line. The value of each element is determined by the congestion status of the corresponding transmission line; the more congested the transmission line, the lower the value. The larger the corresponding element value; FDIA attackers inject spurious data into the real-time electricity market by tampering with measurement values. Misleading state value estimation results This allows them to profit from or disrupt the normal operation of the electricity market by detecting and exploiting bad data. The following is a model for real-time electricity market fake data injection attack based on a bi-objective generative adversarial network: ； in For the attack vector, Measurements taken after a data injection attack. These represent the number of line faults caused by the FDIA and the profits gained by the attackers from the electricity market, respectively. To achieve a comprehensive attack effect, The weights are respectively for the profit target and the line failure target.

8. The system according to claim 6, characterized in that, The construction steps of the bi-objective generative adversarial network model include: Based on the constructed fake data injection attack model, a dual-objective generative adversarial network (GAN) model is built to obtain the optimal attack vector in the fake data injection attack. The objective function of the dual-objective GAN model is as follows: ； in, To identify the network, To generate the network, For the actual load distribution, The distribution of the generator output, The generator outputs spurious measurements. for It is the probability of real data; The fluctuation regularization term added to the generating network is as follows: ； in For fluctuation regularization, The learning threshold is used to control the similarity between fake data and real data; The weights of the fluctuation regularization term are given in two cases. The objective function for generating the final bi-objective generative adversarial network model is as follows: ； Among them, incentive regularization term and These can be used to incentivize DOGAN to generate false measurements that cause large electricity price fluctuations or line faults. and These are electricity price and line fault incentive parameters, used to amplify the losses caused by the attack.

9. The system according to claim 6, characterized in that, The model training module is specifically used for: Training a bi-objective generative adversarial network model using real-world workload data: First, pre-train the discriminator for two rounds. , The loss function is: ； Secondly, the historical real load data is input into the generator G and the discriminator D. The loss function of G is: ； Then, the spurious load data generated by the generator is input into the discriminator D and the electricity market model; Finally, the discriminator's parameters are updated using the discriminator's output, and the generator's parameters are updated using the discriminator's output combined with the electricity market model's output as a bi-objective regularization term.

10. The system according to claim 6, characterized in that, The attack implementation module is specifically used to: obtain the real-time load data of the real-time electricity market. Input it into the generator G after training and optimization in step three; Generator G outputs false measurements Using the false measurement value The original measurement values ​​received by the power system monitoring center were tampered with; wherein, the false measurement values... The constraint of the fluctuation regularization term is satisfied, ensuring that the estimated residual of the monitoring center is less than the threshold for bad data detection, thus passing the bad data detection; simultaneously, the false measurement value... It includes attack characteristics corresponding to incentive regularization terms, misleads the state estimation results of the power system, and induces the real-time power market to generate erroneous dispatch instructions that lead to line congestion or large fluctuations in electricity prices.

Images