A false data injection method for testing and evaluating the availability of smart grids

Abstract

The present invention relates to the field of smart grid security analysis, and more specifically, to a method for injecting false data for testing and evaluating smart grid availability. The method comprises: obtaining historical power load data of the smart grid, performing iterative adversarial training using TimeGAN (TimeGan-like network), and obtaining a false data construction model and a target intervention area; and obtaining real-time power load data of the smart grid, and injecting false data into the target intervention area using the false data construction model. By using TimeGAN to construct false data based on the temporal relationship of the smart grid load data, the method achieves enhanced concealment during temporal detection in smart grid availability analysis.

Description

Technical Field

[0001] The present invention relates to the field of smart grid security analysis, and in particular to a false data injection method for testing and evaluating the availability of a smart grid. Background Art

[0002] In recent years, with the deep integration of industrialized production and information technology, power grid systems have gradually become more complex and intelligent. While smart grids, with their highly integrated computer networks, can operate more efficiently and conveniently, they also pose risks of network intrusion. Attacks targeting smart grids have gradually drawn attention to their security and availability.

[0003] To ensure data security, many smart grids use SCADA (Supervisory Control and Data Acquisition) systems for state estimation and anomaly detection. SCADA systems improve data accuracy by providing redundancy in measurement data, automatically detecting false data caused by random interference, and estimating or forecasting the operational status of the smart grid. These measures filter out false data in the smart grid, reducing its potential disruption to normal operations.

[0004] Following the widespread adoption of SCADA systems, some researchers have begun investigating new cyber threats within smart grids. Among these techniques, false data injection (FDI) is the most widely used and particularly destructive to smart grids. Adversaries, relying on knowledge of the grid topology, can manipulate meter measurements to confuse traditional false data detection strategies, leading to erroneous system calls, resulting in economic losses and even paralysis or collapse of the smart grid, potentially causing significant system security issues. Therefore, research on FDI methods can enhance smart grids' ability to detect and defend against cyber threats such as FDI, thereby ensuring their security and availability.

[0005] The existing FDI design has achieved certain results, but still has the following two defects:

[0006] (1) Existing results mainly focus on the intervention of all nodes, but weaken the possibility and intensity of affecting local areas of smart grids under real conditions;

[0007] (2) Existing results emphasize intervention on the static information characteristics of smart grids, while ignoring the hidden nature of the temporal information characteristics in false data. Therefore, it can be identified through a detection strategy based on temporal analysis (referred to as temporal detection).

[0008] Therefore, it is necessary to study an FDI method that can avoid time series detection and identification, construct false data with strong concealment for time series detection, and provide data support for the detection strategy to improve the false data detection and defense capabilities of smart grids. Summary of the Invention

[0009] In view of the defects existing in the prior art, the purpose of the present invention is to provide a false data injection method for testing and evaluating the availability of smart grids, which generates false data with strong concealment of temporal information characteristics.

[0010] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0011] A false data injection method for testing and evaluating the availability of a smart grid comprises the following steps:

[0012] S1. Obtain historical power load data of the smart grid and use TimeGAN iterative adversarial training to obtain false data to construct a model and the area to be intervened.

[0013] S2, obtain the real-time power load data of the smart grid, and use false data to construct a model in the area to be intervened to inject false data;

[0014] The false data includes static information and timing information.

[0015] Furthermore, step S1 specifically includes the following steps:

[0016] S1-1. Obtain historical power load data from the SCADA system of the smart grid, and serially number the historical power load data;

[0017] S1-2. Generate false data by TimeGAN training on the historical power load data, calculate relative norm values and residual test values for the false data, and select the optimal number of false data;

[0018] S1-3. Randomly select the data with the optimal amount of false data from the historical power load data as the area to be intervened, and perform TimeGAN training on the data in the area to be intervened to obtain the false data construction model. Compared with the TimeGAN training in step S1-2, the TimeGAN training in this step has a lower learning rate and a longer iteration cycle.

[0019] Furthermore, step S2 specifically includes the following steps:

[0020] S2-1. Acquire real-time power load data of the smart grid and perform sequence encoding on the real-time power load data;

[0021] S2-2, selecting corresponding data from the real-time power load data according to the area to be intervened, and then inputting the selected data into a false data construction model to generate false data;

[0022] S2-3. According to the sequence coding, the real-time power load data is replaced by the false data, and the real-time power load data mixed with the false data is then transmitted back to the control center of the smart grid, thereby completing the injection of the false data.

[0023] Furthermore, the relative norm value in step S1-2 is calculated according to formula (1):

[0024]

[0025] Where n is the norm of the data obtained in step S1-1 before intervention, Obtain the norm value of the data after intervention for step S1-1, n ex is the relative norm value;

[0026] The residual test value in step S1-2 is calculated according to formula (2) and formula (3):

[0027]

[0028] Where, is the input false data vector, is the optimal AC state estimate of the state vector in the smart grid, Nonlinear mapping function;

[0029]

[0030] Where r i ∈residual, is the two-norm value of the residual, that is, the residual test value;

[0031] The optimal number of false data is selected in step S1-2, and the relative norm value n is compared. ex and residual test values The number of false data at the intersection of the relative norm value and the residual test value change trend is selected as the optimal number of false data.

[0032] Furthermore, the TimeGAN training process includes:

[0033] The smart grid power load data is input into the embedding function and the generator respectively. The embedding function implicitly encodes the data to obtain information features. The generator scrambles the data and implicitly encodes it to construct false data information features.

[0034] The discriminator compares the static information features and temporal information features of the fake data information features constructed by the generator and the unperturbed information features in the embedding function, and returns the loss value to improve the generator;

[0035] Multiple rounds of iterative optimization are performed with the objective function of minimizing the relative entropy of the original data distribution and the false data distribution.

[0036] The false data injection method for testing and evaluating the availability of a smart grid according to the present invention has the following beneficial effects:

[0037] Compared with the traditional false data injection method that mainly targets the static information characteristics of the smart grid, the method described in the present invention constructs false data based on the time series relationship of the smart grid load data, thereby having stronger concealment in the time series detection of the smart grid availability analysis.

[0038] The method of the present invention not only constructs false data with similar time series information characteristics to the original power grid data, but also ensures that the static information characteristics of the false data meet the topological relationship of the smart grid and the physical relationship between various areas.

[0039] The present invention uses TimeGAN to identify the temporal information features and static information features of the data, so that the data can remain hidden in both static and temporal feature dimensions, making it difficult to be identified by traditional false data detection strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention has the following accompanying drawings:

[0041] Figure 1 Flowchart of the model training phase of the present invention;

[0042] Figure 2 Flowchart of the model implementation phase of the present invention;

[0043] Figure 3 The model training implementation synthesis flow chart of the present invention;

[0044] Figure 4 Schematic diagram of the training process of the temporal generative adversarial network used in the present invention;

[0045] Figure 5 The relative norm value and residual test value of the present invention vary with the amount of false data;

[0046] Figure 6 Comparison of missed detection rates for false data injection using the GAN model and the TimeGAN model;

[0047] Figure 7 Power load variation trend diagram of selected data when the sample is not intervened by the method of the present invention and when it is intervened by the method of the present invention;

[0048] Figure 8 Power load change trend chart of selected data before and after the sample was intervened by the traditional method. DETAILED DESCRIPTION

[0049] This paper addresses the shortcomings of existing FDI design concepts in full-node implementation and the lack of awareness of the hidden nature of time series information features in false data. Based on Time-series Generative Adversarial Networks (TimeGAN), this paper proposes a false data injection method for testing and evaluating the availability of smart grids. This method uses multiple rounds of adversarial training with TimeGAN to generate false data with highly hidden time series information features. This paper addresses the availability analysis needs of smart grids and aims to propose an FDI method that avoids existing time series detection. This approach provides a more comprehensive design approach for time series detection, thereby further improving the smart grid's ability to defend against network threats and providing data support for smart grid availability assessment and defense capability testing.

[0050] The present invention will be further described in detail below with reference to the accompanying drawings.

[0051] like Figure 3 As shown, the false data injection method for testing and evaluating the availability of a smart grid according to the present invention comprises the following steps:

[0052] Step 1: Obtain historical power load data of the smart grid and perform iterative adversarial training using TimeGAN to obtain a false data model and the area to be intervened.

[0053] Step 2: obtain the real-time power load data of the smart grid and inject false data into the model constructed by using false data in the area to be intervened.

[0054] Further, if Figure 1 As shown, step 1 specifically includes the following steps:

[0055] Step 1-1, obtaining historical power load data from the SCADA system of the smart grid and serially numbering it;

[0056] Step 1-2: Generate fake data through a small-scale rapid training of TimeGAN on historical power load data. Calculate the relative norm and residual test value of the fake data, compare and select the optimal number of fake data.

[0057] Small-scale fast training has a higher learning rate, shorter iteration cycle, faster training speed, and rougher training results;

[0058] This step aims to study the implementation method of localized FDI, to make up for the shortcomings of full-node implementation in the design of existing FDI methods, and to more realistically simulate the risks actually faced by smart grids;

[0059] Steps 1-3: Randomly select data with the optimal amount of false data from the historical power load data as the area to be intervened, and perform TimeGAN deep model training on the data in the area to be intervened to obtain a false data construction model;

[0060] The learning rate of deep model training is lower, the iteration cycle is longer, the training speed is slower, and the training results are more refined.

[0061] Further, if Figure 2 As shown, step 2 specifically includes the following steps:

[0062] Step 2-1, obtaining real-time power load data of the smart grid and performing sequence encoding on it;

[0063] Step 2-2, selecting corresponding data in the real-time power load data according to the area to be intervened, and then inputting the selected data into the false data construction model to generate false data;

[0064] In step 2-3, according to the sequence coding, the original real-time power load data is replaced with false data, and then the real-time power load data mixed with the false data is transmitted back to the control center of the smart grid to complete the injection of false data.

[0065] The false data includes static information and time series information of the false data, which respectively represent data information corresponding to the static information characteristics and the time series information characteristics.

[0066] Furthermore, the historical power load data comprises: the historical active power and reactive power flowing into each area of the smart grid before a preset moment; that is, the historical active power load data and the historical reactive power load data together constitute the historical power load data; the real-time power load data comprises: the real-time active power and reactive power flowing into each area of the smart grid starting from the current moment; that is, the real-time active power load data and the real-time reactive power load data together constitute the real-time power load data.

[0067] Furthermore, steps 1-2 specifically include the following steps:

[0068] Step 1-2-1, calculate the norm of the data before and after the intervention obtained in step 1-1 according to formula (1), n and Calculate the relative norm value n of the data ex :

[0069]

[0070] Step 1-2-2, input fake data vector Calculate the residual according to formula (2):

[0071]

[0072] In the formula is the optimal AC state estimate of the state vector in the smart grid, yes The nonlinear mapping function of . During the training process, It is actually a fitting function of a fake data vector, so no input state vector is required.

[0073] Step 1-2-3, set variable r i ∈residual, calculate the residual norm according to formula (3) That is, the residual test value:

[0074]

[0075] Step 1-2-4, the relative norm value n obtained in step 1-2-1 and step 1-2-3 ex Compare with the residual test value and select the optimal number of false data. ex The residual test value is the static information characteristic indicator of the data.

[0076] Furthermore, the comparison process described in step 1-2-4 is a process of balancing the destructiveness of false data and the concealment against bad data detectors. When the destructiveness is greater, the residual test value of the data will also increase, which will cause the false data to be detected by the bad data detector with a higher probability, and the success rate will be reduced; if too high concealment is pursued, the relative norm value of the data will be at a high level, with less destructiveness and weakened effect. The optimal number of false data is when this number of false data is generated, the relative norm and the residual value are one that is constantly increasing and the other one that is constantly decreasing. However, only when both are as small as possible, the better, so the final choice is to select the intersection of the two changing trends, such as Figure 5 As shown in the figure, at this intersection point, the data's relative norm and residual test value are both low, indicating high destructiveness and concealment, and the optimal intervention effect. This comparison ensures that the false data construction model can generate highly destructive and concealed data. This analysis provides guidance for measuring various indicators such as the destructiveness of false data in networked collaborative manufacturing platforms.

[0077] Further, if Figure 4 As shown, the TimeGAN training process is as follows:

[0078] The smart grid power load data is input into the embedding function and the generator respectively. The embedding function implicitly encodes the data to obtain information features. The generator scrambles the data and implicitly encodes it to construct false data information features.

[0079] The discriminator compares the static information features and temporal information features of the fake data information features constructed by the generator and the unperturbed information features in the embedding function, and returns the loss value to improve the generator;

[0080] The above training process is iterated for multiple rounds to realize the optimization process of formula (4), and finally false data with time series information characteristics close to the original data is obtained:

[0081]

[0082] Where s and t are the static information characteristics and time series information characteristics of the power load data, respectively, and D(p||q) is the KL distance between two probabilities, or relative entropy. This process can achieve false data distribution. The information difference between the original data distribution p(s, t) is minimized, thus achieving the similarity in the distribution of the two data, and further achieving the purpose of confusing the false data detection strategy in the smart grid. Ultimately, a false data construction model is obtained that can construct time series information characteristics close to the original data.

[0083] After training is complete, the recovery function can decode the information features in the model to obtain false data. The static information features are the topological relationships of the smart grid and the physical relationships between data areas, while the temporal information features are the changes in power load data over time.

[0084] This training process primarily constructs false data with highly concealed temporal information features. However, because the embedding function in the network preserves the static information features of the data, the resulting false data construction model still effectively maintains the topological and physical relationships of the data. Conventional false data detection strategies in smart grids primarily examine the data residual test values described in steps 1-2-3, which are statistics indicating the static information features of the data. Therefore, these characteristics ensure that the false data constructed using the injection method described in this invention is difficult to detect using conventional false data detection strategies.

[0085] The difference between the TimeGAN small-scale fast training and deep model training described in steps 1-2 and 1-3 lies in the learning rates of the generator and discriminator and the number of training iterations. The small-scale fast training proposed in this paper has a higher learning rate, shorter iterations, faster training speed, and more coarse training results; the deep model training has a lower learning rate, longer iterations, slower training speed, and more refined training results.

[0086] Table 1 compares the expected effects of the injection method of the present invention and the traditional FDI method. The performance indicators include the authenticity of the false data (the degree of simulation of the actual network threats faced by the smart grid), the concealment of static information features, the concealment of time series information features, the success rate of evading traditional false data detection strategies, and the success rate of evading time series detection. The advantages and disadvantages of the injection method of the present invention and the traditional FDI method are compared. As can be seen from Table 1, while maintaining the strong concealment advantage of the traditional FDI method against existing false data detection strategies, the injection method of the present invention also has significant advantages in terms of concealment of false data time series information features and the success rate of evading time series detection.

[0087] Table 1

[0088]

[0089] like Figure 6 As shown in the figure, a comparison chart of the missed detection rates of using the GAN model and the TimeGAN model used in the present invention for false data injection is shown. The TimeGAN model learns the temporal feature information and uses it to inject false data into the smart grid, resulting in a missed detection rate significantly higher than that of the GAN model.

[0090] from Figure 7 、 Figure 8 As can be seen, the injection method described in this paper can successfully evade timing detection with a high success rate. Compared to traditional false data injection methods that primarily target static information characteristics of smart grids, the injection method described in this paper constructs false data based on the timing relationships of smart grid load data, thus achieving greater concealment during timing detection in smart grid availability analysis.

[0091] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A false data injection method for testing and evaluating the availability of a smart grid, characterized in that: The steps include: S1. Obtain historical power load data of the smart grid and use TimeGAN iterative adversarial training to obtain false data to construct a model and the area to be intervened. S2, obtain the real-time power load data of the smart grid, and use false data to construct a model in the area to be intervened to inject false data; Wherein, the false data includes static information and time sequence information; The step S1 specifically includes the following steps: S1-1. Obtain historical power load data from the SCADA system of the smart grid, and serially number the historical power load data; S1-2. Generate false data by TimeGAN training on the historical power load data, calculate relative norm values and residual test values for the false data, and select the optimal number of false data; S1-3. Randomly select data with the optimal amount of false data from the historical power load data as the area to be intervened, and perform TimeGAN training on the data in the area to be intervened to obtain the false data construction model. The TimeGAN training in this step has a lower learning rate and a longer iteration cycle than the TimeGAN training in step S1-2. The TimeGAN training process includes: The smart grid power load data is input into the embedding function and the generator respectively. The embedding function implicitly encodes the data to obtain information features. The generator scrambles the data and implicitly encodes it to construct false data information features. The discriminator compares the static information features and temporal information features of the fake data information features constructed by the generator and the unperturbed information features in the embedding function, and returns the loss value to improve the generator; Multiple rounds of iterative optimization are performed with the objective function of minimizing the relative entropy of the original data distribution and the false data distribution.

2. A false data injection method for testing and evaluating the availability of a smart grid according to claim 1, characterized in that ,Step S2 specifically includes the following steps: S2-1. Acquire real-time power load data of the smart grid and perform sequence encoding on the real-time power load data; S2-2, selecting corresponding data from the real-time power load data according to the area to be intervened, and then inputting the selected data into a false data construction model to generate false data; S2-3. According to the sequence coding, the real-time power load data is replaced by the false data, and the real-time power load data mixed with the false data is then transmitted back to the control center of the smart grid, thereby completing the injection of the false data.

3. The method for injecting false data for testing and evaluating the availability of a smart grid according to claim 1, wherein: The relative norm value in step S1-2 is calculated according to formula (1): Where n is the norm of the data obtained in step S1-1 before intervention, Obtain the norm value of the data after intervention for step S1-1, n ex is the relative norm value; The residual test value in step S1-2 is calculated according to formula (2) and formula (3): Where, is the input false data vector, is the optimal AC state estimate of the state vector in the smart grid, yes Nonlinear mapping function; Where r i ∈ residual, is the two-norm value of the residual, that is, the residual test value; The optimal number of false data is selected in step S1-2, and the relative norm value n is compared. ex and residual test values The number of false data at the intersection of the relative norm value and the residual test value change trend is selected as the optimal number of false data.

Images