Network security control method and device for islanded ac microgrid

Abstract

This invention discloses a network security control method and device for an islanded AC microgrid, relating to the field of microgrid technology. The method includes: acquiring the current state information of the local distributed generation unit and neighboring distributed generation units; generating local spurious data injection attack prediction information and state estimation information based on the state information using an attack prediction network model and a state estimation network model; generating local secondary control information based on the local spurious data injection attack prediction information to control the local distributed generation unit; generating an approximate supervision signal using attack compensation estimation information, and updating the network weights of the attack prediction network model using an online learning algorithm based on the approximate supervision signal. This invention utilizes a dual deep learning model collaborative attack compensation architecture, overcoming the limitations of traditional model dependence and offline training, achieving real-time compensation for FDIA and online updating of network parameters, thus improving the accuracy of attack compensation.

Description

Technical Field

[0001] This invention relates to the field of microgrid technology, and in particular to a network security control method and device for an islanded AC microgrid. Background Technology

[0002] To ensure the stable operation of microgrids, a hierarchical control architecture is typically adopted, including primary droop control that realizes the proportional power distribution among distributed generation units and secondary control responsible for restoring the system frequency and voltage to rated values. Compared to centralized control that relies on a single central node, distributed secondary control only requires local information exchange between adjacent units and achieves global control objectives through local decision-making. It has significant advantages in terms of flexibility, scalability, and fault tolerance, and is therefore widely used in microgrid systems. However, the dependence of distributed secondary control on communication networks also exposes it to significant cybersecurity threats; among them, False Data Injection Attacks (FDIA) are particularly noteworthy due to their high degree of concealment and destructive potential.

[0003] In recent years, model- and data-driven attack compensation strategies have gradually developed. However, model-based methods rely on accurate system dynamics models, which are difficult to accurately obtain in practice due to parameter uncertainties and dynamic operating conditions, thus limiting their robustness. Data-driven methods, on the other hand, often use offline pre-trained networks, lacking online update capabilities and struggling to adapt to real-time evolving attack patterns and system changes. Therefore, how to achieve real-time compensation for FDIAs and online updates of network parameters to improve the accuracy of attack compensation is an urgent problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a network security control method and device for isolated AC microgrids, so as to realize real-time compensation for FDIA and online updating of network parameters, thereby improving the accuracy of attack compensation.

[0005] To address the aforementioned technical problems, this invention provides a network security control method for an islanded AC microgrid, comprising: Obtain the status information of the local distributed generation unit and neighboring distributed generation units at the current moment; wherein, the status information includes frequency information and voltage information, and the local distributed generation unit is communicatively connected to the neighboring distributed generation units; Based on the state information, local fake data injection attack prediction information and state estimation information are generated using an attack prediction network model and a state estimation network model; wherein, both the attack prediction network model and the state estimation network model are deep learning models. Based on the local fake data injection attack prediction information, local secondary control information is generated to control the local distributed power generation unit; Using the state estimation information, an approximate supervision signal is generated, and using the approximate supervision signal, an online learning algorithm is employed to update the network weights of the attack prediction network model.

[0006] On the other hand, the attack prediction network model is a deep learning model trained using simulation data containing multiple fake data injection attack modes, which represents the mapping relationship between local measurement signals and attack signals; the state estimation network model is a deep learning model trained using historical data of normal system operation, which represents the mapping relationship between local measurement signals and control signals.

[0007] On the other hand, the input of the attack prediction network model The output of the attack prediction network model is the local fake data injection attack prediction information, which includes attack prediction signals. The input to the state estimation network model The output of the state estimation network model is the state estimation information, which includes control prediction signals. ; i represents the local distributed generation unit. Let t be the set of distributed generation units that are communicatively connected to the local distributed generation unit. and These are the frequency and voltage information of the local distributed generation unit at the current moment; and These are the frequency and voltage information of the j-th neighboring distributed generation unit at the current time; and These are the estimated values ​​of the dummy data signals injected into the frequency channel and voltage channel of the local distributed generation unit at the current moment, respectively. and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the actual current moment; and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the predicted current time, respectively.

[0008] On the other hand, the online learning algorithm is stochastic gradient descent. The step of generating an approximate supervision signal using the state estimation information and updating the network weights of the attack prediction network model using the approximate supervision signal and the online learning algorithm includes: The approximate monitoring signal is generated based on the state estimation information and the state information of the local distributed generation unit; The network weights are updated using the stochastic gradient descent method based on the approximate supervision signal.

[0009] On the other hand, before updating the network weights using the stochastic gradient descent method based on the approximate supervision signal, the method further includes: Determine whether the system status value after attack compensation is greater than the status reference value; If so, then the step of updating the network weights using the stochastic gradient descent method based on the approximate supervision signal is performed.

[0010] On the other hand, updating the network weights using the stochastic gradient descent method based on the approximate supervision signal includes: pass Update the network weights; wherein, Discrete time step k Time l Layer weight matrix; For discrete time steps ( k +1) time l Layer weight matrix; γ The learning rate; For loss function, ; loss function about The gradient; Discrete time step k The attack prediction signal output by the attack prediction network model at that time. For attack prediction signals Approximate supervisory signal.

[0011] On the other hand, the step of generating local secondary control information based on the local fake data injection attack prediction information includes: pass Generate secondary control voltage commands; and / or, pass Generate secondary control frequency commands; and / or, pass Generates secondary control active power commands; among which, , and These are the secondary control voltage command, the secondary control frequency command, and the secondary control active power command, respectively, where i represents the local distributed generation unit. This refers to the set of distributed generation units that are communicatively connected to the local distributed generation unit; if the local distributed generation unit is communicatively connected to its j-th neighboring distributed generation unit, then... If the local distributed generation unit can receive the reference information sent by the leader node, then... =1, otherwise =0; The most recent trigger communication time for the j-th neighboring distributed generation unit; , and These are the gain coefficients of the voltage, frequency, and active power controllers, respectively. and These are the d-axis components of the voltage information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. , v qi =0; and These are the frequency information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. and These are the active power information of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively. and These are the output voltage reference value and frequency amplitude reference value of the distributed generation unit, respectively; and These are the attack prediction signals for the voltage and frequency controllers of the local distributed generation unit, respectively. and These are the active power droop coefficients of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively.

[0012] On the other hand, the step of generating local secondary control information based on the local fake data injection attack prediction information includes: Obtain the current status information of the local distributed generation unit; Based on the status information of the local distributed generation unit, determine whether the current time has reached a new trigger communication time; If so, the status information of the local distributed generation unit is sent to each of the neighboring distributed generation units, and the status information sent by each of the neighboring distributed generation units is received.

[0013] On the other hand, determining whether a new trigger communication time has been reached based on the status information of the local distributed generation unit includes: pass and Determine the new trigger communication time; wherein, ; , , , ,H = L + B， for H The smallest eigenvalue, || H ‖for H norm, Voltage information of the local distributed generation unit v i d-axis component, This is a reference value for the output voltage. The voltage controller controls the gain. For voltage tracking error, Due to local measurement error, L Let Laplace's matrix be the undirected graph corresponding to the communication topology of the isolated AC microgrid system where the local distributed generation unit is located. B For the leader matrix, B =diag( b 1,..., b N ), N This refers to the number of distributed generation units in the isolated AC microgrid system. The first internal dynamic variable is to satisfy the following dynamics: ; , , β vi and η vi (0) are all positive numbers. for The time derivative; ; , ; , d λ2 and d max It is a degree matrix D The second smallest and largest eigenvalues; For the following internal dynamic variables to satisfy the dynamics: ; , β pi , η pi (0) are all positive numbers; This is the active power droop coefficient. The active power output by the local distributed generation unit. The active power controller controls the gain. For active power tracking error, for The time derivative.

[0014] Furthermore, the present invention also provides a network security control device for an islanded AC microgrid, comprising: The acquisition module is used to acquire the status information of the local distributed generation unit and the neighboring distributed generation units at the current time; wherein, the status information includes frequency information and voltage information, and the local distributed generation unit is communicatively connected to the neighboring distributed generation units; The generation module is used to generate local fake data injection attack prediction information and state estimation information based on the state information, using an attack prediction network model and a state estimation network model; wherein, both the attack prediction network model and the state estimation network model are deep learning models. The compensation module is used to generate local secondary control information based on the attack prediction information injected by the local fake data, so as to control the local distributed power generation unit. The update module is used to generate an approximate supervision signal using the state estimation information, and to update the network weights of the attack prediction network model using an online learning algorithm based on the approximate supervision signal.

[0015] The present invention provides a network security control method for an islanded AC microgrid, comprising: acquiring the current state information of the local distributed generation unit and neighboring distributed generation units; wherein the state information includes frequency information and voltage information, and the local distributed generation unit and the neighboring distributed generation unit are communicatively connected; generating local spoofed data injection attack prediction information and state estimation information based on the state information using an attack prediction network model and a state estimation network model; wherein both the attack prediction network model and the state estimation network model are deep learning models; generating local secondary control information based on the local spoofed data injection attack prediction information to control the local distributed generation unit; generating an approximate supervision signal using the state estimation information, and updating the network weights of the attack prediction network model using an online learning algorithm based on the approximate supervision signal.

[0016] As can be seen, this invention, by generating local spoofed data to inject attack prediction and state estimation information based on state information using an attack prediction network model and a state estimation network model, and employing a collaborative attack compensation architecture with dual deep learning models, overcomes the limitations of traditional model dependence and offline training. This enables real-time compensation for FDIA and online updating of network parameters, improving the accuracy of attack compensation and ensuring the network security of microgrid operation. Furthermore, this invention also provides a network security control device for islanded AC microgrids, which also possesses the aforementioned beneficial effects. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a network security control method for an isolated AC microgrid provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another network security control method for an isolated AC microgrid provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating a dynamic event triggering mechanism for neural network compensation provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the architecture of a voltage-distributed dynamic event triggering control scheme provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the architecture of a frequency-distributed dynamic event triggering control scheme provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the architecture of an active distributed dynamic event triggering control scheme provided in an embodiment of the present invention; Figure 7 This is a structural block diagram of a network security control device for an isolated AC microgrid provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the network security control device for an isolated AC microgrid provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a network security control method for an islanded AC microgrid, provided as an embodiment of the present invention. The method may include: Step 101: Obtain the status information of the local distributed generation unit and neighboring distributed generation units at the current moment.

[0021] The status information includes frequency information and voltage information, and the local distributed generation unit communicates with neighboring distributed generation units.

[0022] It is understood that this embodiment considers an islanded AC microgrid system consisting of a cooperating leader node and N distributed generator (DG) units. The communication topology of this system can be represented by an undirected graph. To describe. Among them, node set ( J 1, J 2,…, J N )( J i Representing the i DG), edge set Represents communication links, adjacency matrix Indicates the connection relationship between nodes: if node i and j It can exchange information (i.e., a communication connection), that is ,but a ij =1; otherwise a ij =0. Matrix A diagonal elements a ii =0. Node i The neighbor set is defined as Degree matrix D =diag{ d 1,..., d N} elements d i =| N i |, where | N i | is a neighborhood group N i ( The number of elements in an undirected graph. Laplace matrix L Defined as L = D - A For undirected graphs , L It is symmetric and positive semi-definite. Define the leader matrix. B =diag( b 1,..., b N ),in bi =1 indicates the first i One DG can receive a reference signal; otherwise b i =0.

[0023] For example, establishing a collection N An islanded AC microgrid model with distributed generators (i.e., distributed generation units). Primary control employs standard droop control, supplemented by distributed secondary control. To achieve consistency in system frequency, voltage, and active power, a distributed secondary controller is designed based on consensus theory, with the following standard form: (1) (2) (3) Assuming the attacker sends an attack signal δ vi and δ fi Injected into the first i In the voltage and frequency secondary control channels of the DG. Then, under FDIA, it acts on the first i The actual input signal of a DG controller can be expressed as: (4) (5) Assumption 1: The communication diagram in this paper It is an undirected connected graph, and there exists at least one node. i satisfy b i =1, meaning that the node can receive information from the reference node.

[0024] Assumption 2: Attack signals in this paper δ and its time derivative Both are bounded, meaning they have positive constants. F d1 and F d2 , making , .

[0025] Correspondingly, in this embodiment, the local distributed generation unit can be the i-th distributed generation unit; the neighboring distributed generation units can be distributed generation units that are communicatively connected to the local distributed generation unit, such as the neighbor set mentioned above. Distributed generation units (nodes) in the system.

[0026] The specific content of the status information in this step can be set by the designer according to the practical scenario and user needs. For example, it may include frequency information and voltage information, as well as power information and other information. This embodiment does not impose any restrictions on this.

[0027] Step 102: Based on the state information, use the attack prediction network model and the state estimation network model to generate local fake data to inject attack prediction information and state estimation information.

[0028] Both the attack prediction network model and the state estimation network model are deep learning models.

[0029] It is understood that the attack prediction network model in this embodiment can be a model used to represent the mapping relationship between local measurement signals and attack signals; the state estimation network model in this embodiment can be a model used to represent the mapping relationship between local measurement signals and control signals. Both the attack prediction network model and the state estimation network model can adopt deep learning models, such as multilayer perceptron (MLP) or long short-term memory (LSTM) networks.

[0030] Correspondingly, this embodiment may also include the training process of an attack prediction network model and a state estimation network model. For example, the attack prediction network model is a deep learning model trained using simulation data containing multiple fake data injection attack modes, which represents the mapping relationship between local measurement signals and attack signals; the state estimation network model is a deep learning model trained using historical data of normal system operation, which represents the mapping relationship between local measurement signals and control signals.

[0031] For example, both the attack prediction network model and the state estimation network model can employ MLPs to achieve a dual-MLP collaborative architecture. The two MLPs are pre-trained offline, and their training data is normalized. The normalization formula is: (6) in, The value after normalizing and mapping the input feature signal. For input characteristic signals (i.e., voltage information) vi and frequency information fi ), and Let represent the minimum and maximum values ​​of each feature in the entire training dataset, respectively.

[0032] Offline pre-training of the Attack Prediction Network (MLP-Attack) model can be performed using simulated data containing various FDIA attack modes to learn the mapping relationship from local measurement signals to attack signals. The input feature vector of the MLP-Attack model is: The input feature vector of MLP-Attack is: (7) The output of MLP-Attack is local spoofing attack prediction information, which includes attack prediction signals: (8) in, F fi ( t )and F vi ( t These correspond to the estimated values ​​of the dummy data signals injected into the frequency channel and voltage channel, respectively.

[0033] The state estimation network (MLP-State) is pre-trained offline using historical data from normal system operation to establish the inherent dynamic relationship between local measurement signals and control signals.

[0034] The input feature vector of MLP-State is: (9) The output of MLP-State is state estimation information, which may include control prediction signals: (10) Where i represents a local distributed generation unit. The set of distributed generation units that are connected to the local distributed generation unit for communication, where t is the current time. and These are the frequency and voltage information of the local distributed generation unit at the current moment; and These are the frequency and voltage information of the j-th neighboring distributed generation unit at the current time; and These are the estimated values ​​of the dummy data signals injected into the frequency channel and voltage channel of the local distributed generation unit at the current moment, respectively. and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the actual current moment; and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the predicted current time, respectively.

[0035] Secondly, during online operation, the two pre-trained networks work collaboratively through a closed-loop mechanism to achieve real-time attack compensation and online updates of network parameters. MLP-Attack is embedded in the controller of each DG, providing real-time attack compensation signals; an online supervised learning mechanism is employed to achieve online updates of network parameters (i.e., network weights). The core of the online learning phase is utilizing the estimates from the state estimation network. y i s ( t ) and the actual measured value without attack compensation y i meas ( t (As shown in equations (10) and (11)) construct an approximate supervisory signal for the attack signal: (11) in, The approximate supervisory signal for the attack signal serves as the supervisory signal for the online learning of the attack prediction network. By minimizing the mean square error between the attack prediction network's output and the approximate supervisory signal, the network weights are updated online using stochastic gradient descent (SGD). The weight update rule is as follows: (12) in, W l ( k () represents the discrete time step k Time l The weight matrix of the layer, i.e., the first layer k When updating network weights for the first time... l Layer weight matrix; γ >0 represents the learning rate, and the loss function is... , loss function about The gradient; Discrete time step k The attack prediction signal output by the attack prediction network model at that time. For attack prediction signals Approximate supervisory signal.

[0036] To ensure the updating of network parameters, the compensation state residual is defined as: (13) in, The compensated system state value (system frequency or voltage). x ref The state reference value. The weight update in (12) can only be performed when... E i ( k )> th Otherwise, the network weights remain unchanged. Here, th >0 is the preset update threshold.

[0037] In this embodiment, the attack signal estimation error percentage can also be defined to quantify the online estimation accuracy, facilitating subsequent evaluation and adjustment, such as: (14) in, e ( t The percentage of the attack signal estimation error is denoted as . This refers to actual attack signals, such as preset attack signals used for performance testing; F i ( t ) is the attack signal estimated by MLP-Attack (Attack Prediction Network).

[0038] Because MLP-Attack undergoes thorough offline pre-training, its initial compensation maintains the attacked system state within its normal operating neighborhood, ensuring that the input to MLP-State conforms to its training domain (normal operating conditions). This allows MLP-State to output reliable estimates of the attack-free state. Correspondingly, the constructed approximate supervision signal... This provides effective supervision signals for online learning, driving further optimization of the attack prediction network. The self-reinforcing closed loop formed by both iteratively optimizes the compensation accuracy and state estimation quality, promoting asymptotic convergence of the system.

[0039] Step 103: Based on the attack prediction information injected from local fake data, generate local secondary control information to control the local distributed power generation unit.

[0040] Understandably, based on the compensation terms provided by the MLP-Attack mentioned above, a traditional secondary control architecture is introduced, and the secondary controller is reconstructed. Details are as follows: (15) (16) Furthermore, the dependence of distributed secondary control on the communication network not only exposes it to attack threats but also leads to significant waste of communication resources. This embodiment can introduce a triggering mechanism to reduce the frequency of communication triggers. For example, in this step, the current status information of the local distributed generation unit can be obtained; based on the status information of the local distributed generation unit, it is determined whether the current time has reached a new trigger communication time (or trigger update time); if so, the status information of the local distributed generation unit is sent to each neighboring distributed generation unit, and the status information sent by each neighboring distributed generation unit (i.e., the latest status information sent by the neighboring distributed generation units) is received to update the network weights of the attack prediction network model of the local distributed generation unit; if not, this process can be terminated, the status information of the local distributed generation unit is not sent to each neighboring distributed generation unit, and the network weights of the attack prediction network model are not updated, thus reducing the frequency of communication triggers. In other words, in this embodiment, the generation of local secondary control information and the update of the network weights of the attack prediction network model can be at different times (i.e., at different frequencies), such as the generation frequency of local secondary control information being greater than the update frequency of the network weights of the prediction network model.

[0041] For example, an event-triggered mechanism can be introduced (such as using...) Figures 4-6 The event-triggered time generator in the model is used to design dynamic triggering conditions for voltage, frequency, and active power channels respectively (e.g., Figure 3 The triggering conditions are detailed below: First, an event triggering mechanism is introduced, such as... Figures 3-6 As shown, the secondary controller for voltage, frequency, and active power can be reconfigured as follows: (17) (18) (19) in, , For the first i The most recent trigger time of each DG This is the time when communication will be triggered next; during this period, the system state remains unchanged until the next trigger time arrives and is updated. , and These are the secondary control voltage command, secondary control frequency command, and secondary control active power command, respectively, where i represents the local distributed generation unit. Let be the set of distributed generation units that are communicatively connected to the local distributed generation unit; if the local distributed generation unit is communicatively connected to its j-th neighboring distributed generation unit, then The value is 1 if the leader node is active and 0 otherwise; if the local distributed generation unit can receive the reference information sent by the leader node, then... =1, otherwise =0; The most recent trigger communication time for the j-th neighboring distributed generation unit; The most recent trigger communication time for the local distributed generation unit; , and These are the gain coefficients of the voltage, frequency, and active power controllers, respectively. and These are the d-axis components of the voltage information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. , v qi =0; and These are the frequency information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively; and These are the active power information of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively. and These are the output voltage reference value and frequency amplitude reference value of the distributed generation unit, respectively; and These are attack prediction signals for the voltage and frequency controllers of the local distributed generation unit, respectively. and are the active power droop coefficients of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively. and These are attack signals injected by malicious attackers into the voltage and frequency controllers of the local distributed generation unit, respectively. During network training, preset simulated attack signals can be used for training, enabling the attack prediction network model to map the corresponding attack prediction signals (such as...). and ), to counteract attack signals; in actual operation, it can be done through Generate secondary control voltage commands; and / or, through Generate secondary control frequency commands; and / or, through Generate secondary control active power commands; utilize attack prediction signals mapped from the attack prediction network model (such as...) and ), to offset unknown attack signals during actual runtime (such as and ).

[0042] For voltage channels, the event triggering time sequence is defined as follows: (20) in, The tracking error for each DG is... Local measurement error is ; , H = L + B It is a mixed matrix. for H The smallest eigenvalue, || H ‖for H norm, Voltage information for local distributed generation units v i d-axis component, This is a reference value for the output voltage. The voltage controller controls the gain. For voltage tracking error, Due to local measurement error, L An undirected graph corresponding to the communication topology of an islanded AC microgrid system where local distributed generation units reside (as shown in the undirected graph above). The Laplace matrix of ) B For the leader matrix defined above, B =diag( b 1,..., b N ), N This refers to the number of distributed generation units in the isolated AC microgrid system. For the following internal dynamic variables to satisfy the dynamics: (twenty one) in, , β vi , η vi (0) is a positive number (i.e., a preset value). For neural network estimation error, To estimate the error boundary values, It can be The time derivative. The dynamic triggering conditions of the frequency channel are consistent with the structure of the voltage channel.

[0043] For active power channels, the trigger time series is defined as follows: (twenty two) in, The tracking error for each DG is... Local measurement error is ; ,d λ2 and d max It is a degree matrix D The second smallest and largest eigenvalues, For the following internal dynamic variables to satisfy the dynamics: (twenty three) in, , β pi , η pi (0) are all positive numbers. This is the active power droop coefficient. The active power output by the local distributed generation unit. The active power controller controls the gain. For active power tracking error, It can be The time derivative.

[0044] In other words, determining whether a new trigger communication time has been reached based on the status information of the local distributed generation unit can include using... and Determine the new trigger time for communication.

[0045] Correspondingly, in some other embodiments, step 101 may include obtaining the status information of the local distributed generation unit at the current moment; determining whether the current moment has reached a new trigger communication moment based on the status information of the local distributed generation unit; if so, sending the status information of the local distributed generation unit to each neighboring distributed generation unit, and receiving the status information of the current moment sent by each neighboring distributed generation unit, and recording it as their latest status information; if not, obtaining the recorded latest status information of each neighboring distributed generation unit as the current status information of each neighboring distributed generation unit, so as to further reduce the communication trigger frequency.

[0046] It should be noted that in this step, the processor (such as the controller) of the local distributed generation unit can inject attack prediction information based on local fake data (as mentioned above). and ), generate local secondary control information (as described above) , and This allows for the control of local distributed generation units, achieving distributed secondary control. The specific method used in this embodiment to control the local distributed generation units using local secondary control information can be implemented in a manner similar to or the same as the distributed secondary control method for microgrids in related technologies; this embodiment does not impose any limitations on this.

[0047] Step 104: Using the state estimation information, generate an approximate supervision signal, and use the approximate supervision signal to update the network weights of the attack prediction network model using an online learning algorithm.

[0048] It is understandable that the specific method for generating the approximate monitoring signal using the state estimation information in this step, i.e., the method for generating the approximate monitoring signal, can be set by the designer. For example, the approximate monitoring signal can be generated based on the state estimation information and the state information of the local distributed generation unit, or the approximate monitoring signal can be calculated using the above formula (11). This embodiment does not impose any restrictions on this.

[0049] Similarly, this embodiment does not limit the specific type of online learning algorithm. For example, stochastic gradient descent or other online learning algorithms can be used.

[0050] Furthermore, before updating the network weights of the attack prediction network model using an approximate supervisory signal and an online learning algorithm (such as stochastic gradient descent), it is also possible to determine the system state value after attack compensation (e.g., ...). Figure 2 In E i ( k Is it greater than the status reference value (e.g.) Figure 2 In th If yes, then the process of updating the network weights of the attack prediction network model using an online learning algorithm with an approximate supervisory signal is executed, such as updating the network weights using the above formula (12); if no, then this process can be terminated and the network weights of the prediction network model are not updated.

[0051] It should be noted that in this embodiment, a Lyapunov function can be constructed to prove the asymptotic convergence of the closed-loop system under bounded attacks, and Zeno behavior can be strictly excluded. For example, by constructing a composite Lyapunov function, it can be proven that under the conditions of a connected communication topology, the existence of at least one leader, and a bounded attack signal, the system tracking error asymptotically converges to zero, i.e., the consistency of system voltage, frequency, and active power is restored. Furthermore, by analyzing the lower bound of the time interval between consecutive triggering moments, Zeno behavior can be strictly excluded, ensuring the feasibility of the control strategy. For example, the stability proof in this embodiment can be performed through the following process.

[0052] a. During the voltage control process, as shown in equation (16), the system stability analysis is as follows: Theorem 1: Under the conditions of satisfying Assumption 1 and Assumption 2, for a given If both conditions are met: (twenty four) (25) (26) Under the distributed dynamic event triggered control protocol (16), even if the microgrid is subjected to an FDIA attack, the voltage of each DG can be restored to the rated value, that is... Furthermore, Zeno behavior does not exist under the trigger condition (19).

[0053] Proof 1: Consider the following Lyapunov function: (27) in, , for transpose, For a compact form of voltage tracking error, obviously V It is positive.

[0054] First, it needs to be proven that: ; According to the triggering conditions (19) and (20): (28) Based on the principle of comparison, we can conclude that: (29) therefore, W ( t It is positive definite. The next step is to prove that it is positive definite. Considering tracking error and local measurement error e vi ( t From this, we can obtain the compact form of the entire system: (30) therefore, V The time derivative is: (31) The second and third terms in equation (30) are derived using inequalities. ,get: (32) Equation (30) can be further derived as follows: (33) Considering the internal dynamic variables, we can further obtain : (34) because , , And satisfy ,therefore That is, under the proposed dynamic triggering scheme, the output voltage of all DGs... Asymptotic synchronization to rated value V ref The proof is complete.

[0055] Proof 2: It needs to be proven that under the event triggering mechanism (19), the time interval between any two consecutive events has a positive lower bound, thus excluding Zeno behavior.

[0056] In the interval Inside, e vi ( t )satisfy: .

[0057] consider ,have: (35) In the Lyapunov stability proof, it was determined that the system is asymptotically stable, therefore the tracking error... It asymptotically converges to zero and has positive constants under bounded initial conditions. Make Furthermore, the event-triggered mechanism ensures local measurement errors are addressed. The prediction error remains bounded between consecutive trigger times. Assumption 2 states, the prediction error caused by the attack signal... It is also bounded. Therefore, from equation (34), we can obtain that there exists a constant C > 0 such that right Established. Due to We can obtain: (36) From equation (19), it can be seen that the following condition is satisfied at the triggering time: (37) Combining equations (35) and (36), we have: (38) Equation (37) shows that Zeno behavior does not exist. Proof complete.

[0058] b. Under the frequency control scheme, i.e., equation (17), the system stability analysis is consistent with the proof of analysis a, so it is omitted here.

[0059] c. Under the active power control scheme, i.e., equation (18), the system stability analysis is as follows: Theorem 2: Under the conditions of satisfying Assumption 1 and Assumption 2, for a given α p1 If the following conditions are met: .

[0060] Under the distributed dynamic event triggering control protocol (18), the active power of the system achieves consistency, that is, there is no Zeno behavior under the triggering condition (21).

[0061] Proof 3: Consider the following Lyapunov function: (39) Using a proof method similar to Theorem 1, we obtain: (40) right W P ( t Find the time derivative: (41) in, According to the conditions in Theorem 2, This ensures the asymptotic convergence of active power sharing.

[0062] Proof 4: Similarly, using the proof method in Theorem 1, it can be proved that there is no Zeno behavior in equation (21).

[0063] In this embodiment, the present invention generates local fake data injection attack prediction information and state estimation information by using an attack prediction network model and a state estimation network model based on state information. By using an attack compensation architecture that coordinates two deep learning models, it overcomes the limitations of traditional model dependence and offline training, and can realize real-time compensation for FDIA and online updating of network parameters, thereby improving the accuracy of attack compensation and ensuring the network security of microgrid operation.

[0064] Corresponding to the above method embodiments, this invention also provides a network security control device for an islanded AC microgrid. The network security control device for an islanded AC microgrid described below and the network security control method for an islanded AC microgrid described above can be referred to in correspondence.

[0065] Please refer to Figure 7 , Figure 7 This is a structural block diagram of a network security control device for an islanded AC microgrid, provided in an embodiment of the present invention. The device may include: The acquisition module 10 is used to acquire the status information of the local distributed generation unit and the neighboring distributed generation unit at the current time; wherein, the status information includes frequency information and voltage information, and the local distributed generation unit and the neighboring distributed generation unit are connected in communication. The generation module 20 is used to generate local fake data injection attack prediction information and state estimation information based on the state information using the attack prediction network model and the state estimation network model; wherein, the attack prediction network model and the state estimation network model are both deep learning models. The compensation module 30 is used to inject attack prediction information based on local false data to generate local secondary control information in order to control the local distributed power generation unit. The update module 40 is used to generate an approximate supervision signal using state estimation information, and to update the network weights of the attack prediction network model using an online learning algorithm based on the approximate supervision signal.

[0066] In some embodiments, the attack prediction network model is a deep learning model trained using simulation data containing multiple fake data injection attack modes, representing the mapping relationship between local measurement signals and attack signals; the state estimation network model is a deep learning model trained using historical data of normal system operation, representing the mapping relationship between local measurement signals and control signals.

[0067] In some embodiments, the input to the attack prediction network model The output of the attack prediction network model is to inject attack prediction information into local fake data. This local fake data injection attack prediction information includes attack prediction signals. Input to the state estimation network model The output of the state estimation network model is state estimation information, which includes control prediction signals. ; i represents a local distributed generation unit. The set of distributed generation units that are connected to the local distributed generation unit for communication, where t is the current time. and These are the frequency and voltage information of the local distributed generation unit at the current moment; and These are the frequency and voltage information of the j-th neighboring distributed generation unit at the current time; and These are the estimated values ​​of the dummy data signals injected into the frequency channel and voltage channel of the local distributed generation unit at the current moment, respectively. and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the actual current moment; and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the predicted current time, respectively.

[0068] In some embodiments, the update module 40 may include: The generation submodule is used to generate approximate monitoring signals based on state estimation information and state information of local distributed generation units; The update submodule is used to update the network weights using stochastic gradient descent based on the approximate supervision signal.

[0069] In some embodiments, the update module 40 may further include: The judgment submodule is used to determine whether the system status value after attack compensation is greater than the status reference value; if so, it sends a start signal to the update submodule.

[0070] In some embodiments, the update submodule may be specifically used to... Update network weights; among which, Discrete time step k Time l Layer weight matrix; For discrete time steps ( k +1) time l Layer weight matrix; γ The learning rate; For loss function, ; loss function about The gradient; Discrete time step k The attack prediction signal output by the attack prediction network model. For attack prediction signals Approximate supervisory signal.

[0071] In some embodiments, the compensation module 30 may include: Voltage compensation submodule, used to... Generate secondary control voltage commands; and / or, The frequency compensation submodule is used to... Generate secondary control frequency commands; and / or, The power compensation submodule is used to... Generates secondary control active power commands; among which, , and These are the secondary control voltage command, secondary control frequency command, and secondary control active power command, respectively, where i represents the local distributed generation unit. Let be the set of distributed generation units that are communicatively connected to the local distributed generation unit; if the local distributed generation unit is communicatively connected to its j-th neighboring distributed generation unit, then The value is 1 if the leader node is active and 0 otherwise; if the local distributed generation unit can receive the reference information sent by the leader node, then... =1, otherwise =0; The most recent trigger communication time for the j-th neighboring distributed generation unit; , and These are the gain coefficients of the voltage, frequency, and active power controllers, respectively. and These are the d-axis components of the voltage information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. , v qi =0; and These are the frequency information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively; and These are the active power information of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively. and These are the output voltage reference value and frequency amplitude reference value of the distributed generation unit, respectively; and These are attack prediction signals for the voltage and frequency controllers of the local distributed generation unit, respectively. and These are the active power droop coefficients for the local distributed generation unit and the j-th neighboring distributed generation unit, respectively.

[0072] In some embodiments, the acquisition module 10 may include: The self-acquisition submodule is used to obtain the status information of the local distributed generation unit at the current moment; The trigger judgment submodule is used to determine whether a new trigger communication time has been reached based on the status information of the local distributed generation unit. The send and receive submodule is used to send the status information of the local distributed generation unit to each neighboring distributed generation unit and receive the status information sent by each neighboring distributed generation unit when a new trigger communication time is reached.

[0073] In some embodiments, the triggering judgment submodule may include: Time generator, used to generate time through and Determine the new trigger communication time; among which, ; , , , , H = L +B， for H The smallest eigenvalue, || H ‖for H norm, Voltage information for local distributed generation units v i d-axis component, This is a reference value for the output voltage. The voltage controller controls the gain. For voltage tracking error, Due to local measurement error, L Let Laplace's matrix be the undirected graph corresponding to the communication topology of the isolated AC microgrid system where the local distributed generation unit is located. B For the leader matrix, B =diag( b 1,..., b N ), N The number of distributed generation units in an isolated AC microgrid system; The first internal dynamic variable is to satisfy the following dynamics: ; , , β vi and η vi (0) are all positive numbers. for The time derivative; ; , ; , d λ2 and d max It is a degree matrix D The second smallest and largest eigenvalues; For the following internal dynamic variables to satisfy the dynamics: ; , β pi , η pi (0) are all positive numbers; This is the active power droop coefficient. The active power output by the local distributed generation unit. The active power controller controls the gain. For active power tracking error, for The time derivative.

[0074] In this embodiment, the present invention uses the generation module 20 to generate local fake data injection attack prediction information and state estimation information based on state information using an attack prediction network model and a state estimation network model. By using an attack compensation architecture that coordinates two deep learning models, it breaks through the limitations of traditional model dependence and offline training, and can realize real-time compensation for FDIA and online updating of network parameters, thereby improving the accuracy of attack compensation and ensuring the network security of microgrid operation.

[0075] Corresponding to the above method embodiments, this invention also provides a network security control device for an islanded AC microgrid. The network security control device for an islanded AC microgrid described below and the network security control method for an islanded AC microgrid described above can be referred to in correspondence.

[0076] Please refer to Figure 8 , Figure 8 This is a schematic diagram of a network security control device for an isolated AC microgrid, provided in an embodiment of the present invention. The device may include: Memory D1 is used to store computer programs; Processor D2 is used to execute computer programs to implement the steps of the network security control method for islanded AC microgrids provided in the above-described method embodiments.

[0077] Specifically, the network security control device for the isolated AC microgrid provided in this embodiment can be a distributed generation unit, such as the local distributed generation unit in the above embodiment.

[0078] Corresponding to the above device embodiments, this invention also provides a network security control system for an islanded AC microgrid. The network security control system for an islanded AC microgrid described below and the network security control device for an islanded AC microgrid described above can be referred to in correspondence.

[0079] A network security control system for an isolated AC microgrid includes: A leader node and N distributed generation units; N is a positive integer greater than or equal to 2; The distributed generation unit is a network security control device for the isolated AC microgrid provided in the above embodiments.

[0080] Corresponding to the above method embodiments, this invention also provides a computer program product. The computer program product described below and the network security control method for an islanded AC microgrid described above can be referred to and correspond to each other.

[0081] A computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the network security control method for an islanded AC microgrid provided in the above-described method embodiments.

[0082] Corresponding to the above method embodiments, this invention also provides a computer-readable storage medium. The computer-readable storage medium described below and the network security control method for an islanded AC microgrid described above can be referred to in correspondence.

[0083] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the network security control method for an islanded AC microgrid as provided in the above-described method embodiments.

[0084] The computer-readable storage medium can specifically be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatuses, devices, systems, computer-readable storage media, and computer program products disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant details can be found in the method section.

[0086] The foregoing has provided a detailed description of the network security control method and apparatus for an isolated AC microgrid provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A network security control method for an isolated AC microgrid, characterized in that, include: Obtain the status information of the local distributed generation unit and neighboring distributed generation units at the current moment; wherein, the status information includes frequency information and voltage information, and the local distributed generation unit is communicatively connected to the neighboring distributed generation units; Based on the state information, local fake data injection attack prediction information and state estimation information are generated using an attack prediction network model and a state estimation network model; wherein, both the attack prediction network model and the state estimation network model are deep learning models. Based on the local fake data injection attack prediction information, local secondary control information is generated to control the local distributed power generation unit; Using the state estimation information, an approximate supervision signal is generated, and using the approximate supervision signal, an online learning algorithm is employed to update the network weights of the attack prediction network model.

2. The network security control method for an islanded AC microgrid according to claim 1, characterized in that, The attack prediction network model is a deep learning model trained using simulation data containing multiple fake data injection attack modes, representing the mapping relationship between local measurement signals and attack signals; the state estimation network model is a deep learning model trained using historical data of normal system operation, representing the mapping relationship between local measurement signals and control signals.

3. The network security control method for an islanded AC microgrid according to claim 1, characterized in that, Input to the attack prediction network model The output of the attack prediction network model is the local fake data injection attack prediction information, which includes attack prediction signals. The input to the state estimation network model The output of the state estimation network model is the state estimation information, which includes control prediction signals. ; i represents the local distributed generation unit. Let t be the set of distributed generation units that are communicatively connected to the local distributed generation unit. and These are the frequency and voltage information of the local distributed generation unit at the current moment; and These are the frequency and voltage information of the j-th neighboring distributed generation unit at the current time; and These are the estimated values ​​of the dummy data signals injected into the frequency channel and voltage channel of the local distributed generation unit at the current moment, respectively. and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the actual current moment; and These are the frequency control signal and voltage control signal of the local distributed generation unit corresponding to the predicted current time, respectively.

4. The network security control method for an islanded AC microgrid according to claim 1, characterized in that, The online learning algorithm is stochastic gradient descent. The step of generating an approximate supervision signal using the state estimation information, and then using the approximate supervision signal to update the network weights of the attack prediction network model using the online learning algorithm, includes: The approximate monitoring signal is generated based on the state estimation information and the state information of the local distributed generation unit; The network weights are updated using the stochastic gradient descent method based on the approximate supervision signal.

5. The network security control method for an islanded AC microgrid according to claim 4, characterized in that, Before updating the network weights using the stochastic gradient descent method based on the approximate supervision signal, the method further includes: Determine whether the system status value after attack compensation is greater than the status reference value; If so, then the step of updating the network weights using the stochastic gradient descent method based on the approximate supervision signal is performed.

6. The network security control method for an islanded AC microgrid according to claim 4, characterized in that, The step of updating the network weights using the stochastic gradient descent method based on the approximate supervision signal includes: pass Update the network weights; wherein, Discrete time step k Time l Layer weight matrix; For discrete time steps ( k +1) time l Layer weight matrix; γ The learning rate; For loss function, ; loss function about The gradient; Discrete time step k The attack prediction signal output by the attack prediction network model at that time. For attack prediction signals Approximate supervisory signal.

7. The network security control method for an islanded AC microgrid according to any one of claims 1 to 6, characterized in that, The step of generating local secondary control information based on the local fake data injection attack prediction information includes: pass Generate secondary control voltage commands; and / or, pass Generate secondary control frequency commands; and / or, pass Generates secondary control active power commands; among which, , and These are the secondary control voltage command, the secondary control frequency command, and the secondary control active power command, respectively, where i represents the local distributed generation unit. This refers to the set of distributed generation units that are communicatively connected to the local distributed generation unit; if the local distributed generation unit is communicatively connected to its j-th neighboring distributed generation unit, then... If the local distributed generation unit can receive the reference information sent by the leader node, then... =1, otherwise =0; The most recent trigger communication time for the j-th neighboring distributed generation unit; , and These are the gain coefficients of the voltage, frequency, and active power controllers, respectively. and These are the d-axis components of the voltage information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. , v qi =0; and These are the frequency information of the j-th neighboring distributed generation unit and the local distributed generation unit, respectively. and These are the active power information of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively. and These are the output voltage reference value and frequency amplitude reference value of the distributed generation unit, respectively; and These are the attack prediction signals for the voltage and frequency controllers of the local distributed generation unit, respectively. and These are the active power droop coefficients of the local distributed generation unit and the j-th neighboring distributed generation unit, respectively.

8. The network security control method for an islanded AC microgrid according to claim 7, characterized in that, The step of generating local secondary control information based on the local fake data injection attack prediction information includes: Obtain the current status information of the local distributed generation unit; Based on the status information of the local distributed generation unit, determine whether the current time has reached a new trigger communication time; If so, the status information of the local distributed generation unit is sent to each of the neighboring distributed generation units, and the status information sent by each of the neighboring distributed generation units is received.

9. The network security control method for an islanded AC microgrid according to claim 8, characterized in that, The step of determining whether a new trigger communication time has been reached based on the status information of the local distributed generation unit includes: pass and Determine the new trigger communication time; wherein, ; , , , , H = L + B， for H The smallest eigenvalue, || H ‖for H norm, Voltage information of the local distributed generation unit v i d-axis component, This is a reference value for the output voltage. The voltage controller controls the gain. For voltage tracking error, Due to local measurement error, L Let Laplace's matrix be the undirected graph corresponding to the communication topology of the isolated AC microgrid system where the local distributed generation unit is located. B For the leader matrix, B =diag( b 1,..., b N ), N This refers to the number of distributed generation units in the isolated AC microgrid system. The first internal dynamic variable is to satisfy the following dynamics: ; , , β vi and η vi (0) are all positive numbers. for The time derivative; ; , ; , d λ2 and d max It is a degree matrix D The second smallest and largest eigenvalues; For the following internal dynamic variables to satisfy the dynamics: ; , β pi , η pi (0) are all positive numbers; This is the active power droop coefficient. The active power output by the local distributed generation unit. The active power controller controls the gain. For active power tracking error, for The time derivative.

10. A network security control device for an isolated AC microgrid, characterized in that, include: The acquisition module is used to acquire the status information of the local distributed generation unit and the neighboring distributed generation units at the current time; wherein, the status information includes frequency information and voltage information, and the local distributed generation unit is communicatively connected to the neighboring distributed generation units; The generation module is used to generate local fake data injection attack prediction information and state estimation information based on the state information, using an attack prediction network model and a state estimation network model; wherein, both the attack prediction network model and the state estimation network model are deep learning models. The compensation module is used to generate local secondary control information based on the attack prediction information injected by the local fake data, so as to control the local distributed power generation unit. The update module is used to generate an approximate supervision signal using the state estimation information, and to update the network weights of the attack prediction network model using an online learning algorithm based on the approximate supervision signal.

Images