Electric power system control method and system based on confrontation training and terminal equipment thereof

Abstract

The invention discloses an electric power system control method and system based on adversarial training and terminal equipment thereof, and belongs to the field of electric power systems. Updating the discriminator based on the generated track set; obtaining a current state-action value function; updating the generator based on the approximate expert strategy probability and the current state-action value function; when the preset training target is not met, repeatedly executing an interaction action based on the current generator and the current discriminator, and obtaining a correction generator and a correction discriminator which meet the preset training target; obtaining a Markov agent based on the correction generator and the correction discriminator; actual operation data of the power system are obtained in real time, and a prevention and control strategy of the power system is obtained based on the Markov agent and the actual operation data. According to the adversarial training-based power system control method and system and the terminal equipment thereof disclosed by the invention, efficient and stable acquisition of a prevention and control strategy in dynamic operation of the power system is realized.

Description

Technical Field

[0001] This invention relates to the field of power systems, and more particularly to power system control methods, systems and terminal equipment based on adversarial training. Background Technology

[0002] With the continuous expansion of power system scale, the high proportion of renewable energy integration and the characteristics of diverse loads have made the transient stability problem of power systems increasingly complex, posing unprecedented challenges to the safe and stable operation of power systems. Traditional transient stability analysis methods mainly rely on time-domain simulation and direct methods to formulate preventive control strategies by solving the Transient Stability Constrained Optimal Power Flow (TSC-OPF) problem. Among them, time-domain simulation has high accuracy, but its computational time is significant, making it difficult to meet the real-time requirements of online decision-making; direct methods have relatively high computational efficiency, but they rely on conservative simplification assumptions, often leading to overly conservative control strategies and poor economic efficiency. In recent years, the application of artificial intelligence technology in the field of power systems has provided new solutions for transient stability analysis. For example, deep reinforcement learning (DRL) methods have powerful dynamic decision-making capabilities and have shown unique advantages in preventive control problems based on TSC-OPF. DRL can learn control strategies that adapt to different operating conditions through continuous interaction between the agent and the environment. However, the electromechanical transient simulation process of power systems is computationally complex, resulting in a large amount of time required for DRL training, which seriously restricts its application effect in practical engineering.

[0003] Generative Adversarial Imitation Learning (GAIL), a cutting-edge technology in artificial intelligence, learns control policies directly from expert demonstration data through an adversarial training mechanism. Employing a Generative Adversarial Network (GAN) architecture, it uses a discriminator to distinguish between expert and generated policies, guiding the generator to continuously optimize its policy output. This approach addresses, to some extent, the slow convergence speed and difficulty in designing reward functions in DRL training. However, existing GAIL methods have significant limitations. The algorithm requires a large number of fresh samples for policy updates, leading to low training sample efficiency, unstable training process, and insufficient robustness in convergence performance. Furthermore, the quality of expert data directly affects learning outcomes, while obtaining high-quality expert demonstration data in the power system field is costly.

[0004] In summary, current power system preventive control technologies primarily employ traditional optimization methods and time-domain simulation, which suffer from low computational efficiency and poor online adaptability. Existing deep reinforcement learning methods are limited by the computational efficiency of electromechanical transient simulation, leading to excessively long training times. While generative adversarial learning can improve training efficiency, the generator faces high variance in dominance function estimation during the initial training phase, failing to provide accurate dominance estimates for the actions of each unit, energy storage, and adjustable load at each time step, resulting in training instability. Therefore, there is an urgent need for a power system preventive control method that can address the aforementioned technical problems. Summary of the Invention

[0005] This invention provides a power system control method, system, and terminal equipment based on adversarial training, which can solve the technical problems of low computational efficiency and poor online adaptability in the prior art, and realize the efficient and stable acquisition of prevention and control strategies in the dynamic operation of the power system.

[0006] This invention provides a power system control method based on adversarial training, comprising:

[0007] Acquire relevant power system data, and based on the power system relevant data, obtain a power system environment model and a Markov process model;

[0008] A generator and a discriminator are constructed based on the Markov process model.

[0009] Perform interactive actions: Based on the interaction between the generator and the power system environment model, obtain the generated trajectory set;

[0010] When the generated trajectory set is determined to meet the preset capacity requirement, the discriminator is updated based on the generated trajectory set to obtain the approximate expert policy probability and the current discriminator.

[0011] The current state-action value function is obtained based on the pre-defined adversarial generation strategy algorithm of the Markov process model.

[0012] The generator is updated based on the approximate expert policy probability and the current state-action value function to obtain the current generator;

[0013] When the approximate expert policy probability judgment does not meet the preset training objective, the interactive action is repeated based on the current generator and the current discriminator until a corrected generator and a corrected discriminator that meet the preset training objective are obtained.

[0014] A Markov agent is obtained based on the aforementioned correction generator and correction discriminator;

[0015] The system acquires real-time operational data of the power system and obtains preventive control strategies for the power system based on the Markov agent and the operational data.

[0016] The power system control method based on adversarial training provided by this invention constructs a power system environment model and a Markov process model. Through interaction with the power system environment model, it achieves iterative interactive training and updates of the generator and discriminator networks, enabling the generator to approximate expert strategies. This allows for the efficient acquisition of preventative control strategies adapted to the dynamic operation of the power system, effectively improving the efficiency of power system preventative control. The game-theoretic mechanism of adversarial training enables continuous optimization of the strategy through iterative generation and discriminator processes, ensuring the adaptability of strategy decisions to the system state. The final control strategy, while conforming to expert experience, can also cope with the dynamic changes of the power system. This method solves the technical problems of low computational efficiency and poor online adaptability in existing technologies, ensuring the efficient acquisition and stable adaptation of preventative control strategies during the dynamic operation of the power system.

[0017] Furthermore, the acquisition of power system-related data, and the acquisition of a power system environment model and a Markov process model based on the power system-related data, includes:

[0018] Acquire relevant power system data, and construct a power system environment model based on the power system using time-domain simulation method;

[0019] A cost function is obtained based on the relevant data of the power system.

[0020] Construct a constrained optimization objective function based on preset optimization objectives, cost functions, and preset constraints;

[0021] The power system control model is obtained based on the power system environment model, differential algebraic equations, and constrained optimization objective function.

[0022] Based on the power system control model, a Markov process model of the power system is obtained.

[0023] In the above scheme, a power system environment model is constructed using time-domain simulation. A constrained optimization objective function is constructed by combining the cost function and constraints. The power system control model is obtained by integrating the environment model and differential-algebraic equations. Based on the control model, a Markov process model is constructed to accurately model the operating characteristics and decision-making logic of the power system, providing a high-fidelity environment and scientific decision-making framework for subsequent adversarial training. By leveraging the ability of time-domain simulation to reproduce the dynamic processes of the power system, and by setting a constrained optimization objective function, the constructed Markov process model can realistically reflect the state changes and decision-making needs of the power system, laying a solid data foundation for subsequent agent training.

[0024] Furthermore, obtaining the Markov process model of the power system based on the power system control model includes:

[0025] For any current time step: obtain the current attribute state, current topology state, and current context information based on the power system control model; obtain the current environment state vector based on the current attribute state, current topology state, and current context information; obtain the current action vector based on the current environment state vector, and add the current environment state vector and the current action vector to the current context information to obtain the context information of the next time step, so as to obtain the environment state vector of the next time step based on the context information of the next time step;

[0026] Construct an environment state vector set based on several current environment state vectors at any current time step;

[0027] Construct a set of action vectors based on preset power adjustment rules;

[0028] The state transition probability is obtained based on the power system control model and the preset transition probability calculation method.

[0029] Construct an instantaneous reward function based on the power system control model and the preset reward algorithm;

[0030] Based on the environmental state vector set, action vector set, state transition probability, instantaneous reward function, and preset decay factor, a Markov process model of the power system is obtained.

[0031] Furthermore, the execution of the interactive action includes:

[0032] The generator includes an action network and an evaluation network;

[0033] The generated trajectory set includes the current time step generated trajectory for all time steps;

[0034] For any time step:

[0035] Get the current time step environment state;

[0036] Based on the current time step environmental state and the evaluation network, obtain the current environmental state value estimate;

[0037] The current time step environment state and the current environment state value estimate are input into the action network to obtain the current generation scheme;

[0038] Based on the interaction between the current generation scheme and the power system environment model, the environmental state at the next time step and the real-time reward at the current time step are obtained.

[0039] Input the current time step environment state into the preset expert strategy library to obtain the current time step expert solution and virtual instant reward;

[0040] The current time step generation trajectory is obtained based on the current time step environment state, the current generation scheme, the next time step environment state, the current time step expert scheme, and the virtual instant reward.

[0041] Further, when the generated trajectory set is determined to meet the preset capacity requirement, the discriminator is updated based on the generated trajectory set to obtain the approximate expert policy probability and the current discriminator, including:

[0042] Based on the generated trajectory set, obtain generation strategy data and expert strategy data;

[0043] The generation strategy data is input into the discriminator to obtain the generation output probability;

[0044] The expert strategy data is input into the discriminator to obtain the expert output probability;

[0045] Based on the generated output probability and the expert output probability, an approximate expert policy probability is obtained;

[0046] The discriminator is updated based on the approximate expert policy probability to obtain the current discriminator.

[0047] In the above scheme, generation strategy and expert strategy data are separated from the generated trajectory set and input into the discriminator to obtain output probabilities. Approximate expert strategy probabilities are then calculated based on the two types of probabilities, and the discriminator is updated using these probabilities. This allows the discriminator to accurately distinguish the technical effects of generation strategies and expert strategies, thereby providing the generator with clear optimization direction signals. The discriminator learns by comparing the two types of strategy data, quantifies the similarity between generation strategies and expert strategies, and enables the updated discriminator to more accurately identify the superiority or inferiority of strategies, providing reliable feedback for the generator's subsequent optimization.

[0048] Furthermore, the step of obtaining the simulated state-action value function based on the Markov process model and the preset adversarial generation strategy algorithm includes:

[0049] The current environmental state is obtained based on the Markov process model.

[0050] Based on the current environmental state, obtain the current time step cost and the current time step backup cost;

[0051] A penalty function is constructed based on the current time step cost and current time step backup cost, preset safe operation constraints, and preset limit violation penalty algorithm, and the total penalty for the current time step is obtained based on the penalty function;

[0052] Construct a reward function based on the current time step cost, the current time step reserve cost, and the current time step total penalty;

[0053] The current state-action value function is obtained based on the reward function and the penalty function.

[0054] In the above scheme, a reward function is constructed by using a Markov process model to obtain the system state, fusion cost, backup cost, and limit violation penalty. Based on the reward function and the penalty function, a state-action value function is obtained to quantify the comprehensive benefits of different state-action pairs, thereby achieving the technical effect of providing a quantitative evaluation standard for generator strategy optimization. Combining the constraint effect of the penalty function on safety constraint violations and the balance between economy and security of the reward function, the constructed state-action value function can comprehensively reflect the comprehensive effect of each decision, enabling the generator to clearly identify which actions are better in which states, thus improving the targeting of network optimization.

[0055] Further, updating the generator based on the approximate expert policy probability and the current state-action value function to obtain the current generator includes:

[0056] Get the current correction factor;

[0057] The simulated state reward is obtained based on the approximate expert policy probability and the current correction factor.

[0058] The modified value function is obtained based on the simulated state reward and the current state-action value function;

[0059] A modified advantage function is obtained based on a preset near-end strategy optimization algorithm and the modified value function;

[0060] The gradient loss function is obtained based on the modified advantage function and the preset error gradient backpropagation algorithm.

[0061] The generator is updated based on the gradient loss function to obtain the current generator.

[0062] In the above scheme, a correction factor is introduced, and the simulated state reward is obtained by combining the approximate expert policy probability to construct a corrected value function and a corrected advantage function. Then, the generator is updated based on the gradient loss function, so as to make the generator continuously approach the expert policy, thereby continuously improving the safety and economy of the generation policy. The constraint of the update amplitude by the near-end policy optimization algorithm avoids the generator policy oscillation. The correction factor and simulated state reward ensure that the optimization direction always follows the expert policy, and the gradient loss function can accurately adjust the generator parameters, so that the generation policy gradually reaches the preset optimization goal, realizing the rapid and effective training of the agent.

[0063] Further, obtaining the current correction factor includes:

[0064] Based on the Markov process model, obtain the instantaneous correction factor, instantaneous return value, and historical average return value of the previous time step;

[0065] The current correction factor is obtained based on the real-time correction factor, real-time return value, historical average return value, and preset correction factor update conditions.

[0066] In the above scheme, the correction factor is dynamically adjusted to adapt to the training process, thereby optimizing the generator update rhythm and improving training stability and convergence speed. By combining the comparison between the current reward and the historical reward, the correction factor can be dynamically adjusted according to the training effect: when the current reward is better than the historical reward, the correction magnitude is reduced, and vice versa, to avoid over-correction or under-correction, so that the generator can obtain the appropriate optimization intensity at different training stages.

[0067] This invention provides a power system control method based on adversarial training. It constructs a power system environment model using time-domain simulation to accurately reproduce the dynamic operating characteristics of the power system. It further quantifies the decision-making logic of state, action, reward, and transition using a Markov process model, providing a simulation environment highly consistent with the actual system for agent training, ensuring that the direction of subsequent strategy optimization aligns with the real needs of the power system. It integrates adversarial training and reinforcement learning, achieving breakthroughs in autonomous optimization strategies through expert experience. In constructing the state-value function, it achieves multi-dimensional collaborative optimization of economy, safety, and stability through multi-objective reward function design, providing comprehensive decision support for power system preventive control. It employs a hybrid value function estimation technique to effectively integrate real value estimation with generator-simulated rewards, significantly improving the algorithm's generalization ability and adaptability. Compared with existing transient stability preventive control methods, this method implements an intelligent adjustment mechanism based on dynamic correction factors, using a generative adversarial network to adaptively adjust the correction magnitude of the advantage function. The correction magnitude is dynamically adjusted based on the comparison between the current reward and the historical average reward, significantly improving training stability and convergence speed.

[0068] This invention also provides a power system control system based on adversarial training, for implementing the aforementioned power system control method based on adversarial training, comprising:

[0069] The model building module is used to acquire power system-related data, obtain a power system environment model and a Markov process model based on the power system-related data, and build a generator and a discriminator based on the Markov process model.

[0070] The intelligent agent training module is used to perform interactive actions: It interacts with the power system environment model based on the generator to obtain a set of generated trajectories; when the generated trajectory set determines that a preset capacity requirement is met, it updates the discriminator based on the generated trajectory set to obtain an approximate expert policy probability and a current discriminator; it obtains a current state-action value function based on a preset adversarial generation strategy algorithm of the Markov process model; it updates the generator based on the approximate expert policy probability and the current state-action value function to obtain a current generator; when the approximate expert policy probability determines that the preset training objective is not met, it returns to repeatedly execute the interactive actions based on the current generator and the current discriminator until a corrected generator and a corrected discriminator that meet the preset training objective are obtained.

[0071] The agent deployment module is used to obtain Markov agents based on the correction generator and the correction discriminator.

[0072] The prevention and control module is used to acquire the actual operating data of the power system in real time, and to acquire the power system prevention and control strategy based on the Markov agent and the actual operating data.

[0073] This invention provides a power system control system based on adversarial training. It constructs a power system environment model and a Markov process model through a model building module. Then, an agent training module, based on interaction with the power system environment model, achieves iterative interactive training and updates the generator and discriminator networks, enabling the generator to approximate expert strategies. This allows for the efficient acquisition of preventative control strategies adapted to the dynamic operation of the power system, effectively improving the efficiency of power system preventative control. The game-theoretic mechanism of adversarial training enables continuous optimization of the strategy through iterative generation of the generator and discriminator, ensuring the adaptability of strategy decisions to the system state. The final control strategy, while conforming to expert experience, can also cope with the dynamic changes of the power system. This solves the technical problems of low computational efficiency and poor online adaptability in existing technologies, ensuring the efficient acquisition and stable adaptation of preventative control strategies during the dynamic operation of the power system. Finally, the agent deployment module and preventative control module implement preventative control during the actual operation of the power system. This invention addresses the technical problems of low computational efficiency, poor online adaptability, low training efficiency, and unstable training in existing technologies.

[0074] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the power system control method based on adversarial training of the present invention.

[0075] By implementing this invention, generative adversarial networks (GANs) are combined with deep reinforcement learning to construct an adversarial training mechanism between the generator and the discriminator. Combining a dynamic correction factor mechanism, multi-objective reward function design, and a hybrid estimation technique for state-action value functions, the intelligent adjustment mechanism based on the dynamic correction factor uses adversarial training to adaptively adjust the correction magnitude of the advantage function. The correction magnitude is dynamically adjusted based on the comparison between the current reward and the historical average reward, significantly improving training stability and convergence speed. Through the design of a multi-objective reward function, multi-dimensional synergistic optimization of economy, security, and stability is achieved, providing comprehensive decision support for power system prevention and control. The use of a hybrid value function estimation technique effectively integrates real value estimation with generator-simulated rewards, greatly enhancing the algorithm's generalization ability and adaptability. Attached Figure Description

[0076] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0077] Figure 1 This is a schematic diagram of the power system control method based on adversarial training provided in this embodiment;

[0078] Figure 2 This is a schematic diagram of a power system control algorithm model structure based on adversarial training provided in this embodiment. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0080] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0081] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0082] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0083] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0084] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0085] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0086] This embodiment provides a power system control method based on adversarial training. Please refer to [link to relevant documentation]. Figure 1 ,include:

[0087] S1. Obtain relevant data of the power system, and obtain a power system environment model and a Markov process model based on the relevant data of the power system;

[0088] S2. Construct a generator and a discriminator based on the Markov process model;

[0089] S3. Perform interactive actions: Based on the interaction between the generator and the power system environment model, obtain the generated trajectory set;

[0090] S4. When the generated trajectory set is determined to meet the preset capacity requirement, the discriminator is updated based on the generated trajectory set to obtain the approximate expert strategy probability and the current discriminator;

[0091] S5. Obtain the current state-action value function based on the pre-set adversarial generation strategy algorithm of the Markov process model;

[0092] S6. Update the generator based on the approximate expert policy probability and the current state-action value function to obtain the current generator;

[0093] S7. If the probability judgment based on the approximate expert strategy does not meet the preset training objective, the interaction action is repeated based on the current generator and the current discriminator until a corrected generator and a corrected discriminator that meet the preset training objective are obtained.

[0094] S8. Obtain a Markov agent based on the correction generator and correction discriminator;

[0095] S9. Obtain real-time actual operating data of the power system, and obtain power system prevention and control strategies based on the Markov agent and the actual operating data.

[0096] The power system control method based on adversarial training provided in this embodiment constructs a power system environment model and a Markov process model. Through interaction with the power system environment model, it achieves iterative interactive training and updates of the generator and discriminator networks, enabling the generator to approximate expert strategies. This allows for the efficient acquisition of preventative control strategies adapted to the dynamic operation of the power system, effectively improving the efficiency of power system preventative control. The game-theoretic mechanism of adversarial training enables continuous optimization of the strategy through iterative generation of the generator and discriminator, ensuring the adaptability of strategy decisions to the system state. The final control strategy, while conforming to expert experience, can also cope with the dynamic changes of the power system. This solves the technical problems of low computational efficiency and poor online adaptability in existing technologies, ensuring the efficient acquisition and stable adaptation of preventative control strategies during the dynamic operation of the power system.

[0097] Optionally, step S1 includes:

[0098] Acquire relevant power system data, and construct a power system environment model based on the power system using time-domain simulation method;

[0099] A cost function is obtained based on the relevant data of the power system.

[0100] Construct a constrained optimization objective function based on preset optimization objectives, cost functions, and preset constraints;

[0101] The power system control model is obtained based on the power system environment model, differential algebraic equations, and constrained optimization objective function.

[0102] Based on the power system control model, a Markov process model of the power system is obtained.

[0103] In practical implementation, constructing a power system environment model requires including the power system's topology, equipment models, and load models, considering various constraints of transient operation. Power system-related data includes the network topology, encompassing the locations and connections of power plants, substations, transmission lines, distribution lines, and load nodes. First, the network topology, including the locations and connections of power plants, substations, transmission lines, distribution lines, and load nodes, is determined using power system simulation software such as MATLAB / Simulink, PowerWorld, DSP, and BPA. In practical engineering, this embodiment uses time-domain simulation to analyze the transient stability of the power system and employs a set of differential-algebraic equations (DAEs) to describe the power system, constructing the power system environment model as shown below:

[0104] The differential equation part is shown in the following equation:

[0105]

[0106] The algebraic equation part is shown in the following equation:

[0107] 0 = g(x,y,p)

[0108] In the formula: x represents the state variable describing the dynamic characteristics of the system, which is used to describe the dynamic characteristics of the system (rotor angle, rotor angular velocity, transient electromotive force, excitation system voltage, etc.); y represents the algebraic variable describing the operating state of the system, which is used to describe the operating state of the network; p represents the parameter vector of the system, such as line impedance, transformer turns ratio, load characteristic parameters, etc., which usually represent constants.

[0109] In the prevention and control of power systems, the core optimization objective of this embodiment (i.e., the preset optimization objective) is to seek the economically optimal solution under the hard constraint of ensuring transient stability. This embodiment constructs a high-dimensional, nonlinear constrained optimization model with this objective, and the corresponding constrained optimization objective function is shown in the following equation:

[0110]

[0111] In the formula: ξ represents the set of generators and the set of loads that can be reduced, respectively, and ΔPG,i (t) represents the adjustment of the active power output of generator i during time period t, ΔL j (t) represents the load reduction of j in time period t, where t represents the time step, and S represents the set of observed vectors of the environmental state vector. These represent the corresponding cost functions, typically quadratic or piecewise functions. Specifically, the cost function used in this embodiment... The calculation process is shown in the following formula:

[0112] C G,i (ΔP G,i ) = a i ×(ΔP G,i ) 2 +b i ×|ΔP G,i | 2 +c i

[0113] C S,j (ΔL j )=k j ×|ΔL j |+d j

[0114] In the formula: a i ,b i ,c i It is the cost coefficient of the i-th generator, usually a i >0, these cost coefficients are obtained by fitting the consumption characteristics of the unit; k j This represents the marginal cost of that output range; d j This represents a constant term, used to ensure the continuity of the function at each segmentation point.

[0115] The preset constraints include equality constraints and inequality constraints. Among them, the equality constraints include the power balance equation of the system and the differential equations describing the dynamic behavior of the system components, as shown in the following equation:

[0116]

[0117] The inequality constraints are shown in the following equation:

[0118]

[0119] V i min ≤V i ≤V i max

[0120] In the formula: P lossLet V(t) and θ(t) represent the system network loss, respectively, and let P represent the node voltage magnitude V and phase angle θ at time step t. l V represents the power on the line load. i Let represent the voltage of generator i during the time interval t. The nodal voltage amplitude at time step t is calculated using a differential equation.

[0121] Optionally, obtaining the Markov process model of the power system based on the power system control model includes:

[0122] For any current time step: obtain the current attribute state, current topology state, and current context information based on the power system control model; obtain the current environment state vector based on the current attribute state, current topology state, and current context information; obtain the current action vector based on the current environment state vector, and add the current environment state vector and the current action vector to the current context information to obtain the context information of the next time step, so as to obtain the environment state vector of the next time step based on the context information of the next time step;

[0123] Construct an environment state vector set based on several current environment state vectors at any current time step;

[0124] Construct a set of action vectors based on preset power adjustment rules;

[0125] The state transition probability is obtained based on the power system control model and the preset transition probability calculation method.

[0126] Construct an instantaneous reward function based on the power system control model and the preset reward algorithm;

[0127] Based on the environmental state vector set, action vector set, state transition probability, instantaneous reward function, and preset decay factor, a Markov process model of the power system is obtained.

[0128] In the specific implementation process, when obtaining the Markov process model of the power system based on the power system control model, agents with different objectives have different state definitions, action definitions, transition mechanisms, and reward mechanisms. In this embodiment, the Markov decision process model of the power system is first defined as a tuple (S, A, P, R, γ), where:

[0129] S is the set of observation vectors of the environmental state, s t Let n be the set of n state observations of the system at time t, as shown in the following equation:

[0130] S = [X, G, C]

[0131] X = [P] t Q t[other]

[0132] G = [N] t E t ]

[0133] In the formula: X represents the attribute state (including the current attribute state at any time step), representing the various attributes of each component of the power grid system (such as generators, loads, energy storage, and lines), including active power P. t reactive power Q t Voltage V t Other information, such as line losses: G represents the topology state (including the current topology state at any time step), that is, the topology of the power grid system at time step t, including the node feature matrix N. t and adjacency matrix E t The node feature matrix and adjacency matrix are obtained through a set of differential equations; C represents the context information of the power grid state (including the current context information corresponding to any current time step), which contains a sequence of historical environment state vectors (current environment state vectors corresponding to all time steps before the current time step) and action vectors (current action vectors corresponding to all time steps before the current time step).

[0134] A represents the action vector set of the current agent, which can be divided into discrete and continuous types based on the attributes of the actions. t This represents the current action vector selected by the system at time t. The preset power adjustment rule used in this embodiment is designed for the coordinated interaction characteristics of heterogeneous devices (source, grid, load, energy storage, etc.) in a large-scale power grid under open conditions. It adjusts the active power of the generator, the active power of the adjustable load, and the charging and discharging power of the energy storage device. The action space constructed by the action vector set A includes the active power adjustment value ΔP of the generator. gen The active power adjustment value ΔP of the adjustable load load and the adjustment value ΔP of the charging and discharging power of energy storage devices store Three operating parameters are used to adjust the power system; the action space constructed in this embodiment can be represented by the following formula:

[0135] A=[ΔP gen ,ΔP load ,ΔP store ]

[0136] P: S×A→P(S) represents the state transition probability of the agent. It is a matrix consisting of the probabilities of transitioning to any subsequent state after taking an action under a specific environmental state of the power system. The preset transition probability calculation method can be expressed as follows: E represents the probability that the current power system will execute action a in state s to reach the next time step state s', where E represents the expected value.

[0137] R: S×A→R corresponds to the agent's immediate reward function, defining the immediate reward value that the system can obtain for actions in a specific state. The preset reward algorithm is shown in the following formula:

[0138] R = R margin +R smooth +R renewable

[0139] Among them, R margin R represents the safety margin bonus based on voltage stability margin and frequency deviation; smooth This indicates a reward for smooth control of actions, preventing frequent device actions; R renewable This indicates that the incentives for renewable energy consumption encourage the maximum utilization of renewable energy.

[0140] γ∈[0,1] is a preset decay factor used to measure the impact of rewards obtained in subsequent system states on the value of the current state.

[0141] In its implementation, this embodiment uses an interactive process and adversarial training to iteratively train the generator and discriminator. The generator network used is an Actor-Critic network, comprising an action network (Actor) and an evaluation network (Critic), used to generate an optimized strategy that approximates an expert policy. The discriminator network determines the probability that the generated data is real data. The discriminator and generator compete against each other; while updating the discriminator network, the discriminator outputs an evaluation of the generated actions.

[0142] Optionally, the execution of the interactive action includes:

[0143] The generator includes an action network and an evaluation network;

[0144] The generated trajectory set includes the current time step generated trajectory for all time steps;

[0145] For any time step:

[0146] Get the current time step environment state;

[0147] Based on the current time step environmental state and the evaluation network, obtain the current environmental state value estimate;

[0148] The current time step environment state and the current environment state value estimate are input into the action network to obtain the current generation scheme;

[0149] Based on the interaction between the current generation scheme and the power system environment model, the environmental state at the next time step and the real-time reward at the current time step are obtained.

[0150] Input the current time step environment state into the preset expert strategy library to obtain the current time step expert solution and virtual instant reward;

[0151] The current time step generation trajectory is obtained based on the current time step environment state, the current generation scheme, the next time step environment state, the current time step expert scheme, and the virtual instant reward.

[0152] In the modeling process, an evaluation network (Critic) is constructed to assess the merits of the current strategy. The environment state s at the current time step is considered. t To evaluate the network input, the current state value function is estimated, outputting a reliable estimate of the current environment state value to optimize the policy update process. Simultaneously, an action network (Actor) is constructed to evaluate the differences between the old and new policies. The environment state s at the observed current time step is used as the basis for this evaluation. t As input to the action network, the output layer limits the mean of the output actions to within 5%, constructs a normal distribution of action probability based on the continuously decaying variance, samples the action to obtain the current generation scheme, maps it to the simulated power system environment, and evaluates the merits of the current strategy based on the evaluation network.

[0153] In the specific implementation process, considering the uncertainties of renewable energy power generation and load, the expert database generates the optimal collaborative optimization strategy for the source-grid-load-storage system based on historical system topology, electricity load, and unit output information. This strategy serves as the expert strategy, and the preset expert strategy library is constructed accordingly. By inputting the current time-step environmental state into the preset expert strategy library, the expert solution for the current time step and the corresponding virtual instantaneous reward are obtained.

[0154] Optionally, when the generated trajectory set is determined to meet the preset capacity requirement, the discriminator is updated based on the generated trajectory set to obtain the approximate expert policy probability and the current discriminator, including:

[0155] Based on the generated trajectory set, obtain generation strategy data and expert strategy data;

[0156] The generation strategy data is input into the discriminator to obtain the generation output probability;

[0157] The expert strategy data is input into the discriminator to obtain the expert output probability;

[0158] Based on the generated output probability and the expert output probability, an approximate expert policy probability is obtained;

[0159] The discriminator is updated based on the approximate expert policy probability to obtain the current discriminator.

[0160] In the specific implementation process, the generated trajectory set is stored in an experience cache pool. The experience cache pool is determined to meet a preset capacity requirement, i.e., based on the generated trajectory set, the preset capacity requirement is met. Once the experience cache pool reaches a certain size, the generated strategy data and expert strategy data are passed to the discriminator. The discriminator then evaluates the two strategies, obtains the state reward of the simulated data, and calculates the corrected state-action value function Q by combining it with the real state-action value function. t GAN (τ,θ Z The modified advantage function is then obtained and used to update the generator network.

[0161] Optionally, obtaining the simulated state-action value function based on the Markov process model and the preset adversarial generation strategy algorithm includes:

[0162] The current environmental state is obtained based on the Markov process model.

[0163] Based on the current environmental state, obtain the current time step cost and the current time step backup cost;

[0164] A penalty function is constructed based on the current time step cost and current time step backup cost, preset safe operation constraints, and preset limit violation penalty algorithm, and the total penalty for the current time step is obtained based on the penalty function;

[0165] Construct a reward function based on the current time step cost, the current time step reserve cost, and the current time step total penalty;

[0166] The current state-action value function is obtained based on the reward function and the penalty function.

[0167] In the specific implementation process, the construction process of the simulated state-action value function of the generation strategy based on the adversarial training deep reinforcement learning optimization algorithm (i.e., the preset adversarial generation strategy algorithm) is as follows:

[0168] The reward function primarily reduces system operating costs and serves as an evaluation metric during the action network update process. It reflects the quality of the agent's output actions and guides the direction of policy updates. The penalty function defines the penalty value for violating system constraints in each state, primarily reducing the penalty for exceeding system limits. The process of obtaining the current state-action value function based on the reward and penalty functions is shown in the following equation:

[0169]

[0170] In the formula: R t This represents the reward value at the current time step. Indicates the cost at the current time step. This represents the system's current time step reserve cost, and its value is related to the system's imbalance. Indicates climbing power. N represents the active power output over time period t. coal N represents the number of thermal power units. ess C represents the energy of the energy storage system. ess Indicates the cost of electricity generation from the energy storage system; F t all F represents the total penalty at the current time step. t unb F t ranp F t line F t v This represents the current time step's power imbalance penalty, ramp penalty, line limit violation penalty, and voltage limit violation penalty. α1, α2, α3, α4, α5, α6, and α7 represent the corresponding coefficients used to unify the dimensions of each reward and penalty and establish their proportional relationships; α8 represents the unit cost of standby capacity; α9 = 100, ensuring that the reward value for no limit violation is always greater than the reward value for limit violation; P t sta This indicates the acceptable limit for unbalanced power.

[0171] During training, if the power system meets the preset safe operation constraints based on the current environmental state, the penalty value for each limit violation is 0. If the system violates the safe operation constraints (i.e., does not meet the preset safe operation constraints), the penalty for exceeding the limit is the absolute value of the limit violation. Finally, the total penalty for exceeding the limit is obtained based on the coefficients of each component. When the agent executes actions, the limit violations of thermal power units, energy storage systems, and load reduction actions are limited to within 5%.

[0172] Optionally, updating the generator based on the approximate expert policy probability and the current state-action value function to obtain the current generator includes:

[0173] Get the current correction factor;

[0174] The simulated state reward is obtained based on the approximate expert policy probability and the current correction factor.

[0175] The modified value function is obtained based on the simulated state reward and the current state-action value function;

[0176] A modified advantage function is obtained based on a preset near-end strategy optimization algorithm and the modified value function;

[0177] The gradient loss function is obtained based on the modified advantage function and the preset error gradient backpropagation algorithm.

[0178] The generator is updated based on the gradient loss function to obtain the current generator.

[0179] Optionally, obtaining the current correction factor includes:

[0180] Based on the Markov process model, obtain the instantaneous correction factor, instantaneous return value, and historical average return value of the previous time step;

[0181] The current correction factor is obtained based on the real-time correction factor, real-time return value, historical average return value, and preset correction factor update conditions.

[0182] In the specific implementation process, the generator continuously updates its generation strategy using the corrected signal during the game with the discriminator. Through repeated iterations, the generated strategy continuously approaches and eventually surpasses the expert level, ultimately achieving the preset training objective: the generator network gradually generates an optimized strategy that approximates the expert strategy. During this process, the discriminator continuously judges the generated strategy, and the simulated state-action value function of the generated strategy gradually approaches the real state-action value function. Therefore, the current state-action value function is corrected as the corrected value function shown in the following equation:

[0183]

[0184] In the formula, G r (θ Z Q represents the simulated state reward generated by the generator network; πθ (s t ,a t ) represents the true current state-action value function at time step t; τ represents the current correction factor, used to adjust the magnitude of the GAN's correction to the advantage function estimate. If τ = 0, the model no longer depends on the manually defined reward function, but the correction direction is determined by the expert database.

[0185] The current correction factor τ is adjusted based on whether the reward of the current event is greater than the average reward of the previous m events. The update method of the correction factor τ, obtained based on the preset correction factor update conditions, is as follows:

[0186]

[0187] In the formula, τ t τ represents the instantaneous correction factor at the current time step. t+1 G represents the current correction factor at the next time step. t This represents the immediate reward value at the current time step. This represents the average of the m returns up to the current time step. When the correction factor increases in the next time step, the magnitude of the GAN's correction to the advantage function decreases; when When the correction factor decreases in the next time step, the magnitude of the GAN's correction to the advantage function increases; when At that time, the correction factor for the next time step remains unchanged.

[0188] The simulated state rewards generated by the generator network are as follows:

[0189] G r (θ Z )=log(D(θ Z ) / (1-τD(θ Z )))

[0190] In the formula, D(θ) Z D(θ) represents the approximate expert policy probability of the current generation strategy, with an interval of [0, 1]. As the generator is continuously optimized, the discriminator will find it increasingly difficult to distinguish between the generation strategy and the expert scheduling strategy, which is reflected in D(θ). Z The value of ) approaches 0.5, meaning the simulated state reward G... r (θ Z () Approaching 0.

[0191] This embodiment employs a pre-defined proximal policy optimization algorithm, namely the PPO algorithm, which uses the advantage function to calculate the probability distribution of the agent selecting the next state action, and then uses the corrected... The action value function Q replaces the actual state at the current time step. πθ The corrected advantage function is shown in the following equation:

[0192]

[0193] In the formula, V πθ (s t ) represents the state value function using the PPO algorithm.

[0194] The generator network is trained in the current context through unsupervised learning. Its parameter updates are not directly derived from sample data, but rather employ a pre-defined error gradient backpropagation algorithm. This involves using the backpropagation of the discriminator network to evaluate the simulated state-action value function generated by the generator network. The resulting error gradient of the state-action value function is then backpropagated to the generator network to adjust its weights. Therefore, the gradient loss function of the improved PPO algorithm is shown in the following equation:

[0195]

[0196] In the formula: Let represent the modified advantage function, r(π) represent the ratio of the action probabilities of the current generated policy to the old policy, and ∈ represent the pruning coefficient.

[0197] In the specific implementation process described above, the influence of generative adversarial networks is introduced when calculating the advantage function. Specifically, when calculating the current correction factor, the update method of the correction factor continuously changes based on whether the reward of the current decision step is better than the average reward of the previous m decision steps. The revised advantage function is calculated, which effectively reduces the algorithm's variance, enhances its generalization ability, and reduces the influence of the artificially defined reward function.

[0198] In summary, this embodiment constructs a power system control algorithm model structure based on adversarial training, as follows: Figure 2 As shown, the power system control algorithm model includes a power system environment constructed based on a power system environment model, a discriminator, a generator's action network and evaluation network, an expert database (i.e., the preset expert database), and an experience cache pool; wherein, in this embodiment, the training process based on interactive actions to achieve iterative updates includes: the agent obtaining the current state s at time step t from the simulated power system environment. t , will state s t The input is fed into the action network to obtain the generated scheme a. G,t ;a G,t Earn instant rewards by interacting with the power system environment. t and the state at the next time step s t+1 The completion flag for each cycle is set as mark. t . State s t Input into the expert database (preset expert strategy database) to obtain the corresponding expert solution a. E,t With virtual instant rewards E,t The experience samples (interaction trajectory set) are stored in the experience cache pool, and it is determined whether the experience cache pool meets the preset capacity requirements. After the experience cache pool reaches a certain size, the actions in the experience cache pool and the expert policy are passed to the discriminator; the discriminator evaluates the two policies, obtains the state reward of the simulated data, and calculates the corrected state-action value function by combining it with the real state-action value function. The corrected advantage function is then obtained, and the generator network is updated to obtain the current generator and the current discriminator. When the agent completes the offline training requirements, the action network parameters from the offline training are saved, the center is tested using the test set, and the results are output, ultimately yielding a Markov agent that meets the preset training objectives.

[0199] By implementing this embodiment, generative adversarial networks (GANs) are combined with deep reinforcement learning to construct an adversarial training mechanism between the generator and the discriminator. Combining a dynamic correction factor mechanism, multi-objective reward function design, and a hybrid estimation technique for state-action value functions, the intelligent adjustment mechanism based on the dynamic correction factor uses adversarial training to adaptively adjust the correction magnitude of the advantage function. The correction magnitude is dynamically adjusted based on the comparison between the current reward and the historical average reward, significantly improving training stability and convergence speed. Through the design of a multi-objective reward function, multi-dimensional synergistic optimization of economy, security, and stability is achieved, providing comprehensive decision support for power system prevention and control. The use of a hybrid value function estimation technique effectively integrates real value estimation with generator-simulated rewards, greatly enhancing the algorithm's generalization ability and adaptability.

[0200] This embodiment provides a power system control system based on adversarial training, used to implement the aforementioned power system control method based on adversarial training, including:

[0201] The model building module is used to acquire power system-related data, obtain a power system environment model and a Markov process model based on the power system-related data, and build a generator and a discriminator based on the Markov process model.

[0202] The intelligent agent training module is used to perform interactive actions: It interacts with the power system environment model based on the generator to obtain a set of generated trajectories; when the generated trajectory set determines that a preset capacity requirement is met, it updates the discriminator based on the generated trajectory set to obtain an approximate expert policy probability and a current discriminator; it obtains a current state-action value function based on a preset adversarial generation strategy algorithm of the Markov process model; it updates the generator based on the approximate expert policy probability and the current state-action value function to obtain a current generator; when the approximate expert policy probability determines that the preset training objective is not met, it returns to repeatedly execute the interactive actions based on the current generator and the current discriminator until a corrected generator and a corrected discriminator that meet the preset training objective are obtained.

[0203] The agent deployment module is used to obtain Markov agents based on the correction generator and the correction discriminator.

[0204] The prevention and control module is used to acquire the actual operating data of the power system in real time, and to acquire the power system prevention and control strategy based on the Markov agent and the actual operating data.

[0205] This embodiment provides a power system control system based on adversarial training. A model building module constructs a power system environment model and a Markov process model. Then, an agent training module, through interaction with the power system environment model, performs iterative interactive training and updates the generator and discriminator networks, aiming to make the generator approximate expert strategies. This allows for the efficient acquisition of preventative control strategies adapted to the dynamic operation of the power system, effectively improving the efficiency of power system preventative control. The game-theoretic mechanism of adversarial training enables continuous optimization of the strategy through iterative generation of the generator and discriminator, ensuring the adaptability of strategy decisions to the system state. The final control strategy, while conforming to expert experience, can also cope with the dynamic changes of the power system. This solves the technical problems of low computational efficiency and poor online adaptability in existing technologies, ensuring the efficient acquisition and stable adaptation of preventative control strategies during the dynamic operation of the power system. Finally, an agent deployment module and a preventative control module implement preventative control during the actual operation of the power system. This addresses the technical problems of low computational efficiency, poor online adaptability, low training efficiency, and unstable training in existing technologies.

[0206] Another embodiment of this invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps of the power system control method based on adversarial training as described in this embodiment.

[0207] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A power system control method based on adversarial training, characterized by, The method comprises the following steps: acquiring power system related data, and acquiring a power system environment model and a Markov process model based on the power system related data; constructing a generator and a discriminator based on the Markov process model; performing an interaction action: interacting with the power system environment model based on the generator, and acquiring a generated trajectory set; when it is determined that the generated trajectory set meets a preset capacity requirement, updating the discriminator based on the generated trajectory set, and acquiring an approximate expert policy probability and a current discriminator; acquiring a current state-action value function based on a preset adversarial generation strategy algorithm of the Markov process model; updating the generator based on the approximate expert policy probability and the current state-action value function, and acquiring a current generator; when it is determined that the approximate expert policy probability does not meet a preset training target, returning to repeat the interaction action based on the current generator and the current discriminator until a modified generator and a modified discriminator that meet the preset training target are acquired; acquiring a Markov agent based on the modified generator and the modified discriminator; acquiring a power system preventive control strategy based on the Markov agent and actual operation data of the power system.

2. The adversarial training based power system control method of claim 1, wherein, The method of acquiring power system related data and acquiring a power system environment model and a Markov process model based on the power system related data comprises the following steps: acquiring power system related data, and constructing a power system environment model of the power system based on the power system related data by using a time domain simulation method; acquiring a cost function based on the power system related data; constructing a constraint optimization objective function based on a preset optimization objective, the cost function and a preset constraint condition; acquiring a power system control model based on the power system environment model, a differential algebraic equation and the constraint optimization objective function; acquiring a Markov process model of the power system based on the power system control model.

3. The adversarial training based power system control method of claim 2, wherein, The method of acquiring a Markov process model of the power system based on the power system control model comprises the following steps: for any current time step: acquiring a current attribute state, a current topology state and current context information based on the power system control model; acquiring a current environment state vector based on the current attribute state, the current topology state and the current context information; acquiring a current action vector based on the current environment state vector, and adding the current environment state vector and the current action vector to the current context information to acquire next time step context information, so as to acquire a next time step environment state vector based on the next time step context information; constructing an environment state vector set based on a plurality of current environment state vectors of any current time step; constructing an action vector set based on a preset power adjustment rule; acquiring a state transition probability based on the power system control model and a preset transition probability calculation method; constructing an immediate reward function based on the power system control model and a preset reward algorithm; acquiring a Markov process model of the power system based on the environment state vector set, the action vector set, the state transition probability, the immediate reward function and a preset decay factor.

4. The adversarial training based power system control method of claim 1, wherein, The method of performing an interaction action comprises the following steps: The generator comprises an action network and an evaluation network; The generated trajectory set comprises a current time step generated trajectory of all time steps; For any time step: An environment state of a current time step is obtained; An environment state value estimation of the current time step is obtained based on the environment state of the current time step and the evaluation network; The environment state of the current time step and the environment state value estimation of the current time step are input into the action network to obtain a current generated scheme; The next time step environment state and the current time step immediate reward are obtained by interacting with the power system environment model based on the current generated scheme; The current time step expert scheme and the virtual immediate reward are obtained by inputting the environment state of the current time step into a preset expert policy library; The current time step generated trajectory is obtained based on the environment state of the current time step, the current generated scheme, the next time step environment state, the current time step expert scheme and the virtual immediate reward.

5. The adversarial training based power system control method of claim 1, wherein, When it is determined that the generated trajectory set meets the preset capacity requirement, the discriminator is updated based on the generated trajectory set to obtain an approximate expert policy probability and a current discriminator, comprising: The generated policy data and the expert policy data are obtained based on the generated trajectory set; The generated output probability is obtained by inputting the generated policy data into the discriminator; The expert output probability is obtained by inputting the expert policy data into the discriminator; The approximate expert policy probability is obtained based on the generated output probability and the expert output probability; The current discriminator is obtained by updating the discriminator based on the approximate expert policy probability.

6. The adversarial training based power system control method of claim 1, wherein, The simulated state-action value function is obtained based on the Markov process model and a preset adversarial generated policy algorithm, comprising: The current environment state is obtained based on the Markov process model; The current time step cost and the current time step reserve cost are obtained based on the current environment state; The penalty function is constructed based on the current time step cost, the current time step reserve cost, a preset safe operation constraint condition and a preset out-of-limit penalty algorithm, and the current time step total penalty is obtained based on the penalty function; The reward function is constructed based on the current time step cost, the current time step reserve cost and the current time step total penalty; The current state-action value function is obtained based on the reward function and the penalty function.

7. The adversarial training based power system control method of claim 1, wherein, The generator is updated based on the approximate expert policy probability and the current state-action value function to obtain a current generator, comprising: A current correction factor is obtained; The simulated state reward is obtained based on the approximate expert policy probability and the current correction factor; The correction value function is obtained based on the simulated state reward and the current state-action value function; The correction advantage function is obtained based on a preset proximal policy optimization algorithm and the correction value function; The gradient loss function is obtained based on the correction advantage function and a preset error gradient back propagation algorithm; The generator is updated based on the gradient loss function to obtain the current generator.

8. The adversarial training based power system control method of claim 7, wherein, The current correction factor is obtained, comprising: The immediate correction factor, the immediate return value and the historical return average value of the previous time step are obtained based on the Markov process model; The current correction factor is obtained based on the instant correction factor, the instant return value, the historical return average value, and a preset correction factor update condition.

9. A power system control system based on adversarial training, characterized in that, The power system control method based on adversarial training according to any one of claims 1-8 is implemented by comprising: a model construction module, configured to obtain power system related data, obtain a power system environment model and a Markov process model based on the power system related data, and construct a generator and a discriminator based on the Markov process model; an agent training module, configured to perform an interaction action: interact with the power system environment model based on the generator to obtain a generated trajectory set; when it is determined that the generated trajectory set meets a preset capacity requirement, update the discriminator based on the generated trajectory set to obtain an approximate expert strategy probability and a current discriminator; obtain a current state-action value function based on a preset adversarial generation strategy algorithm of the Markov process model; update the generator based on the approximate expert strategy probability and the current state-action value function to obtain a current generator; when it is determined that the approximate expert strategy probability does not meet a preset training target, return to repeat the interaction action based on the current generator and the current discriminator until a corrected generator and a corrected discriminator meeting the preset training target are obtained; an agent deployment module, configured to obtain a Markov agent based on the corrected generator and the corrected discriminator; a preventive control module, configured to obtain actual operation data of the power system in real time, and obtain a power system preventive control strategy based on the Markov agent and the actual operation data.

10. A terminal device, comprising: The power system control method based on adversarial training according to any one of claims 1-8 is implemented by comprising: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the power system control method based on adversarial training according to any one of claims 1-8 when executing the computer program.

Images