Method and system for controlling transient voltage of power system

Abstract

The invention provides a power system transient voltage control method and system. The method comprises the following steps: acquiring initial power system operation state parameters in real time and establishing a transient model; obtaining an initial transient voltage severity index based on the parameters and a preset initial voltage; constructing a physical information neural network based on the transient model and exporting a loss function of the physical information neural network; based on the parameters, a preset reward parameter and a noise network, obtaining an initial Rainbow agent, executing a removal action and calculating corresponding loss; and constructing a total loss function based on the total loss function and training to convergence by taking minimization as a target, and obtaining an intelligent agent to realize the control of the transient voltage of the power system. According to the method and the system for controlling the transient voltage of the electric power system, the electric power system can be rapidly controlled to recover stability, the situation that the transient voltage control decision is not accurate due to the fact that the action value is excessively estimated is avoided, and the situation that the instability degree of the electric power system is aggravated is further prevented.

Description

Technical Field

[0001] This invention relates to the field of power system stability control technology, and in particular to a method and system for controlling transient voltage in a power system. Background Technology

[0002] With the continuous expansion of the power grid and the in-depth advancement of the energy internet construction, long-distance and large-capacity power exchange between regional power networks is becoming increasingly frequent, significantly increasing the probability of short-circuit faults and posing a severe challenge to network voltage stability. To address this, some scholars have proposed a transient voltage control method that integrates physical information and reinforcement learning. By combining a power system physical model with a data-driven model, physical model constraints are integrated into the data-driven model to construct Physical-Informed Reinforcement Learning (PIRL) for efficient transient voltage control. However, the Deep Q-Network (DQN) used in this method still suffers from Q-value overestimation and low sampling data utilization efficiency, leading to a sharp decline in agent performance in large-scale state-action space systems. Therefore, it is urgent to propose a scheme that can effectively solve the problem of Q-value overestimation and optimize sampling data utilization.

[0003] Under the current technological background, the DQN algorithm has several shortcomings: First, it uses a greedy algorithm to update action values, which can easily cause maximization bias and lead to overestimation of Q value. Overestimation during Q-value updates can lead to a preference for actions that are overestimated, resulting in excessively high target Q-value errors. Secondly, DQN directly outputs the Q-value of each action through a neural network. When the state space is large, the efficiency of accurately learning action Q-values ​​is low, and the generalization ability is limited. Furthermore, Q-value estimation may become unstable in scenarios with large action spaces or complex environments. Thirdly, experience replay involves uniform random sampling from the buffer, meaning that all experiences have the same probability of being selected. However, different error experiences contribute differently to the agent's learning. Experiences with large errors (such as those that cause significant changes in reward or are key nodes in state transitions) contain more important information. Uniform sampling can prevent these high-value experiences from being fully utilized, leading to long training cycles, increased computational resource requirements, and difficulty in applying them to real-world scenarios where samples are expensive. Finally, the exploration ε-greedy strategy is a random exploration based on hyperparameters, randomly selecting actions with a fixed probability. It is a relatively simple "local jitter" perturbation, with relatively simple exploration behavior, making it difficult to generate the large-scale behavioral patterns required for effective exploration in complex environments. Summary of the Invention

[0004] The present invention aims to provide a method and system for controlling transient voltage in a power system to solve the above-mentioned technical problems. It can quickly control the power system to restore stability after a short-circuit fault occurs, while avoiding the overestimation of the action value, which would lead to inaccurate transient voltage control decisions, thereby preventing the power system from becoming more unstable.

[0005] To address the aforementioned technical problems, this invention provides a method for controlling transient voltage in a power system, comprising:

[0006] Real-time acquisition of the initial power system operating status parameters corresponding to the initial power system;

[0007] A transient model of the power system is established based on the initial power system operating state parameters; the severity index of the initial transient voltage is obtained based on the initial power system operating state parameters and the preset initial voltage.

[0008] Based on the transient model of the power system, a physical information neural network is constructed and its corresponding neural network loss function is derived.

[0009] Based on the initial power system operating state parameters, the initial transient voltage severity index, the preset reward parameters, and the preset noise network, the initial Rainbow agent is obtained.

[0010] Based on the initial Rainbow agent, the initial power system is disconnected, and the corresponding Rainbow loss is calculated;

[0011] Construct a total loss function based on Rainbow loss and neural network loss function;

[0012] The initial Rainbow agent is trained with the goal of minimizing the total loss function until the preset convergence condition is met, thereby obtaining the Rainbow agent to control the transient voltage of the power system.

[0013] In the above scheme, constructing a physical information neural network (PEN) using a power system transient model and deriving its corresponding loss function ensures that the PSN's predictions follow physical laws. Simultaneously, the loss function quantifies the data deviation of the predictions, providing a basis for subsequent agent training and error correction. Next, by using an initial Rainbow agent to perform a disconnection action on the initial power system and calculating the corresponding Rainbow loss, the deviation between the agent's current action decision and the optimal control objective is quantified, reflecting the agent's decision-making effectiveness in the transient voltage control process. This provides decision error data for subsequently constructing the total loss function and training the agent. Finally, by constructing the total loss function, a comprehensive measure is taken of the agent's decision deviation and the PSN's prediction deviation, ensuring that the trained agent meets both the control requirements at the decision-making level and conforms to physical constraints. Finally, the initial Rainbow agent is trained with the goal of minimizing the total loss function until the preset convergence condition is met. The agent is continuously optimized so that the final Rainbow agent has accurate and efficient transient voltage control capabilities. This enables the power system to be quickly controlled to restore stability after a short-circuit fault occurs. At the same time, it avoids the overestimation of the action value, which could lead to inaccurate transient voltage control decisions and thus prevent the power system from becoming more unstable.

[0014] Furthermore, the step of constructing a physical information neural network based on a power system transient model and deriving its corresponding neural network loss function includes:

[0015] Based on the transient model of the power system, an initial physical information neural network is constructed.

[0016] Based on the initial physical information neural network, its corresponding initial neural network loss function is derived.

[0017] The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

[0018] In the above scheme, by constructing an initial physical information neural network, the physical laws of the power system transient process can be embedded into the neural network structure, providing an initial network framework for accurate prediction of the power system state at any time after a fault and for transient voltage analysis under physical constraints. Subsequently, by deriving the corresponding initial neural network loss function, the deviation between the initial network prediction and the actual data, as well as the deviation of the initial network prediction from violating the physical laws of the power system, can be quantified, providing direction for adjusting the physical information neural network parameters and optimizing its performance during subsequent training. Next, the initial physical information neural network is trained with the goal of minimizing the initial neural network loss function, continuously iteratively updating the physical information neural network parameters until a preset convergence condition is reached. This yields a physical information neural network that conforms to physical laws and has high prediction accuracy, and its corresponding neural network loss function is derived, providing reliable physical constraint model support for subsequent integration with the Rainbow agent, construction of the total loss function, and implementation of transient voltage control.

[0019] Furthermore, the process of obtaining the initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and preset noise network includes:

[0020] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed.

[0021] The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

[0022] In the above scheme, an initial Rainbow agent is constructed using initial power system operating state parameters, an initial transient voltage severity index, and preset reward parameters. This allows for the creation of a decision-making framework that meets the requirements of power system transient voltage control scenarios, laying the foundation for subsequently integrating a preset noise network to form a complete initial Rainbow agent. Next, by adding a preset noise network to the initial Rainbow agent, the agent's exploration strategy can be optimized, giving the acquired initial Rainbow agent more efficient exploration capabilities. This provides support for subsequent execution of transient voltage control cutoff actions and achieving accurate decision-making.

[0023] Furthermore, the construction of the original Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters includes:

[0024] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established.

[0025] Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network;

[0026] The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

[0027] In the above scheme, by establishing an initial state space, the agent can accurately perceive the initial operating state of the power system; by establishing a reward function, a clear input and feedback logic can be defined for the agent's subsequent action selection and training. Next, by constructing an estimation network, the action value (Q-value) of each action in the current state can be evaluated in real time; by constructing a target network, a relatively stable target Q-value can be provided to avoid error propagation during estimation network updates; through the cooperation of the estimation network and the target network, the problem of Q-value overestimation in the traditional DQN algorithm can be solved, and by combining the input dimension of the initial state space and the output dimension of the preset action space, the network structure is ensured to match the requirements of the power system's transient voltage control scenario.

[0028] Further, the step of obtaining the initial Rainbow agent based on the original Rainbow agent and a preset noise network includes:

[0029] Based on the original Rainbow agent and the preset noise network, obtain the first action value set;

[0030] Select the action value in the first action value set that satisfies the preset power system constraints and the preset conditions as the first action value, and execute the cut-off action corresponding to the first action value to obtain the state space;

[0031] Based on the original Rainbow agent and state space, obtain the initial Rainbow agent.

[0032] In the above scheme, by obtaining a set of action values ​​(Q-values) through a pre-set noise network, the exploration range of action values ​​can be expanded based on the original agent's action values. This retains the basic decision-making logic of the original Rainbow agent while solving the problem of the single exploration behavior of the traditional DQN algorithm, providing a Q-value basis for subsequent selection of actions that meet the needs of the power system. Next, by selecting action values ​​from the first set of action values ​​that satisfy the pre-set power system constraints and pre-set conditions, it is ensured that the selected actions meet the conditions and are used as the first action value. The corresponding cut-off action is then executed to obtain the state space after the action is executed, thus providing real dynamic state data support for the agent's subsequent policy updates and state perception. Finally, by integrating the state space obtained after executing the cut-off action into the decision framework of the original Rainbow agent, the agent can adjust its internal parameters based on this state space, ultimately forming an initial Rainbow agent with noise exploration capabilities, constraint selection capabilities, and dynamic state adaptation capabilities.

[0033] Furthermore, the step of performing a disconnection action on the initial power system based on the initial Rainbow agent and calculating the corresponding Rainbow loss includes:

[0034] Calculate the reward value based on the initial Rainbow agent and reward function;

[0035] The initial state space, the cut-off action corresponding to the value of the first action, the reward value, and the state space are encapsulated and stored in a preset experience buffer pool. The experience buffer pool is then obtained.

[0036] A preset number of samples are drawn from the experience buffer pool and input into the estimation network to obtain the second action value set corresponding to the preset number of samples.

[0037] Based on the state space and target network, obtain the value of the target action;

[0038] Based on the second action value set and the target action value, calculate the corresponding Rainbow loss.

[0039] In the above scheme, based on the actual state of the power system after the agent performs a cut-off action, the effect of the action is quantified according to the reward function rule to calculate the reward value, providing crucial feedback for subsequent experience storage, Rainbow loss calculation, and agent policy optimization. Next, by encapsulating the initial state space, the cut-off action corresponding to the first action value, the reward value, and the state space and storing them in a preset experience buffer pool, a complete "state-action-reward-new state" experience data is stored during the interaction between the agent and the power system. This provides data reserves for the subsequent priority experience replay mechanism, while avoiding the fragmented loss of experience data and ensuring the integrity and usability of subsequent training data. Finally, a preset number of samples are extracted from the experience buffer pool and input into the estimation network. The estimation network evaluates the action value corresponding to the extracted samples and, combined with the target action value provided by the target network, calculates the corresponding Rainbow loss to quantify the deviation between the current agent policy and the optimal policy, providing core data support for subsequently constructing the total loss function and updating the agent network parameters.

[0040] This invention provides a control system for transient voltage in a power system, comprising a state parameter acquisition module, a transient model construction module, a neural network construction module, an agent construction module, an agent loss calculation module, a total loss function construction module, and an agent training module, specifically:

[0041] The status parameter acquisition module is used to acquire the initial power system operating status parameters corresponding to the initial power system in real time.

[0042] The transient model construction module is used to establish a power system transient model based on the initial power system operating state parameters; and to obtain the initial transient voltage severity index based on the initial power system operating state parameters and the preset initial voltage.

[0043] The neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function;

[0044] The agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters and preset noise network.

[0045] The agent loss calculation module is used to perform a disconnection action on the initial power system based on the initial Rainbow agent and calculate the corresponding Rainbow loss.

[0046] The total loss function construction module is used to construct the total loss function based on Rainbow loss and neural network loss function;

[0047] The agent training module is used to train the initial Rainbow agent with the goal of minimizing the total loss function until a preset convergence condition is met, thereby obtaining the Rainbow agent to control the transient voltage of the power system.

[0048] This invention provides a control system for transient voltage in a power system. In practical applications, it only requires a neural network construction module. A physical information neural network is constructed using a power system transient model, and its corresponding neural network loss function is derived. This ensures that the prediction results of the physical information neural network follow physical laws, and the neural network loss function quantifies the data deviation of the prediction results, providing a basis for subsequent agent training and error correction. Next, an agent loss calculation module is used. An initial Rainbow agent performs a disconnection action on the initial power system and calculates the corresponding Rainbow loss. This quantifies the deviation between the agent's current action decision and the optimal control objective, reflecting the decision-making effect of the agent in the transient voltage control process and providing decision error data for subsequent construction of the total loss function and agent training. Finally, a total loss function construction module is used to comprehensively measure the agent's decision deviation and the physical information neural network prediction deviation, ensuring that the trained agent meets both the control requirements at the decision-making level and conforms to physical law constraints. Finally, an agent training module is used to train the initial Rainbow agent with the goal of minimizing the total loss function until the preset convergence condition is reached. The agent is continuously optimized so that the final Rainbow agent has accurate and efficient transient voltage control capabilities. This enables the power system to be quickly controlled to restore stability after a short-circuit fault, while avoiding overestimation of the action value, which could lead to inaccurate transient voltage control decisions and thus prevent the power system from becoming more unstable.

[0049] Furthermore, the neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function; including:

[0050] Based on the transient model of the power system, an initial physical information neural network is constructed.

[0051] Based on the initial physical information neural network, its corresponding initial neural network loss function is derived.

[0052] The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

[0053] In the above scheme, by constructing an initial physical information neural network, the physical laws of the power system transient process can be embedded into the neural network structure, providing an initial network framework for accurate prediction of the power system state at any time after a fault and for transient voltage analysis under physical constraints. Subsequently, by deriving the corresponding initial neural network loss function, the deviation between the initial network prediction and the actual data, as well as the deviation of the initial network prediction from violating the physical laws of the power system, can be quantified, providing direction for adjusting the physical information neural network parameters and optimizing its performance during subsequent training. Next, the initial physical information neural network is trained with the goal of minimizing the initial neural network loss function, continuously iteratively updating the physical information neural network parameters until a preset convergence condition is reached. This yields a physical information neural network that conforms to physical laws and has high prediction accuracy, and its corresponding neural network loss function is derived, providing reliable physical constraint model support for subsequent integration with the Rainbow agent, construction of the total loss function, and implementation of transient voltage control.

[0054] Furthermore, the agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and a preset noise network; including:

[0055] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed.

[0056] The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

[0057] In the above scheme, an initial Rainbow agent is constructed using initial power system operating state parameters, an initial transient voltage severity index, and preset reward parameters. This allows for the creation of a decision-making framework that meets the requirements of power system transient voltage control scenarios, laying the foundation for subsequently integrating a preset noise network to form a complete initial Rainbow agent. Next, by adding a preset noise network to the initial Rainbow agent, the agent's exploration strategy can be optimized, giving the acquired initial Rainbow agent more efficient exploration capabilities. This provides support for subsequent execution of transient voltage control cutoff actions and achieving accurate decision-making.

[0058] Furthermore, the construction of the original Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters includes:

[0059] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established.

[0060] Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network;

[0061] The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

[0062] In the above scheme, by establishing an initial state space, the agent can accurately perceive the initial operating state of the power system; by establishing a reward function, a clear input and feedback logic can be defined for the agent's subsequent action selection and training. Next, by constructing an estimation network, the action value (Q-value) of each action in the current state can be evaluated in real time; by constructing a target network, a relatively stable target Q-value can be provided to avoid error propagation during estimation network updates; through the cooperation of the estimation network and the target network, the problem of Q-value overestimation in the traditional DQN algorithm can be solved, and by combining the input dimension of the initial state space and the output dimension of the preset action space, the network structure is ensured to match the requirements of the power system's transient voltage control scenario. Attached Figure Description

[0063] Figure 1 A flowchart illustrating a method for controlling transient voltage in a power system, as provided in an embodiment of the present invention;

[0064] Figure 2 This is an architecture diagram of a power system transient voltage control system provided in an embodiment of the present invention. Detailed Implementation

[0065] 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, and 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.

[0066] This embodiment provides a method for controlling transient voltage in a power system. Please refer to the flowchart below. Figure 1 ,include:

[0067] Step S1: Obtain the initial power system operating status parameters corresponding to the initial power system in real time;

[0068] Step S2: Based on the initial power system operating state parameters, establish a power system transient model; based on the initial power system operating state parameters and the preset initial voltage, obtain the initial transient voltage severity index;

[0069] Step S3: Based on the power system transient model, construct a physical information neural network and derive its corresponding neural network loss function;

[0070] Step S4: Based on the initial power system operating state parameters, the initial transient voltage severity index, the preset reward parameters, and the preset noise network, obtain the initial Rainbow agent;

[0071] Step S5: Based on the initial Rainbow agent, perform a disconnection action on the initial power system and calculate the corresponding Rainbow loss;

[0072] Step S6: Construct the total loss function based on Rainbow loss and neural network loss function;

[0073] Step S7: Train the initial Rainbow agent with the goal of minimizing the total loss function until the preset convergence condition is met, and obtain the Rainbow agent to realize the control of the transient voltage of the power system.

[0074] In this embodiment, the initial power system operating state parameters corresponding to the initial power system are acquired in real time, wherein the initial power system operating state parameters include: the mechanical power P of generator i. Ti The electromagnetic power P of generator i Ei The rotor speed ω of generator i relative to the center of inertia i The rotor angle δ of motor i relative to the center of inertia i The active power P of the j-th bus dj The reactive power Q of the j-th bus dj The voltage V of the j-th bus j The actual bus voltage V at time t t Synchronous speed constant ω0, generator i's inertia coefficient T Ji The generator damping coefficient D is typically set to 2.0. This allows for the construction of a transient physical model of the power system after a short-circuit fault.

[0075]

[0076] Based on the initial power system operating state parameters and the preset initial voltage, obtain the initial transient voltage severity index: Where TVSI represents the transient voltage severity index, and T is the transient period; T c This is the fault clearing time; TVDI t It is the transient voltage deviation index, calculated from data obtained after time-domain simulation of the power system, and the formula is as follows:

[0077]

[0078] In the formula, V0 represents the preset initial voltage, V t TVSI represents the actual bus voltage at time t, and η represents the preset deviation threshold, typically set to 0.2. A higher TVSI indicates a more unstable power system. Additionally, V is set... t At the end of the control operation, the voltage returns to the range of 0.7 pu to 1.2 pu, indicating that the actual bus voltage at that moment is considered stable. Then, a physical information neural network is constructed using the power system transient model, and its corresponding neural network loss function is derived. This ensures that the prediction results of the physical information neural network follow physical laws. Simultaneously, the neural network loss function quantifies the data deviation of the prediction results and can be added to the Rainbow loss to form the total loss function, accelerating the training of subsequent Rainbow agents and providing a basis for error correction. Next, the initial Rainbow agent performs a disconnection action on the initial power system, and the corresponding Rainbow loss L is calculated. Rainbow This can quantify the deviation between the agent's current action decision and the optimal control objective, reflecting the agent's decision-making performance during transient voltage control, and providing decision error data for subsequent construction of the total loss function and agent training. Subsequently, the total loss function is constructed: L total =L Rainbow +ρ PINN MSE PINN In the formula, ρ PINN MSE represents the coefficients of the neural network loss function corresponding to the physical information neural network. PINN This represents the neural network loss function corresponding to the physical information neural network. By constructing a total loss function, we comprehensively measure the decision bias of the agent and the prediction bias of the physical information neural network, ensuring that the trained agent meets both the control requirements at the decision level and conforms to the constraints of physical laws. Finally, we train the initial Rainbow agent with the goal of minimizing the total loss function, while updating the target network weights in the Rainbow agent using gradient descent until the preset convergence condition is reached. This process continuously optimizes the agent, enabling the final Rainbow agent to possess precise and efficient transient voltage control capabilities. This allows for rapid control of the power system to restore stability after a short-circuit fault (when a short-circuit fault occurs in the power system, the power bus is disconnected through actions until the transient voltage stabilizes within the range of 0.7 pu to 1.2 pu). Simultaneously, it avoids overestimating the value of actions, which could lead to inaccurate transient voltage control decisions and thus prevent the power system from becoming increasingly unstable.

[0079] Furthermore, the step of constructing a physical information neural network based on a power system transient model and deriving its corresponding neural network loss function includes:

[0080] Based on the transient model of the power system, an initial physical information neural network is constructed.

[0081] Based on the initial physical information neural network, its corresponding initial neural network loss function is derived.

[0082] The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

[0083] In this embodiment, since the dynamic equations of a dynamic system can be expressed in the general form of parameterized and nonlinear partial differential equations, therefore: Where u(t,x) represents the hidden solution, x is the input vector (the rotor angle obtained in the above steps), and N[u] represents a nonlinear operator applied to state u to describe the nonlinear dynamic characteristics of the power system. It can be used to encapsulate a series of physical mathematical formulas (i.e., Therefore, u(t,x)=δ(t); In order to describe the dynamic characteristics of a highly dynamic power system, a deep neural network f(t,x) is constructed based on the transient model of the power system to approximate u, which can be given by the left side of the differential equation: Right now The neural network is fed into the observation vector x and time t. Automatic differentiation is used to process some differential equations, calculating the partial derivatives of u with respect to t and x. u(t,x) and f(t,x) constitute the initial physical information neural network, having the same network parameters but different activation functions. u(t,x) outputs the predicted value of the initial physical information neural network, while f(t,x) outputs the deviation value of the initial physical information neural network under physical constraints. By constructing the initial physical information neural network, the physical laws of the power system transient process can be embedded into the neural network structure, providing an initial network framework for subsequent accurate prediction of the power system state at any time after a fault and for realizing transient voltage analysis under physical constraints. Subsequently, by deriving the corresponding initial neural network loss function, the deviation between the initial network prediction value and the actual data, as well as the deviation of the initial network prediction value from the physical laws of the power system, can be quantified. This provides direction for adjusting the parameters and optimizing the performance of the physical information neural network during subsequent training.

[0084] MSE PINN =MSE u +MSE f

[0085]

[0086] Among them, MSE PINN It is the initial neural network loss function, MSE. uIt is the data error of the initial physical information neural network (the deviation between the predicted values ​​of the initial physical information neural network and the actual data), MSE f It is the physical information error of the initial physical information neural network (the deviation of the predicted value of the initial physical information neural network under physical constraints); N u It is the total number of valid samples used to calculate data error. It is the time point corresponding to the nth data sample. It is the input vector of the nth data sample, u n It is the real time-domain simulation data of the nth data sample, used as a benchmark for comparing the model's predicted values; N f It is the total number of valid samples used to calculate the error in physical information. It is the time point corresponding to the nth data sample. This is the input vector of the nth data sample. During training, the calculated error value is fed back to the optimizer to update the neural network parameters until the preset convergence condition (MSE) is reached. PINN =0), obtain physical information neural networks that conform to physical laws and have high prediction accuracy, and derive their corresponding neural network loss function, providing reliable physical constraint model support for subsequent fusion with Rainbow agent, construction of total loss function and realization of transient voltage control.

[0087] Furthermore, the process of obtaining the initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and preset noise network includes:

[0088] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed.

[0089] The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

[0090] In this embodiment, an initial Rainbow agent is constructed using initial power system operating state parameters, an initial transient voltage severity index, and preset reward parameters. This allows for the creation of a decision framework that meets the requirements of power system transient voltage control scenarios, laying the foundation for subsequently integrating a preset noise network to form a complete initial Rainbow agent. Next, by adding a preset noise network to the initial Rainbow agent, the agent's exploration strategy is optimized, enabling the acquired initial Rainbow agent to possess more efficient exploration capabilities. This provides support for subsequent execution of transient voltage control cutoff actions and the achievement of accurate decision-making.

[0091] Furthermore, the construction of the original Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters includes:

[0092] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established.

[0093] Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network;

[0094] The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

[0095] In this embodiment, voltage control after a short-circuit fault in the power system is modeled as a Markov decision process (i.e., state space, preset action space, and reward function). Specifically, an initial state space is established based on the initial power system operating state parameters: S = {P} Ti ,P Ei ,δ i ,ω i ,P dj Q dj V j ,t};The mechanical power P of generator i Ti The electromagnetic power P of generator i Ei The rotor speed ω of generator i relative to the center of inertia i The rotor angle δ of motor i relative to the center of inertia i The active power P of the j-th bus dj The reactive power Q of the j-th bus dj The voltage V of the j-th bus j By establishing an initial state space, the intelligent agent can accurately perceive the initial operating state of the power system. The preset action space is α = {ΔP}. dj (t)},ΔP dj (t) represents the amount of power removed from the load bus at time t. A reward function is established based on the initial power system operating state parameters, the initial transient voltage severity index, and preset reward parameters:

[0096]

[0097] In the formula, ep represents the number of load reductions. This represents the severity index of the transient voltage of the j-th bus at time t. Preset reward parameters include: r pun It is the penalty coefficient, r stepC1 and C2 are the frequency reward coefficients, and their respective weight coefficients, designed based on empirical values ​​during initial training. These can be adjusted during subsequent training to ensure good agent performance. By establishing a reward function, clear input and feedback logic can be defined for the agent's subsequent action selection and training. If the power system remains unstable (i.e., transient voltage exceeds 0.7 pu to 1.2 pu) after a load shedding operation, the reward is set to negative, related to the number of load shedding operations, to increase the penalty. When the power system becomes temporarily stable after a load shedding operation, the load shedding action is evaluated through three parts: 1. Effectiveness: defined by the effective load shedding degree, i.e., TVSI; 2. Total load shedding; 3. Load shedding count reward, with a larger reward for more load shedding operations. An estimation network and a target network are constructed based on the initial state space, preset action space, and preset network parameters. The estimation network splits the output layer of the original Rainbow agent into an action advantage sub-network and a state value sub-network. In the formula, Q(s,a,w) represents the action value (Q-value), A(s,a,w) represents the action advantage subnetwork, and V(s,w) represents the state value subnetwork. Let represent the average of all action advantage values, a′ represent each action in state s, and w represent the parameters of the action advantage subnetwork. The target network updates y using the TD Target formula. t =r t +γ·argmaxQ′(s t+1 ,a′,w′),r t γ is the immediate reward obtained after performing the action at time step t; γ is the discount factor, typically ranging from 0 to 1; s t+1 The state variable at time step t+1 represents the new environment state after the action is performed, and a′ represents the next state s. t+1The variables for all possible actions are defined, with w' being the parameters of the Q' function. This ensures that if an agent experiences an overestimation error in any network during subsequent training, the error will not propagate to another network, improving the stability and speed of the algorithm's learning. Furthermore, by acquiring a set of action values ​​(Q-values), the exploration scope of action values ​​can be expanded based on the original agent's action values. This retains the basic decision-making logic of the original Rainbow agent while solving the problem of limited exploration behavior in traditional DQN algorithms, providing a Q-value basis for subsequent selection of actions that meet the needs of the power system. By constructing an estimation network, the action value (Q-value) of each action in the current state can be evaluated in real time; by constructing a target network, a relatively stable target Q-value can be provided to avoid error propagation during estimation network updates; through the cooperation of the estimation network and the target network, the problem of Q-value overestimation in traditional DQN algorithms can be solved. Simultaneously, by combining the input dimension of the initial state space with the output dimension of the preset action space, the network structure is ensured to match the scenario requirements of transient voltage control in the power system.

[0098] Further, the step of obtaining the initial Rainbow agent based on the original Rainbow agent and a preset noise network includes:

[0099] Based on the original Rainbow agent and the preset noise network, obtain the first action value set;

[0100] Select the action value in the first action value set that satisfies the preset power system constraints and the preset conditions as the first action value, and execute the cut-off action corresponding to the first action value to obtain the state space;

[0101] Based on the original Rainbow agent and state space, obtain the initial Rainbow agent.

[0102] In this embodiment, a cutoff action is selected through a preset noise network, and an additional parameterized noise is added to each weight of the fully connected layer of the original Rainbow agent, as shown below:

[0103]

[0104] In the formula, y represents the set of first action values ​​after adding parameterized noise, and x represents the state variables s in the initial state space of the original Rainbow agent. t W is the basic weight matrix of the Rainbow agent, b is the bias, and W noisy It is the noise weight matrix, b noisy It is the bias vector of the noise, ε ω and ε bThese are two random noise variables. The noise can adaptively adjust its magnitude through continuous learning by the noise network, which is a greedy exploration method. Compared to the ε-greedy exploration strategy used in the DQN algorithm, it introduces more uncertainty during exploration, expanding the breadth and depth of the agent's exploration and improving exploration efficiency. Next, the action value that satisfies the preset power system constraints and preset conditions from the action value set is selected as the first action value, and the corresponding cut-off action a is executed. t The new state variable s is obtained. t+1 To obtain the corresponding new state space, and the cut-off action corresponding to the first action value is stored in the action space α={ΔP dj In (t)}, the original Rainbow agent is initialized. Since the state space is a one-dimensional vector, a fully connected network is chosen. Next, by selecting action values ​​from the action value set that satisfy the preset power system constraints and preset conditions, it is ensured that the selected actions meet the conditions and are used as the first action value. The corresponding cut-off action is then executed to obtain the state space after the action is executed, thus providing real dynamic state data support for the agent's subsequent policy updates and state perception. Finally, by integrating the state space obtained after executing the cut-off action into the decision framework of the original Rainbow agent, the agent can adjust its internal parameters based on this state space, ultimately forming an initial Rainbow agent with noise exploration capabilities, constraint screening capabilities, and dynamic state adaptation capabilities.

[0105] Furthermore, the step of performing a disconnection action on the initial power system based on the initial Rainbow agent and calculating the corresponding Rainbow loss includes:

[0106] Calculate the reward value based on the initial Rainbow agent and reward function;

[0107] The initial state space, the cut-off action corresponding to the value of the first action, the reward value, and the state space are encapsulated and stored in a preset experience buffer pool. The experience buffer pool is then obtained.

[0108] A preset number of samples are drawn from the experience buffer pool and input into the estimation network to obtain the second action value set corresponding to the preset number of samples.

[0109] Based on the state space and target network, obtain the value of the target action;

[0110] Based on the second action value set and the target action value, calculate the corresponding Rainbow loss.

[0111] In this embodiment, the goal of the control process is to maximize the cumulative reward, and the optimal policy of the Rainbow agent is learned through the Rainbow DQN algorithm. The actual state s of the power system after the agent executes the disconnection action is then considered. t+1 The effect of the action is quantified according to the reward function rules described above, thereby calculating the reward value r. t+1 This provides crucial feedback for subsequent experience storage, Rainbow loss calculation, and agent policy optimization. Next, by encapsulating the initial state space, the cut-off action corresponding to the first action value, the reward value, and the state space, we obtain [s]. t ,a t ,r t+1 ,s t+1 The data is then stored in a pre-defined experience buffer pool, which stores complete experience data from "state-action-reward-new state" during the interaction between the agent and the power system. This provides data reserves for the subsequent priority experience replay mechanism, while avoiding the loss of scattered experience data and ensuring the integrity and availability of subsequent training data. The TD error (ξ) is calculated for each newly stored sample. g =Q(s) t ,a t ,w)-y t , where Q(s) t ,a t ,w) is used to estimate the action value of the network output, y t The target action value output by the target network. Finally, based on... p g =|ξ g |+λ is used to sort the experiences in the experience buffer pool; where p g Priority of experience g; ξ g The training error of the empirical g (the Q-value y calculated by the target network) is... t The difference between the estimated network's calculated Q-value Q(s,a,w) and the calculated Q-value λ; λ is a factor to prevent p gA small integer equal to 0; P(g) is the sampling probability of the experience g; μ is the annealing factor used to adjust the sampling weight; thus, high-priority (high-error) experiences are preferentially extracted as a preset number of samples, ensuring that samples with larger TD errors are sampled first, thereby solving the problem of low efficiency of uniform sampling in the traditional DQN algorithm and reducing the training cycle and computational resource consumption of the agent. The estimation network is used to evaluate the action value corresponding to the extracted samples, obtaining the second action value set corresponding to the preset number of extracted samples. At the same time, the state space is input into the target network to obtain the target action value, and the corresponding Rainbow loss is calculated to quantify the deviation between the current agent policy and the optimal policy, providing core data support for the subsequent construction of the total loss function and updating the agent network parameters.

[0112] This embodiment provides a control system for transient voltage in a power system, including a state parameter acquisition module, a transient model construction module, a neural network construction module, an agent construction module, an agent loss calculation module, a total loss function construction module, and an agent training module, specifically:

[0113] The status parameter acquisition module is used to acquire the initial power system operating status parameters corresponding to the initial power system in real time.

[0114] The transient model construction module is used to establish a power system transient model based on the initial power system operating state parameters; and to obtain the initial transient voltage severity index based on the initial power system operating state parameters and the preset initial voltage.

[0115] The neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function;

[0116] The agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters and preset noise network.

[0117] The agent loss calculation module is used to perform a disconnection action on the initial power system based on the initial Rainbow agent and calculate the corresponding Rainbow loss.

[0118] The total loss function construction module is used to construct the total loss function based on Rainbow loss and neural network loss function;

[0119] The agent training module is used to train the initial Rainbow agent with the goal of minimizing the total loss function until a preset convergence condition is met, thereby obtaining the Rainbow agent to control the transient voltage of the power system.

[0120] In this embodiment, the initial power system operating state parameters corresponding to the initial power system are acquired in real time, wherein the initial power system operating state parameters include: the mechanical power P of generator i. Ti The electromagnetic power P of generator i Ei The rotor speed ω of generator i relative to the center of inertia i The rotor angle δ of motor i relative to the center of inertia i The active power P of the j-th bus dj The reactive power Q of the j-th bus dj The voltage V of the j-th bus j The actual bus voltage V at time t t Synchronous speed constant ω0, generator i's inertia coefficient T Ji The generator damping coefficient D is typically set to 2.0. This allows for the construction of a transient physical model of the power system after a short-circuit fault.

[0121]

[0122] Based on the initial power system operating state parameters and the preset initial voltage, obtain the initial transient voltage severity index: Where TVSI represents the transient voltage severity index, and T is the transient period; T c This is the fault clearing time; TVDI t It is the transient voltage deviation index, calculated from data obtained after time-domain simulation of the power system, and the formula is as follows:

[0123]

[0124] In the formula, V0 represents the preset initial voltage, V t TVSI represents the actual bus voltage at time t, and η represents the preset deviation threshold, typically set to 0.2. A higher TVSI indicates a more unstable power system. Additionally, V is set... t At the end of the control operation, the voltage returns to the range of 0.7 pu to 1.2 pu, indicating that the actual bus voltage at that moment is considered stable. Then, a neural network construction module is used to build a physical information neural network through a power system transient model and derive its corresponding neural network loss function. This ensures that the prediction results of the physical information neural network follow physical laws. Simultaneously, the neural network loss function can quantify the data deviation of the prediction results and can be added to the Rainbow loss to form the total loss function, accelerating the training of subsequent Rainbow agents and providing a basis for error correction. Next, an agent loss calculation module is used to perform a disconnection action on the initial power system using the initial Rainbow agent and calculate the corresponding Rainbow loss L. RainbowThis can quantify the deviation between the agent's current action decision and the optimal control objective, reflecting the agent's decision-making performance during transient voltage control, and providing decision error data for subsequent construction of the total loss function and agent training. Subsequently, the total loss function is constructed: L total =L Rainbow +ρ PINN MSE PINN In the formula, ρ PINN MSE represents the coefficients of the neural network loss function corresponding to the physical information neural network. PINN This represents the neural network loss function corresponding to the physical information neural network. A total loss function construction module is used to construct a total loss function that comprehensively measures the decision-making bias of the agent and the prediction bias of the physical information neural network, ensuring that the trained agent meets both the control requirements at the decision-making level and conforms to physical constraints. Finally, an agent training module is used to train the initial Rainbow agent with the goal of minimizing the total loss function. Gradient descent is used to update the target network weights in the Rainbow agent until a preset convergence condition is reached. This continuous optimization of the agent enables the final Rainbow agent to possess precise and efficient transient voltage control capabilities. This allows for rapid control of the power system to restore stability after a short-circuit fault (when a short-circuit fault occurs in the power system, the power bus is disconnected through actions until the transient voltage stabilizes within the range of 0.7 pu to 1.2 pu). Simultaneously, it avoids overestimating the value of actions, which could lead to inaccurate transient voltage control decisions and thus prevent the power system from becoming increasingly unstable.

[0125] Furthermore, the neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function; including:

[0126] Based on the transient model of the power system, an initial physical information neural network is constructed.

[0127] Based on the initial physical information neural network, its corresponding initial neural network loss function is derived.

[0128] The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

[0129] In this embodiment, since the dynamic equations of a dynamic system can be expressed in the general form of parameterized and nonlinear partial differential equations, therefore: Where u(t,x) represents the hidden solution, x is the input vector (the rotor angle obtained in the above steps), and N[u] represents a nonlinear operator applied to state u to describe the nonlinear dynamic characteristics of the power system. It can be used to encapsulate a series of physical mathematical formulas (i.e., Therefore, u(t,x)=δ(t); In order to describe the dynamic characteristics of a highly dynamic power system, a deep neural network f(t,x) is constructed based on the transient model of the power system to approximate u, which can be given by the left side of the differential equation: Right now The neural network is fed into the observation vector x and time t. Automatic differentiation is used to process some differential equations, calculating the partial derivatives of u with respect to t and x. u(t,x) and f(t,x) constitute the initial physical information neural network, having the same network parameters but different activation functions. u(t,x) outputs the predicted value of the initial physical information neural network, while f(t,x) outputs the deviation value of the initial physical information neural network under physical constraints. By constructing the initial physical information neural network, the physical laws of the power system transient process can be embedded into the neural network structure, providing an initial network framework for subsequent accurate prediction of the power system state at any time after a fault and for realizing transient voltage analysis under physical constraints. Subsequently, by deriving the corresponding initial neural network loss function, the deviation between the initial network prediction value and the actual data, as well as the deviation of the initial network prediction value from the physical laws of the power system, can be quantified. This provides direction for adjusting the parameters and optimizing the performance of the physical information neural network during subsequent training.

[0130] MSE PINN =MSE u +MSE f

[0131]

[0132] Among them, MSE P NN It is the initial neural network loss function, MSE. u It is the data error of the initial physical information neural network (the deviation between the predicted values ​​of the initial physical information neural network and the actual data), MSE f It is the physical information error of the initial physical information neural network (the deviation of the predicted value of the initial physical information neural network under physical constraints); N u It is the total number of valid samples used to calculate data error. It is the time point corresponding to the nth data sample. It is the input vector of the nth data sample, u n It is the real time-domain simulation data of the nth data sample, used as a benchmark for comparing the model's predicted values; N fIt is the total number of valid samples used to calculate the error in physical information. It is the time point corresponding to the nth data sample. This is the input vector of the nth data sample. During training, the calculated error value is fed back to the optimizer to update the neural network parameters until the preset convergence condition (MSE) is reached. PINN =0), obtain physical information neural networks that conform to physical laws and have high prediction accuracy, and derive their corresponding neural network loss function, providing reliable physical constraint model support for subsequent fusion with Rainbow agent, construction of total loss function and realization of transient voltage control.

[0133] Furthermore, the agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and a preset noise network; including:

[0134] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed.

[0135] The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

[0136] In this embodiment, an initial Rainbow agent is constructed using initial power system operating state parameters, an initial transient voltage severity index, and preset reward parameters. This allows for the creation of a decision framework that meets the requirements of power system transient voltage control scenarios, laying the foundation for subsequently integrating a preset noise network to form a complete initial Rainbow agent. Next, by adding a preset noise network to the initial Rainbow agent, the agent's exploration strategy is optimized, enabling the acquired initial Rainbow agent to possess more efficient exploration capabilities. This provides support for subsequent execution of transient voltage control cutoff actions and the achievement of accurate decision-making.

[0137] Furthermore, the construction of the original Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters includes:

[0138] Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established.

[0139] Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network;

[0140] The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

[0141] In this embodiment, voltage control after a short-circuit fault in the power system is modeled as a Markov decision process (i.e., state space, preset action space, and reward function). Specifically, an initial state space is established based on the initial power system operating state parameters: S = {P} Ti ,P Ei ,δ i ,ω i ,P dj Q dj V j ,t};The mechanical power P of generator i Ti The electromagnetic power P of generator i Ei The rotor speed ω of generator i relative to the center of inertia i The rotor angle δ of motor i relative to the center of inertia i The active power P of the j-th bus dj The reactive power Q of the j-th bus dj The voltage V of the j-th bus j By establishing an initial state space, the intelligent agent can accurately perceive the initial operating state of the power system. The preset action space is α = {ΔP}. dj (t)},ΔP dj (t) represents the amount of power removed from the load bus at time t. A reward function is established based on the initial power system operating state parameters, the initial transient voltage severity index, and preset reward parameters:

[0142]

[0143] In the formula, ep represents the number of load reductions. This represents the severity index of the transient voltage of the j-th bus at time t. Preset reward parameters include: r pun It is the penalty coefficient, r stepC1 and C2 are the frequency reward coefficients, and their respective weight coefficients, designed based on empirical values ​​during initial training. These can be adjusted during subsequent training to ensure good agent performance. By establishing a reward function, clear input and feedback logic can be defined for the agent's subsequent action selection and training. If the power system remains unstable (i.e., transient voltage exceeds 0.7 pu to 1.2 pu) after a load shedding operation, the reward is set to negative, related to the number of load shedding operations, to increase the penalty. When the power system becomes temporarily stable after a load shedding operation, the load shedding action is evaluated through three parts: 1. Effectiveness: defined by the effective load shedding degree, i.e., TVSI; 2. Total load shedding; 3. Load shedding count reward, with a larger reward for more load shedding operations. An estimation network and a target network are constructed based on the initial state space, preset action space, and preset network parameters. The estimation network splits the output layer of the original Rainbow agent into an action advantage sub-network and a state value sub-network. In the formula, Q(s,a,w) represents the action value (Q-value), A(s,a,w) represents the action advantage subnetwork, and V(s,w) represents the state value subnetwork. Let represent the average of all action advantage values, a′ represent each action in state s, and w represent the parameters of the action advantage subnetwork. The target network updates y using the TD Target formula. t =r t +γ·argmaxQ′(s t+1 ,a′,w′),r t γ is the immediate reward obtained after performing the action at time step t; γ is the discount factor, typically ranging from 0 to 1; s t+1 The state variable at time step t+1 represents the new environment state after the action is performed, and a′ represents the next state s. t+1The variables for all possible actions are defined, with w' being the parameters of the Q' function. This ensures that if an agent experiences an overestimation error in any network during subsequent training, the error will not propagate to another network, improving the stability and speed of the algorithm's learning. Furthermore, by acquiring a set of action values ​​(Q-values), the exploration scope of action values ​​can be expanded based on the original agent's action values. This retains the basic decision-making logic of the original Rainbow agent while solving the problem of limited exploration behavior in traditional DQN algorithms, providing a Q-value basis for subsequent selection of actions that meet the needs of the power system. By constructing an estimation network, the action value (Q-value) of each action in the current state can be evaluated in real time; by constructing a target network, a relatively stable target Q-value can be provided to avoid error propagation during estimation network updates; through the cooperation of the estimation network and the target network, the problem of Q-value overestimation in traditional DQN algorithms can be solved. Simultaneously, by combining the input dimension of the initial state space with the output dimension of the preset action space, the network structure is ensured to match the scenario requirements of transient voltage control in the power system.

[0144] This embodiment combines the Rainbow algorithm with Physical Information Neural Network (PINN) to propose a control method and system for transient voltage in power systems. It improves upon the shortcomings of the traditional DQN algorithm, achieving breakthroughs in Q-value overestimation optimization, sample utilization efficiency improvement, network structure optimization, and exploration strategy improvement. It can quickly control the power system to restore stability after a short-circuit fault, while avoiding overestimation of action value leading to inaccurate transient voltage control decisions, thereby preventing the power system instability from worsening.

[0145] 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 method for controlling transient voltage in a power system, characterized in that, include: Real-time acquisition of the initial power system operating status parameters corresponding to the initial power system; A transient model of the power system is established based on the initial operating state parameters of the power system. Based on the initial power system operating state parameters and the preset initial voltage, the severity index of the initial transient voltage is obtained; Based on the transient model of the power system, a physical information neural network is constructed and its corresponding neural network loss function is derived. Based on the initial power system operating state parameters, the initial transient voltage severity index, the preset reward parameters, and the preset noise network, the initial Rainbow agent is obtained. Based on the initial Rainbow agent, the initial power system is disconnected, and the corresponding Rainbow loss is calculated; Construct a total loss function based on Rainbow loss and neural network loss function; The initial Rainbow agent is trained with the goal of minimizing the total loss function until the preset convergence condition is met, thus obtaining the Rainbow agent to control the transient voltage of the power system.

2. The method for controlling transient voltage in a power system according to claim 1, characterized in that, The process of constructing a physical information neural network based on a power system transient model and deriving its corresponding neural network loss function includes: Based on the transient model of the power system, an initial physical information neural network is constructed. Based on the initial physical information neural network, its corresponding initial neural network loss function is derived. The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

3. The method for controlling transient voltage in a power system according to claim 1, characterized in that, The initial Rainbow agent is obtained based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and a preset noise network; including: Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed. The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

4. The method for controlling transient voltage in a power system according to claim 3, characterized in that, The initial Rainbow agent is constructed based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters; including: Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established. Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network; The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

5. The method for controlling transient voltage in a power system according to claim 4, characterized in that, The process of obtaining the initial Rainbow agent based on the original Rainbow agent and a preset noise network includes: Based on the original Rainbow agent and the preset noise network, obtain the first action value set; Select the action value in the first action value set that satisfies the preset power system constraints and the preset conditions as the first action value, and execute the cut-off action corresponding to the first action value to obtain the state space; Based on the original Rainbow agent and state space, obtain the initial Rainbow agent.

6. The method for controlling transient voltage in a power system according to claim 5, characterized in that, The step of performing a disconnection action on the initial power system based on the initial Rainbow agent and calculating the corresponding Rainbow loss includes: Calculate the reward value based on the initial Rainbow agent and reward function; The initial state space, the cut-off action corresponding to the value of the first action, the reward value, and the state space are encapsulated and stored in a preset experience buffer pool. The experience buffer pool is then obtained. A preset number of samples are drawn from the experience buffer pool and input into the estimation network to obtain the second action value set corresponding to the preset number of samples. Based on the state space and target network, obtain the value of the target action; Based on the second action value set and the target action value, calculate the corresponding Rainbow loss.

7. A control system for transient voltage in a power system, characterized in that, It includes a state parameter acquisition module, a transient model construction module, a neural network construction module, an agent construction module, an agent loss calculation module, a total loss function construction module, and an agent training module, specifically: The status parameter acquisition module is used to acquire the initial power system operating status parameters corresponding to the initial power system in real time. The transient model construction module is used to establish a power system transient model based on the initial power system operating state parameters; and to obtain the initial transient voltage severity index based on the initial power system operating state parameters and the preset initial voltage. The neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function; The agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters and preset noise network. The agent loss calculation module is used to perform a disconnection action on the initial power system based on the initial Rainbow agent and calculate the corresponding Rainbow loss. The total loss function construction module is used to construct the total loss function based on Rainbow loss and neural network loss function; The agent training module is used to train the initial Rainbow agent with the goal of minimizing the total loss function until a preset convergence condition is met, thereby obtaining the Rainbow agent to control the transient voltage of the power system.

8. A control system for transient voltage in a power system according to claim 7, characterized in that, The neural network construction module is used to construct a physical information neural network based on the power system transient model and derive its corresponding neural network loss function; including: Based on the transient model of the power system, an initial physical information neural network is constructed. Based on the initial physical information neural network, its corresponding initial neural network loss function is derived. The initial physical information neural network is trained with the goal of minimizing the initial neural network loss function until the preset convergence condition is met. The physical information neural network is then obtained and its corresponding neural network loss function is derived.

9. A control system for transient voltage in a power system according to claim 7, characterized in that, The agent construction module is used to obtain an initial Rainbow agent based on initial power system operating state parameters, initial transient voltage severity index, preset reward parameters, and a preset noise network; including: Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an original Rainbow agent is constructed. The initial Rainbow agent is obtained based on the original Rainbow agent and the preset noise network.

10. A control system for transient voltage in a power system according to claim 9, characterized in that, The initial Rainbow agent is constructed based on initial power system operating state parameters, initial transient voltage severity index, and preset reward parameters; including: Based on the initial power system operating state parameters, the initial transient voltage severity index, and the preset reward parameters, an initial state space and reward function are established. Based on the initial state space, the preset action space, and the preset network parameters, construct the estimation network and the target network; The original Rainbow agent is built based on the initial state space, preset action space, reward function, estimation network, and target network.

Images