A direct current receiving end power grid transient voltage stability emergency control method and system and medium

Abstract

The application provides a direct-current receiving end power grid transient voltage stability emergency control method and system, a medium, first constructs a GCN-DDPG fusion model, on the basis, introduces a double evaluation network mechanism and an evaluation network and action network asynchronous update strategy to improve the algorithm effect. Based on the GCN-DDPG fusion model, a transient voltage stability emergency control model is constructed, after the model training is completed in the dispatching center, it is issued to the safety control master station, the transient voltage stability emergency control model based on the GCN-DDPG fusion model after the training is completed will be according to the actual operation mode and fault information of the power grid, the emergency control strategy, namely the direct-current current control scheme is quantitatively calculated online, and is sent to the direct-current control and protection system for execution. The application realizes the rapid given of the emergency control strategy, improves the efficiency and reduces the strategy conservatism, and has important significance for the online generation and application of the emergency control strategy.

Description

Technical Field

[0001] This invention relates to the field of emergency control of transient voltage stability in a novel high-voltage DC receiving-end power system, specifically a method, system, and medium for emergency control of transient voltage stability in a DC receiving-end power grid. Background Technology

[0002] In new power systems, conventional high-voltage direct current (LCC-HVDC) transmission systems enable the long-distance transmission of clean energy across regions to load centers. In the HVDC receiving-end grid, the integration of DC leads to a reduction in synchronous generator operation, resulting in relatively strained dynamic reactive power and a significantly increased risk of transient voltage instability. HVDC endpoints are typically located near the load center of the receiving-end grid. When induction motors are the primary load at the load center, faults in the receiving-end grid, especially near the converter station, can cause sustained voltage drops. To avoid transient voltage instability, emergency control is required after a fault occurs. Common emergency control measures include generator tripping and load shedding. However, generator tripping hinders grid recovery, generator reconnection can cause impact, and load shedding presents economic and reliability issues. HVDC transmission, with its millisecond-level response speed, can rapidly regulate active power in interconnected power systems. Studies have shown that adjusting DC transmission power by reducing DC current can release the reactive power of converter station capacitors, thereby enabling the DC inverter station to function as a large-capacity dynamic reactive power source for the receiving-end power grid. This can improve the voltage level of the receiving-end power grid under unstable conditions dominated by transient voltage instability. Based on this, this invention considers adjusting DC active power to address the transient voltage stability problem of the DC receiving-end power grid.

[0003] Conventional emergency control methods generate one or more conservative emergency control strategies offline to address the numerous different operating modes and fault scenarios in new power systems, resulting in significant control costs. To reduce these costs, it is necessary to acquire emergency control schemes online for specific operating modes and fault scenarios. Existing data-driven methods train deep reinforcement learning (DRL) models using a large number of operating modes and fault scenarios. The trained DRL models can then quickly and automatically generate corresponding emergency control strategies based on the actual operating mode and fault information of the power grid. However, conventional DRL models struggle to account for changes in the power grid topology. If the power grid topology or fault location changes, the effectiveness of the trained DRL-based emergency control model needs further investigation.

[0004] Graph Neural Networks (GCNs) are deep learning methods that reflect the connections between nodes. Compared to conventional fully connected neural networks that only take grid state information (such as node voltages) as input, GCNs can also take into account grid topology information. Therefore, to adapt to different grid operation modes and fault scenarios, especially changes in grid topology, this invention proposes an online generation method for emergency control strategies for transient voltage stability in DC receiving-end grids based on a Deep Deterministic Policy Gradient (DDPG) model (a DRL model) improved from GCN. Specifically, in terms of DRL model construction, the network structure of DDPG is modified based on GCN to construct a GCN-DDPG fusion model. On this basis, a dual evaluation network mechanism and a asynchronous update strategy between the evaluation network and the action network are introduced to improve the algorithm performance. Furthermore, in terms of application, a transient voltage stability emergency control model is constructed based on the GCN-DDPG fusion model. After the model training is completed at the dispatch center, it is sent to the safety control master station. The trained transient voltage stability emergency control model based on the GCN-DDPG fusion model will perform online quantitative calculation of the emergency control strategy, i.e., the DC current control scheme, according to the actual operation mode of the power grid and fault information, and send it to the DC control and protection system for execution. Summary of the Invention

[0005] The purpose of this invention is to address the transient voltage stability problem of DC receiving-end power grids. To efficiently and quickly obtain emergency control strategies for transient voltage stability to meet online application requirements and adapt to different power grid operation modes and fault scenarios, especially changes in power grid topology, this invention proposes an online generation method for emergency control strategies for transient voltage stability of DC receiving-end power grids based on an improved deep reinforcement learning model, based on the idea that DC inverter stations can be used as emergency reactive power sources for receiving-end power grids.

[0006] An emergency control method for transient voltage stability of a DC receiving-end power grid includes the following steps:

[0007] Construct an emergency control model for transient voltage stability in a new type of DC receiving-end power system based on the GCN-DDPG fusion model;

[0008] Real-time acquisition of power grid data and monitoring of power grid status; if a fault occurs, an emergency control strategy is generated based on the emergency control model, quantified and calculated online, and sent to the DC control and protection system for execution; otherwise, monitoring of the power grid status continues until a fault occurs.

[0009] As a preferred option, the GCN-DDPG fusion model includes

[0010] Evaluation Network: Generates evaluations of actions based on the definition of the reward function;

[0011] Action network: Used to generate emergency control plans based on power grid status information.

[0012] Preferably, the evaluation network includes evaluation network one and evaluation network two, wherein evaluation network one and evaluation network two have the same structure and both include:

[0013] Graph convolutional layers: used to input power grid state information and topology information, and to extract features from the input information;

[0014] Fully connected network layer: used for Q-value fitting and generating action evaluation.

[0015] As a preferred option, action networks include

[0016] Graph convolutional layers: used to input power grid state information and topology information, and to extract features from the input information;

[0017] Fully connected network layer: used to generate actions, i.e. emergency control plans.

[0018] As a preferred option, constructing an emergency control model for transient voltage stability of a new type of DC receiving-end power system based on the GCN-DDPG fusion model specifically includes:

[0019] Collect power grid data, including power grid status information (voltage of each node) and topology (composed of adjacency matrix). A express);

[0020] The collected data is divided into training set and test set. The training set data is input into the GCN-DDPG fusion model for training until convergence.

[0021] The test set data is input into the trained GCN-DDPG fusion model, which can then quickly generate an emergency control plan based on the power grid status information, topology, and fault information.

[0022] As a preferred approach, a dual-evaluation network mechanism is employed when training the GCN-DDPG fusion model, specifically:

[0023] Based on the original DDPG evaluation network, an additional evaluation network with the same structure is added. Both evaluation networks are used to estimate the target Q-value, and the smaller Q-value is selected in each update to reduce the impact of noise and overestimation bias. It should be noted that although the two evaluation networks have the same structure, their parameters and updates are independent of each other. These two evaluation networks are denoted as follows: Q 1 and Q 2, whose parameters are respectively and The GCN-DDPG fusion model algorithm is based on the above improvements, and its loss function is shown below:

[0024] (1)

[0025] In the formula, L The loss function; N This represents the training sample size for the network. For parameters The evaluation network below, therefore Indicates the state Apply action at the time The subsequent expected return, i.e., the parameter is The output of the evaluation network; For parameters The action network below represents the parameters. Under, state s The output of the policy network is the input; Perform an action while indicating a state Instant rewards; γ This is the discount factor.

[0026] As a preferred approach, when training the GCN-DDPG fusion model, the evaluation network and the action network are updated asynchronously, specifically:

[0027] During model training, the action network experiences errors when updating parameters, which affects the evaluation network's estimation of the Q-value. The Q-value, in turn, serves as a reference for the action network's updates. Instability in the Q-value can cause significant fluctuations during training. The introduction of a dual evaluation network further increases the number of network parameters in the GCN-DDPG model, leading to increased training complexity. Therefore, fluctuations caused by Q-value instability are more likely to occur during training, increasing the uncertainty of training results. Training the action network only after the evaluation network has stabilized can reduce the occurrence of unreasonable strategies. Therefore, this paper adopts a asynchronous update strategy for the evaluation and action networks. The evaluation network is updated several times and stabilized before the action network is updated once. This avoids fluctuations caused by Q-value instability and reduces the parameter update frequency, improving training efficiency and thus enhancing algorithm performance.

[0028] As a preferred option, the power grid is considered as the environment, and the direct current is... I d As an action, it completes the definition of the state inputs, action outputs, and reward functions of the GCN-DDPG fusion model, specifically including...

[0029] The state variables input for the GCN-DDPG fusion model are as follows:

[0030] (2)

[0031] In the formula, X That is U t , indicating in t The voltage vector at each node of the power grid at any given time. for t The adjacency matrix that characterizes the power grid topology at any given time. for The degree matrix;

[0032] The state outputs of the GCN-DDPG fusion model, i.e., the actions, are as follows:

[0033] (3)

[0034] In the formula, Indicates in t The DC current control value is set at any given time, and the DC current range is set to [0.55, 1.0].

[0035] The following transient voltage stability criterion is adopted: during the transient process after a disturbance to the power system, the load bus voltage can recover to above 0.80 pu within 10 seconds after the fault; considering stability, stability margin, and stability control costs, the reward function... The expression is as follows:

[0036] (4)

[0037] In the formula, The moment of failure; Indicates at time t The lowest voltage value in the entire network, which usually corresponds to the voltage amplitude of the bus near the fault location; Indicates time t The action amount; within 10 seconds after the fault, if the lowest voltage of all nodes in the network is greater than 0.8pu after the action is applied, it indicates that the transient voltage has recovered to above the stable standard, and a positive reward is given, and the larger the voltage value, the greater the reward; if the lowest voltage of the node is still less than 0.8pu, a negative reward is given, and the smaller the voltage value, the smaller the reward; if 10 seconds after the fault, the lowest node voltage of the grid is still less than 0.8pu after the action is applied, a reward of -500 is given, and the training round ends.

[0038] Compared with existing technologies, the beneficial effects of this invention are as follows: The improved GCN-DDPG fusion model of this invention supplements the input of power grid topology information, realizing the perception of power grid topology and its changes; on this basis, by introducing a dual evaluation network mechanism and an asynchronous update strategy between the action network and the evaluation network, the algorithm performance is improved; this method realizes the rapid formulation of emergency control strategies by monitoring the power grid state and topology, improving efficiency and reducing strategy conservatism, which is of great significance for the online generation and application of emergency control strategies. Attached Figure Description

[0039] Figure 1 This is a diagram of the GCN network structure;

[0040] Figure 2 This is the network structure diagram of the GCN-DDPG fusion model;

[0041] Figure 3 This is a flowchart of the emergency control model application;

[0042] Figure 4 This is the system structure diagram of the simulation example;

[0043] Figure 5 This is a graph showing the training process rewards of the GCN-DDPG fusion model.

[0044] Figure 6 This is a reward graph of the standard DDPG model training process;

[0045] Figure 7 This is a graph showing the feedback during the testing process of the GCN-DDPG fusion model.

[0046] Figure 8 This is a return graph of the standard DDPG model test process;

[0047] Figure 9 This is a diagram of a DC current control scheme for a standard test sample;

[0048] Figure 10 This is a voltage curve of node 9 in a standard test sample.

[0049] Figure 11 This is a graph showing the frequency variation of a high-voltage DC receiving-end power grid for a standard test sample.

[0050] Figure 12 This is a graph showing the return of the generalization test process of the GCN-DDPG fusion model;

[0051] Figure 13 This is the return graph of the generalization test process of the standard DDPG model;

[0052] Figure 14 This is a diagram of a DC current control scheme for a generalized test sample;

[0053] Figure 15 This is a voltage curve of node 14 in a generalized test sample. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings:

[0055] Example 1

[0056] Figure 1 This is a diagram of the GCN network structure. Figure 1 It can be seen that, in addition to taking a vector composed of node feature information as input, GCN also takes an adjacency matrix as input. A The topological structure is represented as input, thereby enabling perception of the topological structure.

[0057] Based on this, the present invention proposes a network structure based on an improved standard DDPG model using GCN, and constructs a GCN-DDPG fusion model, such as... Figure 2 As shown. The input and output of the GCN-DDPG fusion model are as follows: the input of the action network is the power grid state variables, corresponding to the input of the feature matrix. X The node voltage information represented by the adjacency matrix A The topology of the power grid is represented, and the output is the DC current control quantity. The initial input of the evaluation network is the same as that of the action network. The first fully connected layer after the graph convolutional layer is composed of the output of the convolutional layer and the action generated by the action network. The output is the Q value of the action applied in this environment state, which is used for action evaluation. The complexity of the GCN-DDPG fusion model is increased, which leads to more serious noise influence and easy overfitting. In addition, the Q value estimation used in the action-comment framework of the standard DDPG model algorithm often has an overestimation bias. To this end, a dual evaluation network mechanism is introduced. On the basis of the original evaluation network, an evaluation network with the same structure is added. Both evaluation networks are used to estimate the target Q value. The smaller Q value is selected each time it is updated, thereby reducing the influence of noise and overestimation bias. Based on this, the loss function is shown in Equation (7). In the training process of the GCN-DDPG fusion model, the large fluctuation caused by the instability of Q value, and the further increase in the network parameters of the GCN-DDPG model due to the introduction of the dual evaluation network, further increase the training complexity. This invention proposes a asynchronous update strategy for the evaluation network and the action network. The evaluation network is updated several times and stabilized before the action network is updated once. This avoids fluctuations caused by the instability of the Q value and reduces the parameter update frequency to improve training efficiency, thereby enhancing the algorithm performance.

[0058] Figure 3This is a flowchart of the emergency control model application. The integration of the GCN-DDPG fusion model with DC current control is reflected in the acquisition and implementation of emergency control measures. Conventional emergency control involves generating a large number of emergency control strategies offline for different new power system operating modes, and then combining them into one or more relatively conservative emergency control strategies to cope with different operating modes and fault scenarios, thus incurring significant control costs. To quickly generate emergency control strategies and reduce their conservatism, an emergency control model based on the GCN-DDPG fusion model is trained online at the dispatch center based on anticipated operating modes and faults. The trained model is periodically distributed to the safety control master station, thereby periodically refreshing the emergency control model in the safety control master station. During the training process, the power grid model provides feature matrices to the GCN-DDPG fusion model. X and adjacency matrix A The GCN-DDPG fusion model provides action data to the power grid model. In application, the emergency control model based on the GCN-DDPG fusion model in the security control master station will generate corresponding emergency control strategies online, i.e., DC current control schemes, based on information such as the actual operation mode of the power grid and the location of the fault. This scheme will then be sent to the DC control and protection system for execution. In DC constant current control mode, the DC current control scheme will be used as a current reference value. I ref , I ref The current obtained from VDCOL I VDCOL In comparison, the smaller value is taken as the final current command. Therefore, the method proposed in this invention transforms emergency control from the traditional offline formulation to online application formulation and execution, realizing the online generation of emergency control strategies, improving efficiency and reducing strategy conservatism.

[0059] Specifically, it includes:

[0060] Step 1: To address the shortcoming of the standard Deep Deterministic Policy Gradient (DDPG) model (a type of deep reinforcement learning model) in its difficulty in accepting topological information as input, this invention introduces a Graph Convolutional Network (GCN) to modify the network structure of the standard DDPG. The fully connected network layer at the beginning of the standard DDPG model network structure is replaced with a graph convolutional layer, thus constructing a GCN-DDPG fusion model that can adapt to the input of topological information.

[0061] Step 2: To address the increased complexity of the GCN-DDPG fusion model, its greater susceptibility to noise, and its susceptibility to overfitting, as well as the shortcomings of the Q-value estimation used in the action-comment framework of the standard DDPG model algorithm, which often suffers from overestimation bias, this invention introduces a dual evaluation network mechanism. On the basis of the original evaluation network, an evaluation network with the same structure is added. Both evaluation networks are used to estimate the target Q-value, and the smaller Q-value is selected in each update, thereby reducing the impact of noise and overestimation bias.

[0062] Step 3: To address the significant fluctuations caused by the instability of Q-values ​​during the training of the GCN-DDPG fusion model, and the further increase in network parameters and training complexity due to the introduction of dual evaluation networks, this invention proposes a asynchronous update strategy for the evaluation network and action network. The evaluation network is updated several times until it stabilizes before updating the action network once. This avoids fluctuations caused by the instability of Q-values ​​and reduces the parameter update frequency to improve training efficiency, thereby enhancing the algorithm's performance.

[0063] Step 4: Treat the power grid as the environment and direct current... I d As an action, the state input, action output and reward function of the GCN-DDPG fusion model are defined, and an emergency control model for transient voltage stability of a new type of DC receiving-end power system is constructed based on the GCN-DDPG fusion model.

[0064] Step 5: Train the emergency control model based on the GCN-DDPG fusion model and send the trained model to the safety control master station. The trained transient voltage stability emergency control model based on the GCN-DDPG fusion model will perform online quantification calculation of the emergency control strategy, i.e., the DC current control scheme, based on the actual operation mode of the power grid and fault information, and send it to the DC control and protection system for execution.

[0065] Specifically, the Graph Convolutional Network (GCN) in step 1 works as follows:

[0066] GCN, as a type of neural network that performs convolution-like operations directly on graphs, can directly process non-Euclidean data. Inputs include... 3D feature matrix X and 3D adjacency matrix A . n The number of nodes; d This represents the number of features at each node. In a graph convolutional layer, node features flow through the graph structure as hidden features. The specific operational expression for a graph convolutional layer is:

[0067] (1)

[0068] In the formula, L This represents the number of convolutional layers. H For input feature data, when l When equal to 0, H for X . W l This is the weight matrix. This is the activation function.

[0069] However, under this transitive rule, since nodes do not have self-connections, the adjacency matrix... A The values ​​at the diagonal positions are all 0, and the node's own information is ignored during feature extraction. Furthermore, due to the adjacency matrix... A Without normalization, after multiple matrix transformations, the magnitude difference between the adjacency matrix and the input features will increase, which is detrimental to the training of neural networks. Therefore, the adjacency matrix... A With the identity matrix I Adding them together adds a self-connect to each node and improves the adjacency matrix. A Normalization, as shown below:

[0070] (2)

[0071] In the formula, For matrix A The degree matrix. Thus, the information transfer rules of GCN are expressed as:

[0072] (3)

[0073] For fully connected layers, the adjacency matrix does not participate in the computation process within the fully connected layer; its computation process is as follows:

[0074] (4)

[0075] Based on this, the structure of the DDPG model (GCN-DDPG fusion model) based on GCN improvement in step 1 is as follows:

[0076] The input to the action network of the GCN-DDPG fusion model is the power grid state quantity, corresponding to the characteristic matrix. X The node voltage information represented by the adjacency matrix A The topology of the power grid is represented, and the output is a DC current control quantity. The initial input of the evaluation network is the same as that of the action network. The first fully connected layer after the graph convolutional layer is composed of the output of the convolutional layer and the action generated by the action network. The output is the Q value of the action applied under this environmental state, which is used for action evaluation.

[0077] Due to the characteristic matrix XThe node characteristic number is used to describe the node voltage at a certain moment. d The difference between graph convolutional layers and fully connected layers in the GCN-DDPG network is whether or not an adjacency matrix is ​​introduced, i.e., the inter-layer information transfer rules are different. For graph convolutional layers, the specific inter-layer information transfer rules are as follows:

[0078] (5)

[0079] In the formula, and For hierarchical input and output, W GCN This represents the weight matrix of the graph convolutional layer. b GCN This represents the bias vector of the graph convolutional layer, and ReLU is the activation function.

[0080] For fully connected layers, the specific rules for information transfer between layers are as follows:

[0081] (6)

[0082] In the formula, W FC and b FC These represent the weight matrix and bias vector of the fully connected layer, respectively.

[0083] Specifically, the training algorithm for the GCN-DDPG fusion model under the dual evaluation network mechanism in step 2 is as follows:

[0084] The increased complexity of the GCN-DDPG fusion model makes it more susceptible to noise and prone to overfitting. Furthermore, the Q-value estimation used in the action-comment framework of the standard DDPG model often suffers from overestimation bias. To address this, a dual-evaluation network mechanism is introduced. An additional evaluation network with the same structure is added to the original evaluation network. Both evaluation networks are used to estimate the target Q-value, and the smaller Q-value is selected in each update to reduce the impact of noise and overestimation bias. It should be noted that these two evaluation networks have identical structures, but their parameters and updates are independent of each other. These two evaluation networks are denoted as... Q 1 and Q 2, whose parameters are respectively and .

[0085] In the standard DDPG model algorithm, the evaluation network is optimized by minimizing the loss function. The GCN-DDPG fusion model algorithm is based on the above improvements, and its loss function is as follows:

[0086] (7)

[0087] In the formula, L The loss function; N This represents the training sample size for the network. For parameters The evaluation network below, therefore Indicates the state Apply action at the time The subsequent expected return, i.e., the parameter is The output of the evaluation network; For parameters The action network below represents the parameters. Under, state s The output of the policy network is the input; Perform an action while indicating a state Instant rewards; γ This is the discount factor.

[0088] The action network and evaluation network each contain a target network and an online network, respectively. For clarity, the target network and its parameters are marked with a superscript "'". During training, to improve training stability, the target network is updated via the online network at regular intervals, denoted as _. K The update strategy is as follows:

[0089] (8)

[0090] In the formula, This refers to the network update rate.

[0091] Specifically, the asynchronous update strategy between the evaluation network and the action network in step 3 is as follows:

[0092] During model training, the action network updates its parameters with some error, which affects the evaluation network's estimation of the Q value. The Q value, in turn, serves as a reference for the action network's update. Instability of the Q value can cause significant fluctuations in the training process. Due to the introduction of the dual evaluation network mechanism, the network parameters of the GCN-DDPG model are further increased, leading to a further increase in training complexity. Therefore, fluctuations caused by Q value instability are more likely to occur during training, increasing the uncertainty of the training results. Training the action network after the evaluation network has stabilized can reduce the occurrence of unreasonable strategies. Therefore, this invention adopts a asynchronous update strategy between the evaluation network and the action network. The evaluation network is updated several times and stabilized before the action network is updated once. This avoids fluctuations caused by Q value instability and reduces the parameter update frequency to improve training efficiency, thereby improving the algorithm's performance. The action network update strategy is shown in equation (9). For details on algorithm improvements, please refer to Algorithm 1: GCN-DDPG Fusion Model Algorithm.

[0093] (9)

[0094]

[0095] Specifically, the emergency control model based on the GCN-DDPG fusion model in step 4 is as follows:

[0096] Treating the power grid as the environment (the application object of the DRL model), and direct current... I d As an action, an emergency control model for transient voltage stability of the HVDC receiving-end power grid is constructed to quickly obtain DC current control schemes. The power grid model and the emergency control model based on the GCN-DDPG model interact periodically (e.g., every 0.2 seconds): power grid data such as node voltages during the transient process after a fault are sent to the GCN-DDPG fusion model. The GCN-DDPG fusion model then generates actions based on this data and applies them to the power grid, evaluating the actions according to changes in the power grid state. The actions are shown below:

[0097] (10)

[0098] In the formula, Indicates in t The DC current control value is set at any time, and the DC current range is set to [0.55, 1.0].

[0099] To reflect power grid topology information, the input to the GCN-DDPG fusion model is a feature vector. X and adjacency matrix A State variables It is expressed as follows:

[0100] (11)

[0101] In the formula, X That is U t , for t The adjacency matrix that characterizes the power grid topology at any given time. for The degree matrix.

[0102] Based on the technical specifications for power system safety and stability calculation, this invention adopts the following transient voltage stability criterion (other stability criteria can be used similarly, only the return function in equation (12) needs to be adjusted accordingly): During the transient process after a power system is disturbed, the load bus voltage can recover to above 0.80 pu within 10 seconds after the fault. Taking into account stability, stability margin, and stability control costs, the return function... The expression is as follows:

[0103] (12)

[0104] In the formula, The moment of failure; Indicates at time t The lowest voltage value in the entire network, which usually corresponds to the voltage amplitude of the bus near the fault location; Indicates time t The action amount is determined as follows: Within 10 seconds after a fault, if the lowest node voltage across the entire network is greater than 0.8 pu after the action is applied, it indicates that the transient voltage has recovered to above the stable standard, and a positive reward is given, with a larger reward for a higher voltage value. If the lowest node voltage is still less than 0.8 pu, a negative reward is given, with a smaller reward for a lower voltage value. If, 10 seconds after the fault, the lowest node voltage is still less than 0.8 pu after the action is applied, a reward of -500 is given, and the training round ends. Furthermore, to ensure the stability of the receiving-end power angle and frequency, if the power angle difference after the fault is greater than 180° or the steady-state frequency exceeds the fluctuation range of ±0.2Hz (taking a 220kV voltage level grid as an example), a reward of -500 is given, and the training round ends.

[0105] Specifically, the training and deployment of the emergency control model based on the GCN-DDPG fusion model in step 5 are as follows:

[0106] At the dispatch center, an emergency control model based on a large number of anticipated operating modes and faults is trained online using the GCN-DDPG fusion model. The trained model is periodically distributed to the safety control master station, thus periodically refreshing the emergency control model in the master station. In the safety control master station, the trained transient voltage stability emergency control model based on the GCN-DDPG fusion model will perform online quantification calculations of the emergency control strategy, i.e., the DC current control scheme, based on information such as the actual operating mode and faults of the power grid, and send it to the DC control and protection system for execution. In the DC constant current control mode, the DC current control scheme is used as the current reference value. I ref , I ref The current obtained from VDCOL I VDCOL In comparison, the smaller value is taken as the final current command.

[0107] Example 2

[0108] This embodiment provides an emergency control system for transient voltage stability of a DC receiving-end power grid, including the following:

[0109] The first module is configured to build an emergency control model for transient voltage stability of a new type of DC receiving-end power system based on the GCN-DDPG fusion model.

[0110] The second module is configured to collect grid data in real time and monitor the grid's operational status. If a fault occurs, it generates an emergency control strategy based on the emergency control model, performs online quantitative calculations, and sends it to the DC control and protection system for execution. Otherwise, it continues to monitor the grid's operational status until a fault occurs.

[0111] Example 3

[0112] This embodiment provides a readable storage medium storing computer-executable instructions, which, when executed by a processor, configure the processor to perform the aforementioned emergency control method for transient voltage stabilization of the DC receiving-end power grid.

[0113] Example 4

[0114] This embodiment provides an application example.

[0115] based on Figure 4 The example system shown verifies the effectiveness and superiority of the proposed method. The generator output and load parameters of the example system are shown in Tables 1 and 2. The DC receiving-end system includes 5 generators, 11 loads, and a voltage reference of 220kV. Node 14 connects to a 90MW aggregate induction motor load; node 7 connects to a 180MW wind farm, which includes 120 wind turbines, each with a rated capacity of 1.5MW. The DC sending-end system is a simplified equivalent two-machine system, including 2 generators, 2 loads, and a voltage reference of 500kV. Node S1 connects to a remote equivalent system. Node S3 of the sending-end system and node 9 of the receiving-end system are respectively connected to the rectifier and inverter stations of the high-voltage direct current transmission system. The DC system operates at rated power unipolar, with a rated power of 1000MW, a rated voltage of 500kV, and a rated current of 2kA.

[0116] Table 1 Generator Output of the Example System

[0117]

[0118] Table 2 System Load of the Example

[0119]

[0120] Different load levels and parameters, generator and renewable energy plant start-up and output status, and fault location can affect the transient voltage stability of the power grid. Therefore, in order to obtain a large number of different operating modes to train the emergency control model based on the GCN-DDPG fusion model, the loads in Table 2 were adjusted to set different overall power grid load levels: including 80%, 90%, 100%, 110%, and 120% of the baseline load level. Under each load level, the output and load type and their proportions of each load were randomly changed, and the generator start-up and output status in Table 1 were adjusted accordingly to maintain power flow convergence. At the same time, the output of wind farms was adjusted to change the penetration level of renewable energy sources. The wind turbines were controlled to start and stop in units of 4 units. Therefore, it can be considered that the overall load characteristics of the power grid also changed through the above adjustments. The fault was considered to be located near node 14 of line 9-14 or near node 9 of line 9-10, both of which were three-phase short-circuit ground faults. In addition, one line of the original topology (lines 4-9, 2-4, 3-4, 1-5, 2-5, 1-2, 6-13, 12-13, 6-12, etc.) was disconnected to consider different power grid topologies. Based on the above analysis, the sample acquisition scheme is shown in Table 3.

[0121] Based on the sample acquisition scheme in Table 3, if the sample types are not repeated, there will be a total of 400 possible combinations. During the training phase, 4000 training samples are generated through time-domain simulation, following the principle of random occurrence of each combination in equal proportion. It should be noted that during the sample generation process, due to adjustments in generator output and load, power flow non-convergence may occur. If power flow non-convergence occurs, the sample is discarded and generation continues until 4000 training samples are obtained.

[0122] During the testing phase, 800 standard test samples were generated in the same manner as in Table 3. Considering the randomness of load distribution under different load levels and the randomness of output distribution under different start-up conditions, it can be assumed that no two test samples are exactly the same. Furthermore, considering two load levels (70% and 130%, both assuming generator 2 is not running), and two topologies with line 2-3 and line 4-5 disconnected, with the fault located near node 14 of line 9-14 or near node 9 of line 9-10, 400 generalized test samples were generated. Therefore, the total number of training samples is 4000, and the total number of test samples is 1200.

[0123] Table 3 Sample Acquisition Scheme

[0124]

[0125] To verify the superiority of the emergency control model based on the GCN-DDPG fusion model, an emergency control model based on the standard DDPG model was trained for comparison. In the above embodiments, the algorithm was run on the PyCharm platform using Python, and the deep learning framework was Tensorflow 2.0; the computer hardware configuration was 32GB RAM, Intel Xeon Platinum 8124M CPU, and GeForce RTX 3090Ti GPU (it should be noted that the hardware and software are not limited to the above configuration).

[0126] The training process rewards of the emergency control model based on the GCN-DDPG fusion model and the standard DDPG model are as follows: Figure 5 and Figure 6 As shown. It can be seen that the total reward of each training round shows a trend of gradually increasing and then stabilizing. Based on the definition of the reward function in equation (12), it can be seen that in the early stage of training, there is still a situation of transient voltage instability after applying the action. In the later stage of training, the reward of each round is maintained at around -100. It can be seen that the control of DC current in the later stage can achieve transient voltage stability of DC receiving-end grid. From Figure 6 It can be seen that the return in each round of the later training stage remained around -150, indicating that the control of DC current in the later stage could achieve transient voltage stability of the DC receiving-end grid. For comparison... Figure 5 and Figure 6 It can be seen that the emergency control model based on the GCN-DDPG fusion model has a faster convergence speed and less fluctuation in the training process. On the one hand, this is due to the asynchronous update strategy of the action and evaluation networks. The action network is trained and its parameters are updated after the evaluation network stabilizes, which avoids unreasonable strategies to a certain extent and improves training efficiency. On the other hand, the introduction of dual evaluation networks reduces the influence of overestimation bias and noise, thus making the training process more stable.

[0127] The test results of the emergency control model based on the GCN-DDPG fusion model and the standard DDPG model are as follows: Figure 7 and Figure 8 As shown, the returns of the conventional test process are around -100 and -150 respectively. According to the return function of equation (12), the overall voltage recovery effect is better after applying the DC current control scheme generated by the GCN-DDPG fusion model.

[0128] To further verify the effectiveness and superiority of the GCN-DDPG fusion model, a sample was randomly selected from the routine test samples. This sample had a load level of 100%, all generator units were running, and lines 2-4 were disconnected. The fault occurred near node 9 of line 9-10. After the fault, the voltage at node 9 was the lowest in the entire network. The DC current control scheme, the voltage at node 9, and the frequency variation curves of the high-voltage DC receiving-end grid are shown below. Figures 9-11 As shown.

[0129] from Figure 10 It can be seen that after applying the DC current control scheme, the voltage recovered to above 0.8 pu within 10 seconds after the fault; without control, the node voltage at the fault location was below 0.8 pu, and the grid transient voltage became unstable. The DC current control scheme generated by the emergency control model based on the GCN-DDPG fusion model resulted in faster node voltage recovery, with a reduction of 3.86 pu in DC transmission energy; while the DC current control scheme generated by the emergency control model based on the standard DDPG model reduced DC transmission energy by 4.32 pu. Therefore, the DC current control scheme generated by the emergency control model based on the GCN-DDPG fusion model can achieve faster and better node voltage recovery with less reduction in DC energy transmission. Other test samples showed similar results. Figure 11 It can be seen that due to the three-phase short-circuit ground fault occurring near node 9 of line 9-10 at t=0.6s, the reactive power consumption of the receiving-end power grid increased while the active power consumption decreased, resulting in a brief period of active power surplus and a frequency spike. At t=0.7s, the relay protection activated, causing the frequency to drop rapidly and recover to around 50Hz within 100ms, without triggering emergency frequency control measures such as load shedding. At t=0.9s, the emergency control for transient voltage stabilization of the receiving-end power grid proposed in this invention was initiated, and the DC current control scheme was issued to the DC control and protection system. Due to the reduced active power transmitted via DC, the frequency of the receiving-end power grid began to drop at t=0.9s, reaching a minimum of 49.83Hz around t=1.1s. The frequency then gradually recovered and eventually remained above 49.8Hz, indicating frequency stability of the receiving-end power grid. For other samples, it was verified that the receiving-end power grid also maintained frequency stability.

[0130] It is evident that, since transient voltage stability is a localized problem, the voltage drop is more pronounced at nodes near the fault. By representing the power grid topology using GCN, the GCN-DDPG fusion model, compared to the standard DDPG model, supplements the information on the connectivity relationships between nodes and characterizes the distance of each node from the fault location. Therefore, it can be considered that replacing some fully connected network layers in the DDPG model's network structure with GCN to represent the power grid topology achieves more effective information filtering and dimensionality reduction during graph convolution, thus facilitating the generation of more reasonable actions.

[0131] The generalization test results of the emergency control model based on the GCN-DDPG fusion model and the standard DDPG model are as follows: Figure 12 and Figure 13 As shown.

[0132] contrast Figure 12 and Figure 13 It can be seen that the generalization test sample return of the GCN-DDPG fusion model is greater than -300, and it can generate an effective emergency control strategy for samples that have not been trained. That is, the voltage recovers to above 0.8 pu in all generalization test samples. However, for the standard DDPG model, there are some generalization test rounds with a return of less than -500. Similarly, it can be seen that the voltage recovery is poor in these samples, and it may even fail to meet the transient voltage stability standard.

[0133] To further verify the effectiveness and superiority of the GCN-DDPG fusion model, a generalization test sample was randomly selected. The load level was set to 70%, generator G2 was not started, steady-state topology line 4-5 was disconnected, and a three-phase short-circuit ground fault occurred near node 14 of line 9-14. After the fault occurred, the voltage at node 14 was the lowest in the entire network. The generated DC current control scheme is as follows: Figure 14 As shown, the voltage at node 14 is as follows Figure 15 As shown.

[0134] from Figure 15 It can be seen that the DC transmission energy is reduced by 4.14 pu and 4.27 pu under the emergency control schemes generated by the GCN-DDPG fusion model and the standard DDPG model, respectively. However, when no control is applied or when the emergency control strategy generated by the standard DDPG model is applied, the node voltage at the fault location is lower than 0.8 pu, and the grid voltage becomes unstable. After applying the emergency control strategy based on the GCN-DDPG fusion model, the node voltage recovers to above 0.8 pu. It can be seen that when dealing with generalized test samples, i.e., sample types that did not appear during training, the emergency control model based on the GCN-DDPG fusion model achieves more stable and superior results in terms of voltage recovery and reduction of DC transmission power decline.

[0135] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0136] It should be understood that the above description of the embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art can make substitutions or modifications under the guidance of this invention without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. An emergency control method for transient voltage stability of a DC receiving-end power grid, characterized in that, Includes the following steps: Construct an emergency control model for transient voltage stability in a new type of DC receiving-end power system based on the GCN-DDPG fusion model; including: Collect power grid data, including power grid status information and topology; The collected data is divided into training set and test set. The training set data is input into the GCN-DDPG fusion model for training until convergence. The test set data is input into the trained GCN-DDPG fusion model, which can then quickly generate an emergency control plan based on the power grid status information, topology, and fault information. When training the GCN-DDPG fusion model, a dual-evaluation network mechanism is adopted, specifically: Based on the original DDPG evaluation network, an additional evaluation network with the same structure is added. These two evaluation networks are used to estimate the target Q-value, and the smaller Q-value is selected at each update. The two evaluation networks are denoted as follows: Q 1 and Q 2, the parameters are respectively and The GCN-DDPG fusion model algorithm is based on the above improvements, and its loss function is shown below: (1) In the formula, L The loss function; N This represents the training sample size for the network. For parameters The evaluation network below, therefore Indicates the state Apply action at the time The subsequent expected return, i.e., the parameter is The output of the evaluation network; For parameters The action network below represents the parameters. Under, state s The output of the policy network is the input; Perform an action while indicating a state Instant rewards; γ Discount factor; Real-time acquisition of power grid data and monitoring of power grid status; if a fault occurs, an emergency control strategy is generated based on the emergency control model, quantified and calculated online, and sent to the DC control and protection system for execution; otherwise, monitoring of the power grid status continues until a fault occurs.

2. The emergency control method for transient voltage stability of a DC receiving-end power grid according to claim 1, characterized in that, The GCN-DDPG fusion model includes Evaluation Network: Generates evaluations of actions based on the definition of the reward function; Action network: Used to generate emergency control plans based on power grid status information.

3. The emergency control method for transient voltage stability of a DC receiving-end power grid according to claim 1, characterized in that, The evaluation network includes evaluation network one and evaluation network two, which have the same structure and both include: Graph convolutional layers: used to input power grid state information and topology information, and to extract features from the input information; Fully connected network layer: used for Q-value fitting and generating action evaluation.

4. The emergency control method for transient voltage stability of a DC receiving-end power grid according to claim 1, characterized in that, Action networks include Graph convolutional layers: used to input power grid state information and topology information, and to extract features from the input information; Fully connected network layer: used to generate actions, i.e. emergency control plans.

5. The emergency control method for transient voltage stability of a DC receiving-end power grid according to claim 1, characterized in that, When training the GCN-DDPG fusion model, the evaluation network and the action network are updated asynchronously, specifically: During model training, an asynchronous update strategy is adopted between the evaluation network and the action network. The evaluation network is updated several times and stabilized before the action network is updated once.

6. The emergency control method for transient voltage stability of a DC receiving-end power grid according to claim 1, characterized in that, Treating the power grid as an environment, direct current I d As an action, it completes the definition of the state inputs, action outputs, and reward functions of the GCN-DDPG fusion model, specifically including... The state variables input for the GCN-DDPG fusion model are as follows: (2) In the formula, X That is U t , indicating in t The voltage vector at each node of the power grid at any given time. for t The adjacency matrix that characterizes the power grid topology at any given time. for The degree matrix; The state outputs of the GCN-DDPG fusion model, i.e., the actions, are as follows: (3) In the formula, Indicates in t The DC current control value is set at any given time, and the DC current range is set to [0.55, 1.0]. The following transient voltage stability criterion is adopted: during the transient process after a disturbance to the power system, the load bus voltage can recover to above 0.80 pu within 10 seconds after the fault; considering stability, stability margin, and stability control costs, the reward function... The expression is as follows: (4) In the formula, The moment of failure; Indicates at time t The lowest voltage value in the entire network, which usually corresponds to the voltage amplitude of the bus near the fault location; Indicates time t The action amount; within 10 seconds after the fault, if the lowest voltage of all nodes in the network is greater than 0.8pu after the action is applied, it indicates that the transient voltage has recovered to above the stable standard, and a positive reward is given, and the larger the voltage value, the greater the reward; if the lowest voltage of the node is still less than 0.8pu, a negative reward is given, and the smaller the voltage value, the smaller the reward; if 10 seconds after the fault, the lowest node voltage of the grid is still less than 0.8pu after the action is applied, a reward of -500 is given, and the training round ends.

7. An emergency control system for transient voltage stability of a DC receiving-end power grid, characterized in that, Including the following: Module 1: Configured for constructing an emergency control model for transient voltage stability in novel DC receiving-end power systems based on the GCN-DDPG fusion model; including: Collect power grid data, including power grid status information and topology; The collected data is divided into training set and test set. The training set data is input into the GCN-DDPG fusion model for training until convergence. The test set data is input into the trained GCN-DDPG fusion model, which can then quickly generate an emergency control plan based on the power grid status information, topology, and fault information. When training the GCN-DDPG fusion model, a dual-evaluation network mechanism is adopted, specifically: Based on the original DDPG evaluation network, an additional evaluation network with the same structure is added. These two evaluation networks are used to estimate the target Q-value, and the smaller Q-value is selected at each update. The two evaluation networks are denoted as follows: Q 1 and Q 2, the parameters are respectively and The GCN-DDPG fusion model algorithm is based on the above improvements, and its loss function is shown below: (5) In the formula, L The loss function; N This represents the training sample size for the network. For parameters The evaluation network below, therefore Indicates the state Apply action at the time The subsequent expected return, i.e., the parameter is The output of the evaluation network; For parameters The action network below represents the parameters. Under, state s The output of the policy network is the input; Perform an action while indicating a state Instant rewards; γ Discount factor; The second module is configured to collect grid data in real time and monitor the grid's operational status. If a fault occurs, it generates an emergency control strategy based on the emergency control model, performs online quantitative calculations, and sends it to the DC control and protection system for execution. Otherwise, it continues to monitor the grid's operational status until a fault occurs.

8. A readable storage medium, characterized in that, The system stores computer-executable instructions that, when executed by a processor, configure the processor to perform the method according to any one of claims 1-6.

Images