A method, device and equipment for generator tripping decision of a power system and a storage medium

Abstract

The application discloses a generator tripping decision method and device of a power system, equipment and a storage medium, relates to the technical field of power system control, and obtains topological structure data and running state data of a power grid and a preset fault set; the topological structure data and the preset fault set are used for homophonic division of generator sets in the power system, and a homophonic machine group is obtained; a double-layer Markov decision process model is constructed according to the homophonic machine group and the running state data; a hierarchical decision model is constructed according to the double-layer Markov decision process model; and the running state data is input into the hierarchical decision model, so that the hierarchical decision model determines a target homophonic machine group and a generator tripping control instruction, and the structural dimension reduction of an emergency generator tripping decision action space is realized, and the timeliness and accuracy of the emergency generator tripping decision are improved.

Description

Technical Field

[0001] This application relates to the technical field of power system control, and in particular to a power system tripping decision method, apparatus, equipment and storage medium. Background Technology

[0002] With the large-scale grid connection of high-proportion renewable energy sources, the rotational inertia of the power system decreases and its damping characteristics weaken, significantly increasing the difficulty of transient stability control. Emergency generator tripping, as the second line of defense for the safe and stable operation of the power system, directly impacts grid security through the accuracy and timeliness of its decision-making.

[0003] Currently, emergency generator tripping decisions mainly employ two technical solutions. One is based on offline pre-decision strategies, which generate a "fault-policy" lookup table by combining heuristic optimization algorithms with time-domain simulation for typical operating modes and anticipated fault sets, and execute the strategy when an actual fault occurs. The other solution introduces deep reinforcement learning, modeling generator tripping decisions as Markov decision processes, and achieving end-to-end control command generation through interaction between an agent and the simulation environment. However, these solutions still face many challenges in practical applications. For offline pre-decision schemes, the actual operating conditions of the power grid are complex and variable, and the static strategies generated offline are difficult to cover a massive number of unknown operating conditions. When encountering unexpected fluctuations in renewable energy output or complex faults, strategy mismatch is likely to occur. For deep reinforcement learning schemes, large-scale power grids contain dozens or even hundreds of generator units, and the generator tripping action space expands exponentially. The agent's exploration efficiency is extremely low in the early stages of training, making it difficult to converge to an effective strategy. In addition, the action output of existing data-driven models lacks physical constraints such as the power grid topology and generator synchronization characteristics, which can easily generate invalid generator tripping combinations across regions. The economic efficiency and reliability of the decision results are difficult to meet the requirements of practical engineering.

[0004] Therefore, improving the timeliness and accuracy of emergency aircraft switching decisions is an urgent problem that needs to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, device, and storage medium for power system tripping decisions, aiming to solve the technical problem of how to improve the timeliness and accuracy of emergency tripping decisions.

[0006] To achieve the above objectives, this application proposes a power system tripping decision method, the method comprising: Acquire power grid topology data, operational status data, and preset fault sets; Based on the topology data and the preset fault set, the generator sets in the power system are divided into coherent groups to obtain coherent generator groups. A two-level Markov decision process model is constructed based on the coherent machine group and the operating status data. A hierarchical decision model is constructed based on the described two-level Markov decision process model; The operational status data is input into the hierarchical decision model so that the hierarchical decision model can determine the target coherent machine group and the machine switching control command.

[0007] In one embodiment, the step of constructing a two-level Markov decision process model based on the coherent machine group and the operating state data includes: Based on the real-time relative power angle and preset stability threshold of each generator set in the operating status data, the normalized relative power angle feature value is calculated, and the state space is constructed based on the normalized relative power angle feature value. The action space is decoupled into a first action space and a second action space according to the coordinating machine group. The first action space is used to select the target coordinating machine group, and the second action space is used to determine the machine group to be cut off within the target coordinating machine group. Obtain the current decision step count and the total unit capacity cumulatively removed when the simulation reaches the preset total simulation duration during the iterative decision-making process based on the preset time-domain simulation state transition strategy; A comprehensive reward function is constructed based on the normalized relative power angle feature value, the preset stability threshold, the current decision step number, and the cumulative total capacity of the units removed. A two-layer Markov decision process model is constructed based on the state space, the first action space, the second action space, the comprehensive reward function, and the preset time-domain simulation state transition strategy.

[0008] In one embodiment, the step of decoupling the action space into a first action space and a second action space according to the coherent machine group includes: The coherent machine group is used as the first discrete action option, and a first action space is constructed based on the first discrete action option; The generator sets included within the synchronized generator group are used as the second discrete action option, and a second action space is constructed using the second discrete action option.

[0009] In one embodiment, the step of constructing a comprehensive reward function based on the normalized relative power angle feature value, the preset stability threshold, the current decision step number, and the cumulative total capacity of the units removed includes: The stability reward value is determined based on the normalized relative work angle characteristic value and the preset stability threshold. The time response penalty is determined based on the current decision step number and the preset single-step penalty constant. The tripping control cost is determined based on the cumulative total capacity of the units to be cut and the preset tripping cost weighting coefficient; A comprehensive reward function is constructed based on the stability reward value, the time response penalty, and the machine switching control cost.

[0010] In one embodiment, the step of dividing the generator sets in the power system according to the topology data and the preset fault set to obtain a coherent generator group includes: Based on the topology data, the generator sets in the power system are clustered to obtain an initial coherent generator group; The relative power angle discrete sampling sequence of each generator set within the observation window is obtained according to the preset fault set; Calculate the correlation coefficient of the discrete sampling sequence of relative power angle between any two generator sets within the initial synchronized generator group; The initial coherent machine group is divided according to the correlation coefficient to obtain the coherent machine group.

[0011] In one embodiment, the step of constructing a hierarchical decision model based on the two-level Markov decision process model includes: Based on the state space of the aforementioned two-layer Markov decision process model, a value network is constructed. Based on the first action space of the two-layer Markov decision process model, construct the first policy network; Based on the second action space of the two-layer Markov decision process model and the target coherent machine group output by the first policy network, a second policy network is constructed. A hierarchical decision-making model is constructed based on the value network, the first policy network, and the second policy network.

[0012] In one embodiment, the step of inputting the operating status data into the hierarchical decision model to enable the hierarchical decision model to determine the target coherent machine group and the machine switching control command includes: The operating status data is input into the first policy network in the hierarchical decision model so that the first policy network outputs the selection probability distribution of each coherent machine group, and the target coherent machine group is determined according to the selection probability distribution. The operating status data and the identification information of the target coherent machine group are concatenated to obtain a joint feature; The joint features are input into the second policy network in the hierarchical decision model so that the second policy network outputs the initial probability distribution of each generator set in the target coordinating generator group. The initial probability distribution of the machine switching is input into the mask matrix in the hierarchical decision model and multiplied element by element and renormalized to determine the machine switching control command based on the probability distribution after renormalization.

[0013] Furthermore, to achieve the above objectives, this application also proposes a power system tripping decision device, the device comprising: The data acquisition module is used to acquire power grid topology data, operating status data, and preset fault sets; The coherence partitioning module is used to partition the generator sets in the power system according to the topology data and the preset fault set to obtain a coherent generator group. The decision process construction module is used to construct a two-layer Markov decision process model based on the coherent machine group and the operating status data. The decision model construction module is used to construct a hierarchical decision model based on the two-level Markov decision process model. The decision execution module is used to input the operating status data into the hierarchical decision model so that the hierarchical decision model can determine the target coordinating machine group and the machine switching control command.

[0014] In addition, to achieve the above objectives, this application also proposes a power system tripping decision device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the power system tripping decision method as described above.

[0015] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the power system tripping decision method as described above.

[0016] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the power system tripping decision method as described above.

[0017] This application provides a method for power system generator tripping decision-making. The method includes: acquiring power grid topology data, operating status data, and a preset fault set; classifying generator units in the power system according to their coherence based on the topology data and the preset fault set to obtain coherent generator groups; constructing a two-layer Markov decision process model based on the coherent generator groups and the operating status data; constructing a hierarchical decision model based on the two-layer Markov decision process model; and inputting the operating status data into the hierarchical decision model to determine the target coherent generator group and the generator tripping control command. In summary, this application, by dividing all generator units in the power grid into several groups according to their coherence characteristics, decouples the global high-dimensional generator tripping action space into a two-layer action space, and constructs a hierarchical decision model for progressive decision-making. This achieves structured dimensionality reduction of the emergency generator tripping decision action space, solves the problem of decision-making difficulties caused by the high-dimensional action space of a large-scale power grid, and improves the timeliness and accuracy of emergency generator tripping decisions. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the first embodiment of the power system tripping decision method of this application; Figure 2 This is an improved IEEE 39-node system topology diagram in one embodiment of the power system tripping decision method of this application; Figure 3 This is a topology diagram of a hierarchical decision model in one embodiment of the power system tripping decision method of this application; Figure 4 A flowchart illustrating the second embodiment of the power system tripping decision method of this application; Figure 5 This is a logic diagram of two-layer action space decoupling and topology mask transmission in one embodiment of the power system tripping decision method of this application; Figure 6 A flowchart illustrating the third embodiment of the power system tripping decision method of this application; Figure 7 This is a schematic diagram of the algorithm training and network parameter update process in one embodiment of the power system tripping decision method of this application; Figure 8 This is a schematic diagram of the module structure of the power system tripping decision device according to an embodiment of this application; Figure 9 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the power system tripping decision method in the embodiments of this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0023] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0024] It should be noted that the executing entity in this embodiment can be a power system tripping decision system, a computing service device with data processing, network communication, and program execution functions, or an electronic device capable of realizing the aforementioned power system tripping decision function, etc. This embodiment does not specifically limit it in this way. The following uses a power system tripping decision system as an example to describe this embodiment and the following embodiments.

[0025] Based on this, embodiments of this application provide a power system tripping decision method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the power system tripping decision method of this application.

[0026] In this embodiment, the power system tripping decision method includes steps S10 to S50: Step S10: Obtain the topology data and operating status data of the power grid, as well as the preset fault set.

[0027] It should be noted that the power grid topology data and operational status data are obtained from the power grid dispatch automation system or data acquisition and monitoring system. The preset fault set is pre-configured based on the actual operational needs of the power grid or historical fault statistics. Specifically, topology data refers to data describing the electrical connections between various power devices in the power grid (including generator sets, transformers, transmission lines, buses, etc.) and the inherent attribute parameters of each device, including but not limited to: the node-branch relationships of the power grid, the connection location of each generator set to the bus, the impedance parameters of transmission lines, and the turns ratio and impedance parameters of transformers. Operational status data refers to the monitoring data of various electrical quantities during the real-time operation of the power grid, including but not limited to: the real-time power angle, active power output, reactive power output, and terminal voltage of each generator; the voltage amplitude and phase angle of each bus; and the active and reactive power flow of each line. The preset fault set refers to a set of anticipated fault scenarios used to test and train the trip decision model, including but not limited to: fault location (such as which transmission line or busbar the fault occurs on), fault type (such as three-phase short circuit, two-phase short circuit, single-phase ground fault, etc.), fault duration, and fault occurrence time.

[0028] Step S20: Based on the topology data and the preset fault set, perform coordination division of the generator sets in the power system to obtain a coordinating generator group.

[0029] It should be noted that in this step, the system will classify all generator sets in the power system according to the acquired topology data and preset fault set, and group generator sets with close electrical connections and consistent transient response characteristics into several co-coordinated groups.

[0030] Additionally, it should be noted that coherence classification refers to the process of grouping generator sets in a power system that exhibit similar power angle swing characteristics after being subjected to disturbances, with their relative swing angles not exceeding a given threshold, into the same group. A coherent generator group refers to a collection of generators whose power angle trajectories maintain synchronous swing during the transient process after a significant disturbance to the power system; generator sets within the same coherent generator group have similar dynamic response characteristics during transient processes and can be controlled as a whole in emergency tripping control.

[0031] In one feasible implementation, step S20 specifically includes: Step S201: Cluster the generator sets in the power system according to the topology data to obtain the initial synchronized generator group.

[0032] It should be noted that the system will cluster electrically connected generator units into initial synchronized groups based on the inherent topology and regional operation characteristics of the power grid, combined with the backbone network layout, substation electrical connections, and the geographical distribution of generator units. Furthermore, it should be noted that the initial synchronized groups refer to preliminary clustering results obtained solely based on the static topology of the power grid, and have not yet undergone transient trajectory correlation verification. Under specific fault conditions, the units within these initial synchronized groups may exhibit inconsistent dynamic responses.

[0033] Step S202: Obtain the discrete sampling sequence of the relative power angle of each generator set within the observation window according to the preset fault set.

[0034] It should be noted that, for each fault scenario in the preset fault set, power system transient simulation software (such as DSP (Digital Simulator Power)) is called to perform time-domain simulation calculations. Within the set observation window, the power angle data of each generator set relative to the system inertia center are collected at a fixed sampling interval to form a discrete sampling sequence.

[0035] Additionally, it should be noted that the relative power angle discrete sampling sequence refers to the discrete data sequence formed by arranging the relative power angle values ​​of each generator set relative to the system's inertia center reference coordinate system in chronological order during the transient process. It is used to reflect the degree to which each generator set deviates from the system's average motion state.

[0036] Step S203: Calculate the correlation coefficient of the discrete sampling sequence of relative power angle between any two generator sets in the initial synchronized generator group.

[0037] It should be noted that, for each initial synchronized generator group, the system calculates the Pearson correlation coefficient between the discrete sampling sequences of the relative power angles of any two generator units within the group, in order to quantify the degree of consistency in the dynamic response of the two units during transient processes. Understandably, this step is used to identify units that are topologically adjacent but exhibit significant differences in their dynamic response under specific fault conditions.

[0038] Specifically, the Pearson correlation coefficient is a statistic that measures the degree of linear correlation between two discrete sampling sequences of relative work angles, denoted as . The calculation formula is as follows:

[0039] Where N is the total number of discrete sampling points in the trajectory sequence; k is the sampling point index; , These are the average values ​​of the relative power angle sequences of units i and j within the observation window, respectively. , are the standard deviations of the relative power angle sequences of the i-th and j-th generators, respectively; , These are the relative power angle sampling values ​​of the i-th and j-th generators at the k-th sampling time, respectively. It is understood that the correlation coefficient is a statistical indicator used to measure the degree of linear correlation between two variables; in this application, it is used to quantify the similarity of the power angle trajectories of any two generator sets during a transient process.

[0040] Step S204: Divide the initial coherent machine group according to the correlation coefficient to obtain the coherent machine group.

[0041] It should be noted that the system compares the calculated correlation coefficients with preset thresholds. If the correlation coefficients of two units within an initial synchronized generator group are lower than the preset threshold, the initial synchronized generator group is split into two subgroups; conversely, if the correlation coefficients are higher than or equal to the preset threshold, the current grouping remains unchanged. After pairwise verification and iterative splitting, the final synchronized generator group division result is obtained. This step ensures a high degree of consistency in the dynamic response of units within the synchronized generator group under specific fault conditions.

[0042] Additionally, it should be noted that the preset threshold is a critical value of the correlation coefficient set according to the power grid scale, unit characteristics and control accuracy requirements. The higher the threshold, the more refined the grouping and the more significant the dimensionality reduction effect of the action space, but the number of units will also increase accordingly.

[0043] Step S30: Construct a two-layer Markov decision process model based on the coherent machine group and the operating status data.

[0044] It should be noted that in this step, the system models the power system emergency generator tripping control problem as a two-level Markov decision process model based on the results of the coherent generator group partitioning and operating status data. This specifically includes state space construction, hierarchical action space construction, and comprehensive reward function design. Understandably, the purpose of this step is to provide a formal mathematical framework for the learning and decision-making of the hierarchical decision model.

[0045] Additionally, it should be noted that the two-level Markov decision process model refers to extending the traditional single-level Markov decision process into a Markov decision process model with a hierarchical decision structure. Its decision process is decomposed into two levels: upper-level decision and lower-level decision. The upper-level decision is responsible for selecting the target object at the macro level (in this application, selecting the target synchronized generator group requiring generator cut-off control). The lower-level decision is responsible for executing specific micro-actions within the target object (in this application, determining which generator units to cut off within the selected synchronized generator group). The two levels of decisions are executed sequentially, with the lower-level decision based on the output of the upper-level decision, forming a progressive hierarchical decision structure. For example, as... Figure 2As shown, in the improved IEEE 39-node system, five coherent machine groups are obtained through coherence partitioning, including group 1 (G8, G9, G19), group 2 (G10, G11, G20), group 3 (G17, G18, G21), group 4 (G1, G2) and group 5 (G6, G7).

[0046] Step S40: Construct a hierarchical decision model based on the two-layer Markov decision process model.

[0047] It should be noted that in this step, the system constructs a hierarchical decision model based on a two-layer Markov decision process model to solve the decision process. This hierarchical decision model is based on the Actor-Critic reinforcement learning framework and includes a value network (i.e., the Critic network), a first policy network (i.e., the upper-layer Actor network), and a second policy network (i.e., the lower-layer Actor network). The value network is used to evaluate the long-term expected return of the current grid state; the first policy network is used to select the target coordinating generator group from the first action space based on the current system state; and the second policy network is used to select specific generator tripping control commands from the second action space based on the current system state and the target coordinating generator group output by the first policy network.

[0048] In one feasible implementation, step S40 specifically includes: Step S401: Construct a value network based on the state space of the two-layer Markov decision process model.

[0049] It should be noted that, as Figure 3 As shown, the system constructs a fully connected neural network as the value network, consisting of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is consistent with the dimension of the state space, and it receives the normalized relative power angle feature vectors of all generators in the network. The hidden layer adopts a two-layer fully connected structure, with 128 nodes in the first hidden layer and 64 nodes in the second hidden layer. Each hidden layer uses the ReLU (Rectified Linear Unit) activation function for nonlinear feature transformation. The output layer contains a linear neuron, which outputs a scalar state value estimate of the current system state. .

[0050] Additionally, it should be noted that the value network primarily evaluates the current power grid state (i.e., the global state vector). The long-term expected return, the network parameters are denoted as The output is used to compute the generalized advantage estimate to reduce the variance of the policy gradient estimate.

[0051] Step S402: Construct a first policy network based on the first action space of the two-layer Markov decision process model.

[0052] It should be noted that, as Figure 3 As shown, the system constructs a fully connected neural network containing an input layer, hidden layers, and an output layer as the first policy network. The first policy network is the top-level Actor module for hierarchical decision-making, and its network parameters are... Its core function is to select the target cluster at a macroscopic level. The number of neurons in the input layer is consistent with the dimension of the state space, and it is used to receive the global state vector of the system. The hidden layers employ a three-layer fully connected structure, with 128, 128, and 64 neurons in each layer, respectively. Each layer uses the ReLU activation function. The number of neurons in the output layer is the same as the number of coordinating machines, i.e., equal to the dimension of the first action space. The output layer uses the Softmax activation function to map the output value of each neuron to a probability value between 0 and 1, and the sum of all output probability values ​​is 1, thus forming a discrete probability distribution of each coordinating machine being selected as the target. .

[0053] Step S403: Construct a second policy network based on the second action space of the two-layer Markov decision process model and the target coherent machine group output by the first policy network.

[0054] It should be noted that, as Figure 3 As shown, the system constructs a fully connected neural network containing an input layer, hidden layers, and an output layer as the second policy network. The second policy network is responsible for outputting specific switching instructions, and its parameters are denoted as... The number of neurons in the input layer is equal to the dimension of the state space plus one. It receives the joint feature vector formed by concatenating the system's global state vector with the identifier information of the target coherent generator group output by the first policy network. The hidden layers employ a three-layer fully connected structure, with 256, 128, and 64 neurons in each layer, respectively. Each layer uses the ReLU activation function. The number of neurons in the output layer is equal to the total number of generator groups in the system, and it outputs the initial probability of each generator group being disconnected.

[0055] Additionally, it should be noted that the second policy network also includes a masking layer, used to generate a mask matrix corresponding to the system's generator topology based on the target coherent generator group. Specifically, if a generator belongs to the target coherent generator group, the element corresponding to that generator in the mask matrix is ​​1; if a generator does not belong to the target coherent generator group, the element corresponding to that generator in the mask matrix is ​​0. The initial probability matrix output by the second policy network's output layer is then... With mask matrix Element-wise multiplication (Hadamard product) is performed, then the action probabilities corresponding to non-target generator groups are set to zero. Finally, the remaining probabilities are renormalized so that the sum of all action probabilities is 1, thus obtaining the final probability distribution of each generator unit being disconnected. Specific disconnection commands are then sampled and generated based on this final probability distribution. .

[0056] Step S404: Construct a hierarchical decision model based on the value network, the first policy network, and the second policy network.

[0057] It should be noted that, as Figure 3 As shown, in this step, the system combines the constructed value network, first policy network, and second policy network according to the Actor-Critic framework to form a complete hierarchical decision-making model. Specifically, the input of the hierarchical decision-making model receives the global state vector. Global state vector Simultaneously input the value network, the first policy network, and the second policy network. The first policy network is based on... Output the target coherent cluster selection probability distribution and sample to obtain the target coherent cluster index. The input to the second policy network simultaneously receives the global state vector. Index of target coherent clusters The two are concatenated to form a joint feature vector. , The second strategy network outputs the initial switching probability matrix. Through the mask matrix The final machine switching command is output after processing. The value network receives the global state vector. Output state value estimate The parameters of the hierarchical decision model are updated through end-to-end collaborative training.

[0058] Additionally, it should be noted that the optimization objectives of the hierarchical decision-making model include: generalized advantage estimation, upper-level policy loss function, lower-level policy loss function, and value network loss function.

[0059] Specifically, generalized advantage estimation refers to the method used to evaluate the current policy in the current state (i.e., the global state vector). The generalized advantage estimation (GOE) is a quantitative indicator of the superiority or inferiority of an action relative to the average level. Its core function is to reduce the variance of gradient estimation during policy gradient updates while maintaining a reasonable level of bias, thereby improving training stability. GOE is based on the value network's prediction of state value and is calculated using an exponentially weighted average of temporal difference residuals. The formula is:

[0060] in, As a discount factor, The immediate reward obtained after performing the action at time t. For the value network to estimate the value of the state in the next time step, For the value network, the current state (i.e., the global state vector) is... Value estimate.

[0061] Then, define the generalized advantage estimate at step t. For the exponentially weighted average of the multi-step TD residuals:

[0062] in, This is a smoothing factor for generalized dominance estimation, used to control the balance weights between bias and variance. Let T be the temporal difference residual at step t+k, and let T represent the total number of decision steps in a single simulation round (or trajectory).

[0063] Understandably, the calculated generalized dominance estimate... As a global evaluation signal, it synchronously guides the gradient updates of the upper and lower policy networks: when When the value is positive, it indicates that the current action is better than the average level, and the sampling probability of the corresponding layer action should be increased; when When the value is negative, it indicates that the current action is worse than the average level, and the probability of choosing this action should be reduced.

[0064] Specifically, the upper-level policy loss function refers to a quantitative function used in reinforcement learning training to measure the difference between the current decision policy and the historical decision policy of the upper-level policy network (i.e., the first policy network). Its core objective is to maximize the expected cumulative reward while limiting the magnitude of policy updates through a pruning mechanism to avoid training instability or policy collapse due to excessively large single updates.

[0065] The optimization of the upper-level policy network aims to maximize the expected return while avoiding excessive parameter updates. First, the policy probability ratio between the old and new policies is defined:

[0066] in, The current upper-layer policy network parameters Below, in state Select target harmonic group The probability of; To update the network parameters of the old upper-layer policy Below, in state Select target harmonic group The probability of; The probability ratio between the two reflects the magnitude of the policy update.

[0067] Loss function of upper-layer policy network It consists of a policy pruning loss term and an entropy regularization term, and its expression is:

[0068] in, Indicates the probability ratio Crop to range [ ]Inside, This is the pruning threshold, used to limit the maximum magnitude of policy updates; Shannon entropy for the upper-level strategy, used to measure the degree of exploration of the strategy; This is the entropy regularization coefficient, used to control the balance weight between exploration and exploitation.

[0069] Specifically, the lower-level policy loss function refers to a quantification function used during reinforcement learning training to measure the difference between the current decision policy and the historical decision policies of the lower-level policy network (i.e., the second policy network). Similar to the upper-level policy loss function, the lower-level policy loss also adopts the form of combining pruned agent target and entropy regularization term. However, the difference is that the output probability of the lower-level policy needs to be corrected by topological mask and renormalized. Therefore, only the final execution probability after mask correction is used when calculating the probability ratio.

[0070] Lower-level policy network The output is constrained by the selection of the upper-layer cluster, and its action probability needs to be corrected by the topology mask and renormalized. Therefore, when calculating the probability ratio of the old and new strategies, only the final execution probability after mask correction is used, as shown in the formula:

[0071] in, This represents the probability ratio of the lower-level strategy.

[0072] Since the lower-level policy uses the same optimization logic as the upper-level policy, its loss function also combines the pruning of the proxy objective with an entropy regularization term. The expression is:

[0073] The meanings of each symbol are consistent with the corresponding symbols in the upper-level policy loss function; the only difference lies in the probability ratio of the lower-level policy. It is calculated based on the final probability after mask correction.

[0074] Specifically, the value network loss function refers to a quantitative function used to measure the accuracy of the value network (Critic network) in estimating the state value of the system. Its optimization objective is to minimize the mean square error between the predicted state value and the actual cumulative discounted return, so that the value network can provide a more accurate state value benchmark for the policy network and reduce the bias of the advantage function estimation.

[0075] Value Network Responsible for fitting the system state The long-term expected return provides a benchmark for advantage function estimation. Value network loss function. The expression is:

[0076] in, This represents the actual cumulative discount reward obtained from the current time t until the end of the simulation.

[0077] The final algorithm will use the Adam optimizer to simultaneously update the parameters of the upper-level policy, lower-level policy, and value network by minimizing the three types of loss functions mentioned above.

[0078] Step S50: Input the operating status data into the hierarchical decision model so that the hierarchical decision model can determine the target coordinating machine group and the machine switching control command.

[0079] It should be noted that in this step, the system will convert the real-time collected power grid operation status data into a normalized relative power angle feature vector according to the normalization processing method, and use this feature vector as input data to construct the hierarchical decision model. The hierarchical decision model generates the generator tripping decision according to the following process: The first strategy network receives the input state vector, outputs the selection probability distribution of each coordinating generator group, and determines the target coordinating generator group based on the sampling of the probability distribution; The second strategy network receives the joint feature vector after concatenating the state vector and the target coordinating generator group identification information, outputs the initial generator tripping probability distribution of each generator group, and after element-wise multiplication and renormalization processing by the mask matrix, determines the specific target coordinating generator group and generator tripping control command that need to be executed for emergency generator tripping based on the sampling of the final probability distribution.

[0080] This embodiment provides a method for power system generator tripping decision-making. The method includes: acquiring power grid topology data, operating status data, and a preset fault set; classifying generator units in the power system according to their coherence based on the topology data and the preset fault set to obtain coherent generator groups; constructing a two-layer Markov decision process model based on the coherent generator groups and the operating status data; constructing a hierarchical decision model based on the two-layer Markov decision process model; and inputting the operating status data into the hierarchical decision model to determine the target coherent generator group and the generator tripping control command. In summary, this embodiment, by dividing all generator units in the power grid into several groups according to their coherence characteristics, decouples the global high-dimensional generator tripping action space into a two-layer action space, and constructs a hierarchical decision model for progressive decision-making, achieves structured dimensionality reduction of the emergency generator tripping decision action space. This solves the problem of decision-making difficulties caused by the high-dimensional action space of a large-scale power grid, and improves the timeliness and accuracy of emergency generator tripping decisions.

[0081] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 4 , Figure 4 This is a flowchart illustrating the second embodiment of the power system tripping decision method of this application. Step S30 includes: Step S301: Calculate the normalized relative power angle feature value based on the real-time relative power angle and preset stability threshold of each generator set in the operating status data, and construct the state space based on the normalized relative power angle feature value.

[0082] It should be noted that in this step, the system extracts the real-time relative power angle of each generator unit relative to the system's center of inertia coordinate system at time t from the operating status data. Substituting the real-time relative power angle of each unit into the normalization calculation formula, the system obtains the normalized relative power angle characteristic value of each unit. The normalized relative power angle characteristic values ​​of all generator units are then combined into a multi-dimensional state vector according to the unit number order, which serves as the agent's observed state at time t.

[0083] Additionally, it should be noted that the preset stability threshold is a power angle limit set according to the power system power angle stability engineering criterion, used to normalize the original power angle data and determine the transient stability of the system. The state space is the set of all possible states that the agent can observe during the Markov decision-making process; in this application, it is represented by a multi-dimensional continuous space composed of the normalized relative power angle eigenvalues ​​of all generating units in the network.

[0084] Specifically, the formula for calculating the normalized relative work angle eigenvalue is as follows:

[0085] in, The total number of generator sets in the system. Let be the real-time relative power angle of the i-th generator unit at time t. Let be the normalized relative power angle characteristic value of the i-th generator unit at time t. This is a preset stability threshold.

[0086] Finally, the agent's global state vector at time t From the entire network The normalized relative power angle eigenvalues ​​of the generator sets are represented as follows:

[0087] Step S302: Decouple the action space into a first action space and a second action space according to the coherent machine group. The first action space is used to select the target coherent machine group, and the second action space is used to determine the machine group to be cut off within the target coherent machine group.

[0088] It should be noted that the first action space refers to the upper-level discrete action space composed of all coordinating generator groups as optional actions, and its dimension is equal to the total number of coordinating generator groups; the second action space refers to the lower-level discrete action space composed of the generator units contained within the target coordinating generator group as optional actions, and its dimension is equal to the number of generator units within the target generator group. In this step, based on the division result of the coordinating generator groups, the system decouples the original global high-dimensional action space containing all generator unit combinations into two nested low-dimensional action subspaces according to the hierarchical logic of "selecting groups first, then selecting generators". The upper-level action space is responsible for selecting the target disturbed generator group from several coordinating generator groups, and the lower-level action space is responsible for determining the specific generator units to be disconnected within the selected target generator group.

[0089] Understandably, this hierarchical decoupling mechanism compresses the exponentially expanding global action space into a combination of local subspaces, effectively alleviating the solution bottleneck caused by the high-dimensional action space and reducing the agent's ineffective exploration in the huge solution space.

[0090] In one feasible implementation, step S302 specifically includes: Step A10: Use the coherent machine group as the first discrete action option, and construct the first action space with the first discrete action option.

[0091] It should be noted that, based on the aforementioned grouping results based on generator unit compatibility, the system pre-divides all generator units in the network into M coherent generator groups with strong electrical coupling characteristics. Each coherent generator group is treated as an independent first discrete action option, and all M coherent generator groups together constitute the first action space. Specifically, the system assigns a unique index number to each coherent generator group, and the set of all index numbers constitutes the first action space. Here, the first discrete action option refers to the upper-level optional action with a single coherent generator group as the basic unit, and each option corresponds to a generator group set containing several strongly electrically coupled generator units; the first action space refers to the set composed of M first discrete action options, i.e., {group 1, group 2, ..., group M}, in which the first strategy network performs generator group selection decisions.

[0092] Additionally, it should be noted that in subsequent steps, based on the aforementioned clustering results, the first policy network will use the global state vector... As input, the output is an M-dimensional discrete probability distribution, which is used to select the target aircraft group index to be executed for emergency aircraft switching control at the current moment. The expression is:

[0093] in, For the target coherent cluster index, This represents the first policy network (i.e., the upper-level policy network).

[0094] Step A20: Take the generator sets included in the synchronized generator group as the second discrete action option, and construct the second action space with the second discrete action option.

[0095] It should be noted that for each synchronized generator group, the system treats all generators within that synchronized generator group as second discrete action options, constructing a second action space corresponding to that group. Specifically, each generator in the group is assigned a unique generator number, and the set of all generator numbers constitutes the group's second action space. Here, the second discrete action option refers to a lower-level selectable action based on a single generator group, with each option corresponding to a specific generator group; the second action space refers to the set of all generators within the target synchronized generator group, where the second policy network (i.e., the lower-level policy network) makes generator removal decisions.

[0096] Additionally, it should be noted that in subsequent steps, the lower-level policy network... The input is the global state. Index of machine clusters in sync with upper-level targets The concatenated vector. To strictly constrain the tripping control actions within the selected unit group, the model introduces a masking mechanism based on topology mapping. Let the two-dimensional initial probability matrix of each unit's tripping action output by the lower-layer network be... Based on this, a mask matrix corresponding to the unit topology is constructed. The mask vector corresponding to the i-th generator is defined as follows:

[0097] The initial probability matrix output by the lower layer network After applying the mask constraints described above, the final probability distribution of the machine switching action is obtained through probability renormalization. As shown in the following formula:

[0098] in, It represents the Hadamardi (or Hadama) stack.

[0099] Understandably, the above steps can directly reduce the probability of tripping actions for non-target aircraft groups to zero, while fully preserving the relative proportions of action probabilities for each unit within the target aircraft group, thus achieving rigid transmission and unbiased execution of hierarchical control commands between upper and lower layers. Specifically, the decoupling of the two-layer action space and the topology mask transmission logic based on homogeneous grouping are as follows: Figure 5 As shown.

[0100] Step S303: Obtain the current decision step number and the total unit capacity cumulatively removed when the simulation reaches the preset total simulation duration during the iterative decision-making process based on the preset time-domain simulation state transition strategy.

[0101] It should be noted that during the closed-loop interaction between the agent and the power system simulation environment, the system records two parameters in real time: the current decision-making step count, i.e., the number of decisions the agent has executed since the start of this round of decision-making; and the cumulative total capacity of units disconnected, i.e., the sum of the rated capacities of all units disconnected during this round of decision-making. These two parameters are used to subsequently calculate the time response penalty and the unit disconnection control cost, respectively.

[0102] Additionally, it should be noted that the current decision step count refers to the cumulative number of decisions executed by the agent from the initial state during the current training round or online decision-making process, used to measure the timeliness of the decision response; the cumulative total capacity of the units cut off refers to the sum of the rated active power capacities of all generator units that have been cut off during the current decision-making process, used to quantify the economic cost of the unit cut-off control; the time-domain simulation state transition strategy refers to a strategy mechanism that uses power system transient simulation software as an environmental interaction medium, deduces the electromechanical transient response process of the system under the action of unit cut-off through time-domain numerical integration, and feeds back the new state and immediate rewards.

[0103] Step S304: Construct a comprehensive reward function based on the normalized relative power angle feature value, the preset stability threshold, the current decision step number, and the cumulative total capacity of the units removed.

[0104] It should be noted that the comprehensive reward function is used to evaluate the quality of the actions performed by the agent at each decision-making moment. Its design goal is to guide the agent to minimize the cost of tripping control while satisfying the system's transient stability constraints. Specifically, the comprehensive reward function consists of three components: the stability final value reward, the time response penalty term, and the tripping control cost. The stability final value reward guides the agent to prioritize the system's transient power angle stability as the primary decision-making constraint; the time response penalty term prompts the agent to quickly execute effective tripping actions in the early stages of a fault; and the tripping control cost guides the agent to minimize the impact on power supply and economic losses while satisfying stability constraints.

[0105] In one feasible implementation, step S304 specifically includes: Step B10: Determine the stability bonus value based on the normalized relative work angle characteristic value and the preset stability threshold.

[0106] It should be noted that the system stability determination result is a binary reward / penalty term set based on the system's final stable state at the end of the simulation time domain. Its core function is to guide the agent to take the system's transient power angle stability as the primary constraint in decision-making. The expression for the stability reward value is:

[0107] in, , These are the preset reward values ​​for when the system maintains stability and the penalty values ​​for when instability occurs; This represents the total duration of the transient simulation. Let be the normalized relative power angle characteristic value of the i-th unit at time t.

[0108] Step B20: Determine the time response penalty based on the current decision step number and the preset single-step penalty constant.

[0109] It should be noted that emergency power system shutdown is highly time-sensitive. If the agent cannot provide an effective strategy within a short period, the system will quickly cross the stability boundary. To guide the agent to achieve stable control with the fewest decision steps, the system constructs a time response penalty term linearly related to the current number of decision steps. The expression for the time response penalty is:

[0110] in, The time response penalty constant for a pre-defined single-step decision is... This represents the current decision-making step.

[0111] Understandably, the purpose of this step is to force the agent to make an effective switching decision as soon as possible in the early stages of a failure by gradually accumulating negative penalties, thereby avoiding repeated trial and error that could lead to a continuous deterioration of the system state and improving the timeliness of the decision.

[0112] Step B30: Determine the cut-off control cost based on the cumulative total capacity of the units cut off and the preset cut-off cost weighting coefficient.

[0113] It should be noted that the expression for the cutting machine control cost is:

[0114] in, The weighting coefficients are preset for the cost of switching machines. This represents the total capacity of units that have been cumulatively removed.

[0115] Understandably, the larger the total capacity of units cumulatively disconnected, the greater the absolute value of the disconnection control cost. This penalty term is used to guide the agent to minimize the disconnection capacity while satisfying stability constraints, thereby reducing the power supply impact and economic losses caused by disconnection operations.

[0116] Step B40: Construct a comprehensive reward function based on the stability reward value, the time response penalty, and the machine switching control cost.

[0117] It should be noted that the system will perform a weighted summation of the obtained stability reward value, time response penalty, and switchover control cost to construct a comprehensive reward function. The expression for the comprehensive reward function is:

[0118] in, For the total prize, This is a stability bonus value; Time-based response penalty item; To incentivize cost control during the cutting process.

[0119] Step S305: Construct a two-layer Markov decision process model based on the state space, the first action space, the second action space, the comprehensive reward function, and the preset time-domain simulation state transition strategy.

[0120] It should be noted that the electromechanical transient processes of power systems after large disturbances exhibit strong nonlinearity and strong coupling characteristics, making it difficult for conventional mathematical analytical models to accurately describe their complex state transition processes. Therefore, this solution utilizes a Python-DSP co-simulation platform as the interactive environment for reinforcement learning, implementing state transitions in Markov decision-making processes through time-domain simulation. In this step, the system integrates the state space, first action space, second action space, comprehensive reward function, and time-domain simulation state transition strategy constructed in the preceding steps to build a two-layer Markov decision-making process model.

[0121] Specifically, in closed-loop interactive training, the specific execution steps of a single state transition are as follows: (1) Instruction issuance: The agent outputs a tripping action instruction at time t, and the interface program automatically converts it into a control card that the DSP software can recognize and updates it in the simulation configuration file. (2) Simulation calculation: The DSP transient calculation kernel is called to execute the electromechanical transient time-domain numerical integration and calculate the transient response process of the power grid under the tripping action. The DSP transient simulation kernel is called to execute the tripping instruction, and the electromechanical transient response process of the system after being controlled is deduced through time-domain numerical integration. (3) State feedback loop: After the simulation of the current step size is completed, the program will automatically parse the DSP output results and extract the relative power angle data of the entire grid units. This data is used on the one hand to construct the new observation state of the agent at time t+1, and on the other hand to evaluate the transient stability of the system and calculate the immediate reward, thereby completing a complete state transition loop.

[0122] In this embodiment, by decoupling the global high-dimensional action space into a two-layer structure of upper-layer group selection and lower-layer machine selection, the dimensionality of the action space is effectively reduced. At the same time, by introducing a comprehensive reward function based on the system's stable state, the number of decision steps, and the cost of machine switching, the problems of convergence difficulties caused by action space explosion and lack of physical constraints on decision results in conventional deep reinforcement learning methods in power grid emergency machine switching scenarios are solved, thereby improving the training efficiency and decision speed of the agent in large-scale power system emergency control scenarios.

[0123] Based on the first and second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to that in embodiments one and two above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 , Figure 6 This is a flowchart illustrating the third embodiment of the power system tripping decision method of this application. Step S50 includes: Step S501: Input the operating status data into the first policy network in the hierarchical decision model so that the first policy network outputs the selection probability distribution of each coherent machine group, and determine the target coherent machine group according to the selection probability distribution.

[0124] It should be noted that the selection probability distribution refers to the discrete probability distribution of all coordinating generator groups output by the first strategy network based on the current system state. Each component represents the probability value of the corresponding coordinating generator group being selected as the control target, and the sum of the probability values ​​of all components is 1. The target coordinating generator group refers to the coordinating generator group selected as the emergency generator tripping control object at the current moment, which is the output result of the upper-level decision. Subsequent lower-level decisions will only select specific generator units to be cut off within this group. In this step, the system will input the normalized operating state data of the power grid at the current moment, which is a state vector composed of the normalized relative power angle characteristic values ​​of all generator units in the network. This state vector is input into the first strategy network (i.e., the upper-level strategy network). This network adopts a fully connected feedforward architecture. The input layer is connected to the global state vector of the corresponding dimension, and the hidden layer is set with a three-layer fully connected structure with 128, 128, and 64 neurons respectively. Each layer is configured with the ReLU activation function, and the output layer dimension is consistent with the number of coordinating generator groups M in the system. After normalization using the Softmax function, the output layer generates an M-dimensional discrete probability distribution, representing the probability that each coherent aircraft group is selected as the current emergency takeover target aircraft group. Subsequently, sampling is performed based on this probability distribution to obtain the index of the target coherent aircraft group.

[0125] Step S502: The operating status data and the identification information of the target coherent machine group are concatenated to obtain the joint feature.

[0126] It should be noted that the joint feature refers to the composite input feature formed by vector concatenation of global operational state data and target coherent machine group identification information. Its dimension is the sum of the state vector dimension and the number of coherent machine groups, used to characterize the joint decision-making conditions of the current system state and the selected target machine group. In this step, the global state vector and the target coherent machine group identification information are concatenated. Specifically, the target coherent machine group identification information is represented using one-hot encoding, i.e., generating an M-dimensional identification vector, where the element corresponding to the target machine group is 1, and the elements at other positions are 0. This identification vector is then concatenated with the state vector to form the joint feature.

[0127] Step S503: Input the joint features into the second policy network in the hierarchical decision model so that the second policy network outputs the initial probability distribution of each generator set in the target synchronized generator group.

[0128] It should be noted that the system inputs the joint features into the second policy network (i.e., the lower-level conditional policy network). This network adopts a conditional input mechanism, with the hidden layer capacity expanded to a three-layer fully connected structure, with 256, 128, and 64 nodes respectively. Each layer is configured with the ReLU activation function. The output layer generates an initial action probability matrix, where each row corresponds to a generator unit, and the two elements represent the initial probabilities of the unit being disconnected and not disconnected, respectively.

[0129] Step S504: Input the initial probability distribution of the machine switching into the mask matrix in the hierarchical decision model and perform element-wise multiplication and renormalization to determine the machine switching control command based on the probability distribution after renormalization.

[0130] It should be noted that the mask matrix is ​​constructed as follows: for the i-th generator set, if the generator set belongs to the target coherent generator group, its corresponding mask vector element is 1; if the generator set does not belong to the target coherent generator group, its corresponding mask vector element is 0. The initial generator tripping probability matrix is ​​multiplied element-wise by the mask matrix (Hadamard product operation) to force the tripping probability of generator sets corresponding to non-target generator groups to zero, while fully preserving the relative proportions of the action probabilities of each generator set within the target generator group. Subsequently, the tripping probabilities of each generator set within the target generator group are renormalized so that their sum is restored to 1, resulting in the final generator tripping action probability distribution. Based on this probability distribution, sampling is performed to determine the tripping state of each generator set, and tripping control commands are generated.

[0131] In this embodiment, a complete closed loop of aircraft switching decisions, from macro-level target cluster location to micro-level aircraft selection, is achieved through a hierarchical decision-making process and the rigid constraints of a masking mechanism. The first strategy network completes the target cluster location based on the homogeneous clustering results, while the second strategy network performs aircraft switching selection only within the target cluster under masking constraints. This reduces the dimension of the action space and the scope of ineffective exploration, and improves the consistency of the spatial logic of the decision and the engineering feasibility.

[0132] like Figure 7 As shown, Figure 7 Here is a flowchart of the algorithm training and network parameter update process. (1) First, the system will complete the basic configuration before training: first, complete the random initialization of the DSP simulation environment interface and the hierarchical Actor-Critic network parameters, and then build an experience replay buffer to temporarily store the complete state-action trajectory data generated by the current strategy. (2) In each training round, the intelligent experience completes the closed-loop interaction with the DSP simulation environment. The execution steps at time t are as follows: State observation: obtain the normalized state vector of the current power grid. Hierarchical decision-making: The upper-level network makes decisions based on the state. The output cluster selection probability distribution is used to obtain the target cluster index after sampling. The lower-level network will and After concatenation, the original action probabilities are output, combined with the topological mask. Actions that eliminate non-candidate aircraft groups are ultimately sampled to generate specific aircraft switching commands. Simulation execution: The machine switching command will be executed. The system is converted into a simulation control card and written to a configuration file. Then, time-domain simulation is invoked to advance the system to the next time step, thereby obtaining a new system state. Data Records: Temporary State Transition Tuples and instant rewards (3) Leveraging value networks Value prediction is performed for all states within the trajectory, and the TD error at each step is calculated recursively in reverse order. and generalized dominance estimate This process relies on a discount factor. and smoothing factor The system stability / instability results at the end of the round are backpropagated to the initial decision point, so as to effectively evaluate the contribution of early key actions to the final system stability. (4) After a batch of trajectory data is collected, small-batch stochastic gradient descent is used to update the network parameters. The specific process is as follows: Data sampling: Randomly extract sample data of batch size M from the temporary interactive buffer; Critic update: Minimize the mean square error of the state value prediction value and the actual cumulative return. Revise the network's assessment benchmark for system state security level; Actor update: based on the same batch of samples and generalized dominance value. Simultaneously calculate the PPO cutoff loss and entropy regularization term of the upper and lower layer policies, and update the parameters respectively using gradient descent. and Experience archiving and buffer reset: To improve sample utilization efficiency and support subsequent hybrid training, a long-term experience archiving mechanism is designed. After updating the parameters of the current batch, the complete temporal trajectory of the temporary buffer is appended to the global self-exploration experience pool, then the temporary buffer is cleared, and the next round of environmental interaction sampling is started. The algorithm iterates according to the above process until the preset maximum number of training rounds is reached or the convergence judgment condition is met.

[0133] This application also provides a power system tripping decision device; please refer to [reference needed]. Figure 8 The power system tripping decision device includes: The data acquisition module 10 is used to acquire the topology data and operating status data of the power grid, as well as a preset fault set; The coherence partitioning module 20 is used to partition the generator sets in the power system according to the topology data and the preset fault set to obtain a coherent generator group. Decision process construction module 30 is used to construct a two-level Markov decision process model based on the coherent machine group and the operating status data; The decision model construction module 40 is used to construct a hierarchical decision model based on the two-level Markov decision process model. The decision execution module 50 is used to input the operating status data into the hierarchical decision model so that the hierarchical decision model can determine the target coordinating machine group and the machine switching control command.

[0134] The power system tripping decision device provided in this application, employing the power system tripping decision method in the above embodiments, can solve the technical problem of how to improve the timeliness and accuracy of emergency tripping decisions. Compared with the prior art, the beneficial effects of the power system tripping decision device provided in this application are the same as those of the power system tripping decision method provided in the above embodiments, and other technical features in the power system tripping decision device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0135] In one embodiment, the coherence division module 20 is further configured to cluster the generator sets in the power system according to the topology data to obtain an initial coherent generator group; obtain the relative power angle discrete sampling sequence of each generator set within the observation window according to the preset fault set; calculate the correlation coefficient of the relative power angle discrete sampling sequence between any two generator sets in the initial coherent generator group; and divide the initial coherent generator group according to the correlation coefficient to obtain a coherent generator group.

[0136] In one embodiment, the decision-making process construction module 30 is further configured to: calculate normalized relative power angle feature values ​​based on the real-time relative power angles of each generator set and a preset stability threshold in the operating status data; construct a state space based on the normalized relative power angle feature values; decouple the action space into a first action space and a second action space based on the coordinating generator group, wherein the first action space is used to select a target coordinating generator group and the second action space is used to determine the generator set to be cut off within the target coordinating generator group; obtain the current decision step count and the total capacity of generator sets cumulatively cut off when the simulation reaches a preset total simulation duration, recorded during the iterative decision-making process according to the preset time-domain simulation state transition strategy; construct a comprehensive reward function based on the normalized relative power angle feature values, the preset stability threshold, the current decision step count, and the cumulative total capacity of generator sets cut off; and construct a two-layer Markov decision process model based on the state space, the first action space, the second action space, the comprehensive reward function, and the preset time-domain simulation state transition strategy.

[0137] In one embodiment, the decision process construction module 30 is further configured to use the coherent generator group as a first discrete action option and construct a first action space based on the first discrete action option; and to use the generator set included in the coherent generator group as a second discrete action option and construct a second action space based on the second discrete action option.

[0138] In one embodiment, the decision process construction module 30 is further configured to determine a stability reward value based on the normalized relative power angle characteristic value and the preset stability threshold; determine a time response penalty based on the current decision step number and the preset single-step penalty constant; determine a tripping control cost based on the cumulative total capacity of the units cut off and the preset tripping cost weighting coefficient; and construct a comprehensive reward function based on the stability reward value, the time response penalty, and the tripping control cost.

[0139] In one embodiment, the decision model construction module 40 is further configured to construct a value network based on the state space of the two-layer Markov decision process model; construct a first policy network based on the first action space of the two-layer Markov decision process model; construct a second policy network based on the second action space of the two-layer Markov decision process model and the target coherent machine group output by the first policy network; and construct a hierarchical decision model based on the value network, the first policy network, and the second policy network.

[0140] In one embodiment, the decision execution module 50 is further configured to input the operating status data into the first strategy network in the hierarchical decision model, so that the first strategy network outputs the selection probability distribution of each coordinating generator group, and determines the target coordinating generator group based on the selection probability distribution; concatenate the operating status data and the identification information of the target coordinating generator group to obtain joint features; input the joint features into the second strategy network in the hierarchical decision model, so that the second strategy network outputs the initial tripping probability distribution of each generator set in the target coordinating generator group; input the initial tripping probability distribution into the mask matrix in the hierarchical decision model for element-wise multiplication and renormalization processing, so as to determine the tripping control command based on the probability distribution after renormalization processing.

[0141] This application provides a power system tripping decision device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the power system tripping decision method in the above embodiment 1.

[0142] The following is for reference. Figure 9The diagram illustrates a structural schematic of a power system tripping decision device suitable for implementing embodiments of this application. The power system tripping decision device in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), tablets, PMPs (Portable Media Players), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 9 The power system tripping decision device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0143] like Figure 9 As shown, the power system tripping decision-making device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the power system tripping decision-making device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touch screens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the power system's tripping decision device to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a power system tripping decision device with various systems, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems may be implemented alternatively.

[0144] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0145] The power system tripping decision-making device provided in this application, employing the power system tripping decision-making method in the above embodiments, can solve the technical problem of how to improve the timeliness and accuracy of emergency tripping decisions. Compared with the prior art, the beneficial effects of the power system tripping decision-making device provided in this application are the same as those of the power system tripping decision-making method provided in the above embodiments, and other technical features in this power system tripping decision-making device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0146] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0147] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0148] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the power system tripping decision method in the above embodiments.

[0149] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0150] The aforementioned computer-readable storage medium may be included in the power system's tripping decision-making equipment; or it may exist independently and not be assembled into the power system's tripping decision-making equipment.

[0151] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by the power system's generator tripping decision-making device, the power system's generator tripping decision-making device: acquires the power grid's topology data, operating status data, and a preset fault set; performs coordination classification of the generator units in the power system based on the topology data and the preset fault set to obtain a coordinating generator group; constructs a two-level Markov decision process model based on the coordinating generator group and the operating status data; constructs a hierarchical decision model based on the two-level Markov decision process model; and inputs the operating status data into the hierarchical decision model so that the hierarchical decision model determines the target coordinating generator group and the generator tripping control command.

[0152] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0153] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0154] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0155] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the power system tripping decision method described above. This addresses the technical problem of improving the timeliness and accuracy of emergency tripping decisions. Compared to the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the power system tripping decision method provided in the above embodiments, and will not be elaborated upon here.

[0156] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the power system tripping decision method as described above.

[0157] The computer program product provided in this application can solve the technical problem of how to improve the timeliness and accuracy of emergency power tripping decisions. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the power system power tripping decision method provided in the above embodiments, and will not be repeated here.

[0158] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A power system tripping decision method, characterized in that, The method includes: Acquire power grid topology data, operational status data, and preset fault sets; Based on the topology data and the preset fault set, the generator sets in the power system are divided into coherent groups to obtain coherent generator groups. A two-level Markov decision process model is constructed based on the coherent machine group and the operating status data. A hierarchical decision model is constructed based on the described two-level Markov decision process model; The operational status data is input into the hierarchical decision model so that the hierarchical decision model can determine the target coherent machine group and the machine switching control command.

2. The method as described in claim 1, characterized in that, The steps of constructing a two-level Markov decision process model based on the coherent machine group and the operating status data include: Based on the real-time relative power angle and preset stability threshold of each generator set in the operating status data, the normalized relative power angle feature value is calculated, and the state space is constructed based on the normalized relative power angle feature value. The action space is decoupled into a first action space and a second action space according to the coordinating machine group. The first action space is used to select the target coordinating machine group, and the second action space is used to determine the machine group to be cut off within the target coordinating machine group. Obtain the current decision step count and the total unit capacity cumulatively removed when the simulation reaches the preset total simulation duration during the iterative decision-making process based on the preset time-domain simulation state transition strategy; A comprehensive reward function is constructed based on the normalized relative power angle feature value, the preset stability threshold, the current decision step number, and the cumulative total capacity of the units removed. A two-layer Markov decision process model is constructed based on the state space, the first action space, the second action space, the comprehensive reward function, and the preset time-domain simulation state transition strategy.

3. The method as described in claim 2, characterized in that, The step of decoupling the motion space into a first motion space and a second motion space based on the coherent machine group includes: The coherent machine group is used as the first discrete action option, and a first action space is constructed based on the first discrete action option; The generator sets included within the synchronized generator group are used as the second discrete action option, and a second action space is constructed using the second discrete action option.

4. The method as described in claim 2, characterized in that, The step of constructing a comprehensive reward function based on the normalized relative power angle feature value, the preset stability threshold, the current decision step number, and the cumulative total capacity of the units removed includes: The stability reward value is determined based on the normalized relative work angle characteristic value and the preset stability threshold. The time response penalty is determined based on the current decision step number and the preset single-step penalty constant. The tripping control cost is determined based on the cumulative total capacity of the units to be cut and the preset tripping cost weighting coefficient; A comprehensive reward function is constructed based on the stability reward value, the time response penalty, and the machine switching control cost.

5. The method as described in claim 1, characterized in that, The step of classifying the generator sets in the power system according to the topology data and the preset fault set to obtain a coherent generator group includes: Based on the topology data, the generator sets in the power system are clustered to obtain an initial synchronized generator group; The relative power angle discrete sampling sequence of each generator set within the observation window is obtained according to the preset fault set; Calculate the correlation coefficient of the discrete sampling sequence of relative power angle between any two generator sets within the initial synchronized generator group; The initial coherent machine group is divided according to the correlation coefficient to obtain the coherent machine group.

6. The method as described in claim 1, characterized in that, The steps for constructing a hierarchical decision model based on the two-level Markov decision process model include: Based on the state space of the aforementioned two-layer Markov decision process model, a value network is constructed. Based on the first action space of the two-layer Markov decision process model, construct the first policy network; Based on the second action space of the two-layer Markov decision process model and the target coherent machine group output by the first policy network, a second policy network is constructed. A hierarchical decision-making model is constructed based on the value network, the first policy network, and the second policy network.

7. The method as described in claim 1, characterized in that, The step of inputting the operating status data into the hierarchical decision model to enable the hierarchical decision model to determine the target coherent machine group and the machine switching control command includes: The operating status data is input into the first policy network in the hierarchical decision model so that the first policy network outputs the selection probability distribution of each coherent machine group, and the target coherent machine group is determined according to the selection probability distribution. The operating status data and the identification information of the target coherent machine group are concatenated to obtain a joint feature; The joint features are input into the second policy network in the hierarchical decision model so that the second policy network outputs the initial probability distribution of each generator set in the target coordinating generator group. The initial probability distribution of the machine switching is input into the mask matrix in the hierarchical decision model and multiplied element by element and renormalized to determine the machine switching control command based on the probability distribution after renormalization.

8. A power system tripping decision device, characterized in that, The device includes: The data acquisition module is used to acquire power grid topology data, operating status data, and preset fault sets; The coherence partitioning module is used to partition the generator sets in the power system according to the topology data and the preset fault set to obtain a coherent generator group. The decision process construction module is used to construct a two-layer Markov decision process model based on the coherent machine group and the operating status data. The decision model construction module is used to construct a hierarchical decision model based on the two-level Markov decision process model. The decision execution module is used to input the operating status data into the hierarchical decision model so that the hierarchical decision model can determine the target coordinating machine group and the machine switching control command.

9. A power system tripping decision device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the power system tripping decision method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the power system tripping decision method as described in any one of claims 1 to 7.

Images