State identification and emergency control method, system and device for time-space cooperation of power system, and storage medium

Abstract

The invention relates to the technical field of intelligent control of a power system, in particular to a state identification and emergency control method, system and device for time-space cooperation of the power system and a storage medium. The method comprises the following steps: collecting operation data of a power system, constructing a spatial topological code fusing branch power flow characteristics, and performing time sequence slicing processing on the operation data to extract time sequence evolution characteristics; establishing a collaborative decision network comprising a state identification path and an emergency control path, performing unified identification and future prediction on a security domain state and a stability domain state by the state identification path through fusing spatial topological coding and time sequence evolution characteristics, and automatically generating an emergency control strategy by the emergency control path when an unstable state is detected; the collaborative decision-making network is deployed in a simulation environment for multiple rounds of interaction to obtain interaction data, an experience pool containing full samples and key samples is constructed, and a state identification path and an emergency control path are optimized respectively; and outputting an operation state identification result and a corresponding emergency control strategy in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of power systems, and in particular to a power system space-time coordinated state identification and emergency control method, system, device and storage medium. BACKGROUND

[0002] With the continuous expansion of the scale of the power system and the continuous increase of the proportion of new energy access, the operation state of the power system presents the characteristics of high complexity and dynamic. The operation safety of the power system involves two core aspects of safety domain and stability domain. The safety domain focuses on whether the system operation parameters are within the allowed range, and the stability domain focuses on whether the system can maintain synchronous operation. The traditional power system state monitoring method mainly relies on manual analysis and offline calculation, which is difficult to meet the real-time and accuracy requirements of modern power systems.

[0003] In the prior art, machine learning methods such as support vector machines and recurrent neural networks are used for power system state identification, which has become a research hotspot. Some research introduces graph neural networks to process the topological relationship between nodes in the power system, and inputs the matrix composed of node features into the network for state estimation. However, these methods generally have the following shortcomings: on the one hand, the existing methods cannot effectively fuse the topological structure features between nodes and the dynamic features of node parameters changing over time, and the state identification accuracy is not high under the complex operation scenario of the power system, and the generalization ability is weak; on the other hand, the existing technology usually only realizes the functions of state identification or state prediction, and cannot automatically generate the corresponding emergency control strategy when detecting abnormal state, which makes it difficult to take effective intervention measures in time when the system appears warning or instability state.

[0004] In terms of data acquisition, the data samples of extreme fault conditions in the actual operation of the power system are relatively scarce. Most of the time, the system is in normal operation state, and the duration of the warning state and the instability state is short, resulting in a serious lack of fault samples in the training data. In order to obtain sufficient fault samples, it is usually necessary to rely on simulation platforms to generate a large amount of data, but how to effectively combine the simulation data with the actual operation scenario and improve the applicability of the model in the real environment is still a technical problem to be solved. In addition, the existing model training method mostly adopts one-time offline training, and lacks the continuous interaction and dynamic optimization mechanism with the system operation environment, so the model performance is difficult to continuously improve with the accumulation of operation data. SUMMARY

[0005] In view of the problems existing in the prior art, the present application is proposed.

[0006] Therefore, the problem to be solved by the present application is how to solve the problems of insufficient fusion of spatial topological features and time dynamic features in power system state identification, separation of state identification and emergency control decision, scarcity of fault sample data, and lack of dynamic optimization ability of the model, etc.

[0007] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the embodiments of the present application provide a power system space-time coordinated state identification and emergency control method, which comprises: collecting power system operation data and constructing a space topology code for representing the electrical connection relationship and mutual influence mode between nodes; performing time series slicing processing on the operation data to extract time series evolution features within a continuous time window for describing the dynamic change law of the system operation state; A collaborative decision-making network is established, which includes a state identification path and an emergency control path; the state identification path fuses the space topology code and the time series evolution features to uniformly identify and predict the future of the safe domain state and the stable domain state of the power system; the emergency control path automatically generates an emergency control strategy when an unstable state is detected based on the prediction results of the state identification path; The collaborative decision-making network is trained through interactive samples; the interactive samples are deployed in a simulation environment for multiple rounds of interaction to obtain interaction data, and the interaction data are used to optimize the state identification path and the emergency control path respectively; Based on the trained collaborative decision-making network, the operation state identification result of the power system and the corresponding emergency control strategy are output in real time.

[0008] As a preferred scheme of the power system space-time coordinated state identification and emergency control method of the present application, the collection of power system operation data and the construction of a space topology code comprises: identifying the direct electrical connection relationship between each node in the power system to generate an adjacency matrix, and embedding branch flow characteristics into the weight of the adjacency matrix to obtain a node representation vector that fuses the topological relationship.

[0009] As a preferred scheme of the power system space-time coordinated state identification and emergency control method of the present application, the time series slicing processing of the operation data comprises: setting a time window length, and cutting a data segment containing multiple consecutive time steps from the operation data sequence; The voltage amplitude, voltage phase angle, active power and reactive power of each node in each data segment are arranged in matrix form to generate a node feature matrix of a single time window; The node feature matrices of multiple time windows are stacked along the time dimension to construct a three-dimensional time series feature tensor.

[0010] As a preferred scheme of the state identification and emergency control method of the power system space-time coordination, wherein: the time sequence evolution feature extraction in the continuous time window comprises: slicing the continuously collected operation data by a fixed length time window, each time window containing operation parameters of multiple continuous time steps; the node operation parameters in each time window are constructed into a time sequence feature tensor, and the trend feature and periodic feature of the operation state changing with time are extracted according to the time sequence feature tensor.

[0011] As a preferred scheme of the state identification and emergency control method of the power system space-time coordination, wherein: the establishment of the cooperative decision network includes a state identification path and an emergency control path, the state identification path outputs the safe domain state and the stable domain state at the current time and the future time through the classification layer; The emergency control path outputs the emergency control strategy through the decision layer when the instability risk is identified, and the emergency control strategy includes the generator tripping control strategy, the load shedding control strategy or the energy storage scheduling control strategy.

[0012] The beneficial effects of the preferred technical scheme are: the classification layer of the state identification path outputs the dual-domain state information at the current and future time, the emergency control path automatically outputs the generator tripping, load shedding or energy storage scheduling and other control strategies through the decision layer when the instability risk is detected, and a complete closed loop from state identification to control decision is established; the dual-path architecture is functionally differentiated on the basis of sharing the feature extraction capability, shortens the response time from anomaly detection to control implementation, and provides a fast disposal means for the power system to deal with sudden faults.

[0013] As a preferred scheme of the state identification and emergency control method of the power system space-time coordination, wherein: the training of the cooperative decision network through interactive samples comprises: deploying the cooperative decision network after initial training in the power system simulation environment, and generating new sample data containing state labels and control actions through multiple simulation interactions; The new sample data is stored in an experience pool, and the experience pool includes full samples covering all operation states and key samples mainly in instability scenarios; The full samples in the experience pool are used to continuously optimize the state identification path, and the key samples in the experience pool are used to specially train the emergency control path.

[0014] The beneficial effects of the preferred technical solution are: by deploying the preliminarily trained collaborative decision network in a simulation environment and performing multiple rounds of interaction, an experience pool containing real running scene data is constructed, and the full-amount samples in the experience pool are used to continuously train the state recognition path; the emergency control path is specially trained for key samples of unstable scenes, so that the control strategy generation is more accurate and timely; through the interactive sample, a dynamic optimization channel for model performance is established, so that the network can continuously improve the prediction accuracy and control decision quality with the accumulation of running data.

[0015] As a preferred scheme of the power system space-time collaborative state recognition and emergency control method, the corresponding emergency control strategy includes triggering the emergency control path when the state recognition path outputs a warning state or an unstable state. According to the current system running condition and the predicted state evolution trend, the target area that needs to take control measures is determined. Output the generator tripping control amount, load shedding control amount or energy storage scheduling control amount for the target area to restore the power system to a safe and stable operating state.

[0016] In a second aspect, the embodiments of the present application provide a power system space-time collaborative state recognition and emergency control system, which includes a data acquisition module that acquires power system running data and constructs a spatial topology code for representing the electrical connection relationship and mutual influence mode between nodes; the running data is subjected to time series slicing processing to extract time series evolution features within a continuous time window for describing the dynamic change law of system running state; A collaborative decision module establishes a collaborative decision network including a state recognition path and an emergency control path; the state recognition path fuses the spatial topology code and the time series evolution features to uniformly recognize and predict the future of the safe domain state and the stable domain state of the power system; the emergency control path automatically generates an emergency control strategy when detecting an unstable state based on the prediction result of the state recognition path; An interactive training module trains the collaborative decision network through interactive samples; the interactive samples deploy the collaborative decision network in a simulation environment for multiple rounds of interaction to obtain interactive data, and use the interactive data to optimize the state recognition path and the emergency control path respectively; An output module outputs the running state recognition result of the power system and the corresponding emergency control strategy in real time based on the trained collaborative decision network.

[0017] In a third aspect, an embodiment of the present application provides a computer device comprising a memory and a processor, and the memory stores a computer program, wherein the computer program instructions are executed by the processor to implement the steps of the power system spatiotemporal coordination state identification and emergency control method according to the first aspect of the present application.

[0018] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program instructions are executed by the processor to implement the steps of the power system spatiotemporal coordination state identification and emergency control method according to the first aspect of the present application.

[0019] The present application has the following beneficial effects: by constructing the spatial topology coding fusing branch power flow characteristics and the dynamic evolution characteristics extracted by time sequence slicing, the synchronous capture of the electrical coupling relationship between nodes of the power system and the time evolution law of the operation parameters is realized; a collaborative decision network including a state identification path and an emergency control path is established, the state identification path realizes the unified identification and future prediction of the safe domain and stable domain state by fusing the topological space characteristics and the time sequence evolution characteristics, the emergency control path automatically generates control strategies such as generator tripping, load shedding or energy storage scheduling when detecting instability risk based on the state prediction result, and the emergency response time is shortened; by deploying the collaborative decision network in the simulation environment for multiple rounds of interaction, an experience pool including full samples and key samples is constructed, the full samples are used to continuously optimize the state identification path to improve the identification ability of various operation states, and the key samples of instability scenarios are used to specially train the emergency control path to make the control strategy generation more accurate; and the real-time and accurate identification of the operation state of the power system and the rapid and intelligent control in the emergency situation are realized as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0021] Figure 1 Flowchart of the power system spatiotemporal coordination state identification and emergency control method; Figure 2 Computer device diagram of the power system spatiotemporal coordination state identification and emergency control method; Figure 3 Specific model framework of the power system spatiotemporal coordination state identification and emergency control method; Figure 4 Mamba module schematic diagram of the power system spatiotemporal coordination state identification and emergency control method. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, a lot of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from the description, and those skilled in the art can make similar generalizations without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent of or mutually exclusive with other embodiments.

[0025] Embodiment 1 Reference Figure 1 Figure 2 For the first embodiment of the present application, the embodiment provides a power system space-time coordination state identification and emergency control method, comprising, S100: Collecting power system operation data and constructing spatial topology coding for representing the electrical connection relationship and mutual influence mode between nodes; performing time series slicing processing on the operation data to extract time series evolution features within a continuous time window for describing the dynamic change law of system operation state.

[0026] S200: Establishing a collaborative decision-making network including a state identification path and an emergency control path; the state identification path fuses the spatial topology coding and the time series evolution features to uniformly identify and predict the future of the safe domain state and the stable domain state of the power system; the emergency control path automatically generates an emergency control strategy when detecting an unstable state based on the prediction result of the state identification path.

[0027] S300: Training the collaborative decision-making network through interactive samples, deploying the collaborative decision-making network in a simulation environment for multiple rounds of interaction to obtain interactive data, and using the interactive data to optimize the state identification path and the emergency control path respectively.

[0028] S400: Based on the trained collaborative decision-making network, real-time output of the operation state identification result of the power system and the corresponding emergency control strategy.

[0029] ​It should be noted that the complexity of the power system in actual operation mainly reflects three aspects: the electrical connection relationship between nodes forms a complex topological network, the state change of the nodes will be transmitted and influenced through the network structure; the system operation parameters present dynamic evolution characteristics over time, load fluctuations, generator output adjustment and fault disturbances will cause the time sequence change of the parameters; the state of the security domain and the stability domain is coupled, the system may face multiple risks such as voltage out-of-limit and power angle instability. The traditional method only focuses on single dimension characteristics or independently carries out state evaluation, lacks the joint modeling capability of topological space correlation and time dynamic evolution, and cannot automatically generate targeted intervention measures when the state deteriorates.

[0030] Therefore, in view of the above problems, through the steps of S100-S400, a spatial topological code is constructed by fusing the node connection relationship and the branch flow characteristics, and the dynamic evolution characteristics are extracted by time window slicing of continuous operation data; a collaborative decision network is established, which includes state identification and emergency control double paths, outputs the joint extraction of topological space features and time sequence features, the unified identification of security domain and stability domain states, and the automatic generation of emergency control strategies; by deploying the network in the simulation environment for interactive training, the experience pool data is used to optimize the two paths respectively, and the learning ability of the model to the scarce fault samples and the adaptability to the complex operation scene are improved; finally, the real-time and accurate identification of the operation state of the power system and the rapid intelligent control in the emergency are realized.

[0031] Embodiment 2 With reference to Figure 2 - Figure 4 This is a second embodiment of the present application.

[0032] In this embodiment, the power system operation data is collected and the spatial topological code is constructed in step S100, which is used to represent the electrical connection relationship and mutual influence mode between nodes; the operation data is processed by time sequence slicing, and the time sequence evolution characteristics in the continuous time window are extracted, which are used to describe the dynamic change law of the system operation state, including the following steps A1-A3: A1: collecting power system operation data and constructing spatial topological code includes identifying the direct electrical connection relationship between each node in the power system to generate an adjacency matrix, embedding the branch flow characteristics into the weight of the adjacency matrix, and obtaining the node representation vector fused with the topological relationship.

[0033] Specifically, in the actual operation of the power system, first, the topological connection information of the system is obtained through the power grid dispatching automation system or the simulation platform. For the IEEE-68 node system, the system contains 68 nodes and 86 branches, and the node types include generator nodes, load nodes and intermediate transmission nodes. Based on the topological file, the direct electrical connection relationship between nodes is identified, for example, node 1 is connected to node 2 through branch 1, node 2 is connected to node 5 through branch 3, and so on. According to these connection relationships, an initial adjacency matrix is constructed, which is a two-dimensional matrix of 68 rows and 68 columns. The initial value of the elements in the matrix is set as follows: if there is a branch directly connecting node i and node j, the matrix element value is 1, otherwise it is 0. This binary adjacency matrix can only express the connectivity between nodes, and cannot reflect the strength of the connection.

[0034] To enhance the expression ability of the adjacency matrix, the branch power flow characteristics are embedded as weights, and the calculation formula is: wherein, and are the active power flow and the reactive power flow of the branch, respectively, is the system reference power.

[0035] A2: The time series slicing processing of the operation data includes setting a time window length, and cutting a data segment containing a plurality of continuous time steps from the operation data sequence; arranging the voltage amplitude, voltage phase angle, active power and reactive power of each node in each data segment into a matrix form to generate a node feature matrix of a single time window; stacking the node feature matrices of multiple time windows along the time dimension to construct a three-dimensional time series feature tensor.

[0036] Specifically, the operation data of the power system is continuously recorded by the monitoring and data acquisition system at fixed time intervals. In the simulation environment, the typical sampling period is set to 10 milliseconds, that is, 100 data points are recorded per second, and such high-frequency sampling can capture the rapid changes in the dynamic process of the power system. For a simulation lasting 10 seconds, a total of 1000 data snapshots are generated at time points, and each snapshot records the operation parameters of all 68 nodes in the system at that time. In order to convert the continuous data stream into a sample format suitable for neural network training, a fixed-length sliding time window is used for slicing processing. The length of the time window is a key parameter that needs to be balanced between capturing sufficient time information and maintaining sample independence. If the window is too short, the time information will be incomplete and it will be difficult to reflect the evolution trend of the state; if the window is too long, the calculation burden will be increased and irrelevant historical information may be included. The present application preferably sets the length of the time window to 9 time steps, corresponding to a time span of 90 milliseconds, which can cover the key stage of the system response after the fault and meet the time requirements of real-time control.

[0037] The samples are extracted from the complete data sequence using a sliding window method. The sliding step is set to 1 sampling period, that is, the window moves forward by one time point each time. For a complete data sequence containing 1000 time points, using a sliding window with a length of 9 and a step of 1, 992 independent samples can be generated. Each sample corresponds to a continuous 9-time-step data segment, the first sample contains time points 1 to 9, the second sample contains time points 2 to 10, and so on. This high-overlap slicing method can make full use of limited simulation data and greatly increase the number of training samples.

[0038] For each time window, the 9-time-step data contained is organized into a standard matrix format. At each time step, the 68 nodes of the system each have 4 operation parameters to be recorded: the voltage amplitude reflects the voltage level of the node, the voltage phase angle reflects the phase position of the node voltage, the active power reflects the active power injection or consumption of the node, and the reactive power reflects the reactive power condition of the node. Arrange these 4 parameters by column and 68 nodes by row to form a node feature matrix of 68 rows and 4 columns. The i-th row of the matrix represents the complete operation state of node i at that time, and the j-th column represents the j-th parameter of all nodes. For the 9 time steps in a single time window, 9 node feature matrices are constructed, each with a shape of 68 rows and 4 columns.

[0039] A3: Extracting time evolution features in continuous time windows includes slicing the continuously collected operation data with a fixed-length time window, each time window containing operation parameters of multiple consecutive time steps; constructing node operation parameters in each time window into a time feature tensor, and extracting trend features and periodic features of the operation state over time from the time feature tensor.

[0040] Specifically, the time sequence evolution characteristics of the power system operation state contain multiple levels of information. The most basic is the instantaneous change of the parameter, reflecting the change direction and amplitude of the current time relative to the previous time; further is the short-term trend, reflecting the change law of the parameter within several time steps, such as continuous rise, continuous decline or oscillation fluctuation; a higher level is the periodic pattern, reflecting the daily cycle, periodic change characteristics of load or power generation; the most critical is the mutation feature, reflecting the parameter jump and abnormal fluctuation caused by faults or disturbances. In order to comprehensively extract these complex time sequence characteristics, the Mamba time sequence modeling module is used as the core component.

[0041] Specifically, the Mamba module is a new type of sequence modeling architecture, which has higher computational efficiency and stronger long sequence modeling ability compared with traditional recurrent neural networks and Transformers. The core mechanism of Mamba is the selective state space model, which records historical information by maintaining a hidden state vector, and calculates the current output according to the current input and historical state. Unlike the standard state space model, Mamba introduces a selective mechanism, so that the model can dynamically adjust the state update mode according to the content of the input. Specifically, the Mamba module contains three key components: input projection layer, selective state update mechanism and output projection layer.

[0042] Specifically, the input projection layer is responsible for projecting the node features of each time step in the time sequence feature tensor to the hidden space. For an input tensor with a shape of 9x68x4, the projection layer linearly transforms the 4 features in the third dimension, expanding the feature dimension from 4 to 64. The projection uses a fully connected layer, and the weight matrix has a shape of 4x64. Through matrix multiplication, the 4-dimensional features of each node are mapped to 64-dimensional hidden features. The shape of the projected tensor becomes 9x68x64, where the 64-dimensional hidden features contain richer expression capabilities. The parameters of the projection layer are automatically learned through the backpropagation algorithm during the training process, so that the hidden features can optimally express the information useful for the task in the original features.

[0043] In an alternative embodiment, the time-evolution feature extraction can also employ a long short-term memory network (LSTM). LSTM is particularly suitable for processing long sequence data by designing a sophisticated gating unit structure to selectively remember and forget information. LSTM contains three gating units: the forget gate determines how much historical information to discard, the input gate determines how much new information to receive, and the output gate determines how much current state information to output. For the input time-series feature tensor, LSTM processes in time-step order, updating the cell state and hidden state once per time step. The cell state serves as long-term memory, maintaining information across multiple time steps; the hidden state serves as short-term memory, updating rapidly over time. The advantage of LSTM is its ability to capture long-term dependencies, such as the gradual transition of a system from a steady state to instability, which can span tens of time steps. LSTM can establish a connection between early weak signs and later instability results.

[0044] In another alternative embodiment, the time-evolution feature extraction can also employ a gated recurrent unit (GRU). GRU is a simplified version of LSTM, combining the forget gate and input gate into an update gate, and eliminating the separate cell state, only maintaining the hidden state. GRU has about 25% fewer parameters than LSTM, is faster to compute, and easier to train. In the power system state identification scenario, since the time window is relatively short (9 time steps), the degree of long-term dependence is limited, and the simplified structure of GRU is sufficient to capture key time-series features while enjoying higher computational efficiency. The choice between GRU and LSTM can be based on the specific scenario's computational resources and accuracy requirements.

[0045] In the present embodiment, the establishment of the collaborative decision-making network in step S200 includes a state identification path and an emergency control path, including the following steps B1-B2: B1: The state identification path outputs the safe domain state and the stable domain state at the current time and in the future through the classification layer.

[0046] Specifically, the overall architecture of the state identification path adopts an encoder-classifier structure mode. The encoder part is responsible for deep fusion and abstraction of the aforementioned extracted spatio-temporal features, generating high-level semantic feature representations; the classifier part is responsible for mapping high-level features to specific state categories. The encoder part includes three main modules: a graph convolution module, a time-series modeling module, and a spatio-temporal fusion module. The graph convolution module processes the spatial topology encoding generated in step A1, extracting topological association features between nodes; the time-series modeling module processes the time-series feature tensor generated in steps A2 and A3, extracting dynamic evolution features in the time dimension; the spatio-temporal fusion module integrates spatial features and temporal features, generating a comprehensive representation of spatio-temporal coupling.

[0047] Specifically, the input of the graph convolution module is a weighted adjacency matrix and a time sequence feature tensor. The shape of the weighted adjacency matrix is 68x68, and the shape of the time sequence feature tensor is 9x68x4. The calculation process of the graph convolution is independent for each time step. For the node feature matrix of each time step, the graph convolution operation is performed in combination with the adjacency matrix. The basic principle of the graph convolution is neighborhood aggregation of node features. The new features of each node are generated by aggregating its own features and the features of all neighbor nodes. In the aggregation process, the weight of the adjacency matrix determines the contribution degree of different neighbors, and the neighbor with a larger weight contributes more. In the specific calculation, first, the node feature matrix is multiplied by a learnable weight matrix to perform feature transformation; then, it is multiplied by the normalized adjacency matrix to realize neighborhood aggregation; and finally, a nonlinear activation function (such as ReLU) is used to enhance the expression ability.

[0048] The propagation rule of the graph convolution is represented as: wherein, is a normalized adjacency matrix, is a degree matrix, is the node feature of the lth layer, is a weight matrix, is an activation function.

[0049] The application adopts a two-layer graph convolution network. The first layer expands the feature dimension from 4 to 32, and the second layer expands the feature dimension from 32 to 64. The two-layer graph convolution enables each node to aggregate the information of two-hop neighbors, and the coverage is sufficient to capture the local and medium-range topological correlation.

[0050] Specifically, the time sequence modeling module adopts the Mamba module in step A3 for processing. The input of the Mamba module is the feature tensor processed by the graph convolution, and the shape is 9x68x64. The Mamba module performs time sequence modeling independently for each node. For each node, the 64-dimensional features of the node at 9 time steps are input as a sequence. Through the selective state update mechanism of the Mamba, the variation law of the running parameters of the node over time is captured. The output of the Mamba module maintains the shape of 9x68x64, but the output features have fused time sequence information. The features at each time step are no longer isolated instantaneous states, but contain dynamic information accumulated over time.

[0051] Specifically, the classifier part maps the global feature vector to the state category. The operating state of the power system involves two dimensions: the security domain and the stability domain. To achieve unified recognition of the dual-domain state, the classifier is designed as a dual-output head structure. The first output head is responsible for security domain classification and contains a fully connected layer that maps 64-dimensional features to 2-dimensional logits (logits), with two dimensions corresponding to the security state and the alert state. The logits are converted to probability distribution by the softmax function, and the sum of the two probability values is 1, representing the likelihood of the system being in a safe state and an alert state. The second output head is responsible for stability domain classification and also contains a fully connected layer that maps to 2-dimensional logits, corresponding to the stable state and the unstable state, and then obtains the probability distribution through softmax. The two output heads input from the same global feature, but use independent weight parameters, The state classification adopts the cross-entropy loss function: wherein, is the number of samples, is the number of categories, is the category weight coefficient, is the true label, is the predicted probability.

[0052] This shared input and independent output design enables the model to simultaneously learn two related but not completely identical classification tasks.

[0053] Specifically, to achieve future state prediction function, the state recognition path adopts multi-time output design. The aforementioned global feature vector is mainly used for current state recognition, while for future state prediction, the features of each time step are used instead of only using the global feature. In the specific implementation, before the time dimension attention pooling of the spatio-temporal fusion module, the feature tensor of 9 time steps is retained. The node dimension pooling is performed on the feature of each time step to obtain 9 time step feature vectors, each with a shape of 64 dimensions. Then, the classification is performed on each time step feature vector to generate 9 time step state prediction results. In this way, the model not only outputs the state of the current time (the 9th time step), but also outputs the state reconstruction of the previous 8 time steps and the future prediction based on the current trend. By observing the time series change curve of the state probability, the evolution trend of the system can be judged, for example, the instability probability gradually rises from 0.1 to 0.3 and then to 0.6, indicating that the system is developing towards instability and needs early warning.

[0054] B2: When the emergency control path identifies the instability risk, the emergency control strategy is output through the decision layer, including the generator tripping control strategy, the load shedding control strategy or the energy storage scheduling control strategy.

[0055] Specifically, the emergency control path and the state identification path share the feature extraction network at the bottom layer, including the graph convolution module, the time series modeling module, and the spatio-temporal fusion module. The shared architecture design is based on two considerations: one is to reduce the number of model parameters and avoid repeated calculation; the other is to ensure that the control decision is based on the same feature understanding as the state identification, maintaining the consistency of the decision. After the spatio-temporal fusion module, the emergency control path branches out an independent decision network. The input of the decision network includes two parts: one is the global feature vector output by the spatio-temporal fusion module, which contains the comprehensive information of the current running state of the system; the other is the state probability distribution output by the state identification path, which clearly indicates the state category and instability risk level of the system.

[0056] Specifically, the decision network adopts a multi-layer perceptron structure, which realizes the nonlinear mapping from the feature space to the control action space through the stacking of fully connected layers. The first fully connected layer concatenates the 64-dimensional global feature and the 4-dimensional state probability (one dimension for each of safe, alert, stable, and unstable) into a 68-dimensional input, which is mapped to a 128-dimensional hidden feature. The second fully connected layer maps the 128-dimensional hidden feature to a 256-dimensional one, further enhancing the expression ability. ReLU activation function is added after each fully connected layer to introduce nonlinearity, and dropout layer is added to prevent overfitting. The output layer is designed differently according to the control type.

[0057] Specifically, for the energy storage scheduling control, the output layer design is different from the previous two, because the control of energy storage is not a binary cut-off / retention, but a continuous power adjustment. Assuming that there are 5 energy storage stations in the system, the output layer is a 5-dimensional vector, each dimension representing the power adjustment amount of the energy storage station. The output layer does not use sigmoid but uses tanh activation function, which maps the output value to between -1 and 1. Negative values represent energy storage charging (absorbing system power), positive values represent energy storage discharging (injecting power to the system), and zero represents no scheduling. The output value is multiplied by the rated power of the energy storage station to get the specific scheduling power value. For example, if the rated power of an energy storage station is 200 megawatts and the output value is 0.5, the scheduling strategy is to discharge 100 megawatts.

[0058] Specifically, the training of the emergency control path adopts a special loss function design. Traditional classification or regression loss cannot fully reflect the goal of emergency control, because the control decision needs to balance two contradictory goals: one is to maximize the control success rate, i.e., the system recovers to stability after control; the other is to minimize the control cost, i.e., the total amount of generators or loads cut off is as small as possible. The loss function is designed as a weighted combination of success reward and cost penalty. The success reward is a binary indicator variable, which is 1 when the control is successful and 0 when the control fails, and the negative sign in front of it means maximizing the success reward is equivalent to minimizing the loss. The cost penalty is the sum of the magnitudes of all control actions, representing the total cost of control. The weight coefficients of the two items are determined through hyperparameter search, and the typical values are success reward weight 10 and cost penalty weight 1, indicating that the importance of control success is 10 times that of reducing cost.

[0059] The emergency control loss function is represented as: wherein, is a control success indicator variable, is the i-th control action, is the number of control objects, and is a weight coefficient.

[0060] In an optional embodiment, the emergency control path can also introduce a reinforcement learning framework. The power system simulation environment is regarded as a Markov decision process, the control policy network is regarded as an agent, the state is the system operation data, the action is the control instruction, and the reward is the control effect. Through the policy gradient algorithm or Q learning algorithm, the agent learns by trial and error in the simulation environment, and gradually discovers the control strategy with high reward. However, the training process of reinforcement learning is long and unstable, requiring a large number of simulation interactions, and is suitable for offline training scenarios.

[0061] In another optional embodiment, the emergency control path can also introduce an explainability design. An attention mechanism is added to the decision network to output the contribution of each node to the control decision. The specific implementation is to calculate the attention weight of the features of the 68 nodes in the middle layer of the decision network, representing the influence degree of each node on the decision. The attention weight can be visualized as a heat map superimposed on the power system topology map, which intuitively shows the key areas that the model focuses on.

[0062] In the present embodiment, the collaborative decision network is trained by interactive samples in step S300, including the following steps C1-C3: C1: deploy the initially trained collaborative decision network in the power system simulation environment, and generate new sample data containing state labels and control actions through multiple simulation interactions.

[0063] Specifically, the initial trained collaborative decision network is deployed into the power system simulation platform, and a real-time interface between the network and the simulation environment is established. The simulation platform uses professional power system analysis software and can accurately simulate the dynamic response process of the system. During the running process of the simulation platform, the system operation data are read in real time, and the data are transmitted to the collaborative decision network. The network outputs the state identification result and the control strategy, and the simulation platform receives the control instructions and acts on the simulation system. The whole interaction process is realized in the form of software interface, and the delay of data transmission and instruction execution is controlled within milliseconds, meeting the real-time requirement.

[0064] Specifically, the workflow of interactive training includes five links: scenario generation, simulation running, network decision, effect verification, and data recording. The scenario generation link randomly generates simulation conditions according to the preset parameter range, including initial load level, generator output distribution, fault type, fault location, and fault duration. The load level is randomly selected between 70% and 110% of the rated value, the generator output is randomly distributed within the respective output range, the fault type is mainly three-phase short circuit, the fault location covers all branches, and the fault duration is selected between 0.1 seconds and 0.5 seconds. Each set of parameters corresponds to an independent running scenario, and a diversified scenario library is generated by a large number of random combinations.

[0065] Specifically, the simulation running link starts the power system dynamic simulation, and the stability determination condition of the effect verification link is: wherein, , is the maximum and minimum power angle of the system, is the frequency deviation, is the voltage per unit of node i. If all conditions are met, the control is determined to be successful; if any condition is not met, the control is determined to be failed.

[0066] The simulation system runs under the given scenario. The simulation duration is set to 10 seconds, which is sufficient to cover the complete process of fault occurrence, system response, and state evolution. During the simulation process, the system operation data are recorded at a period of 10 milliseconds, and 1000 time point data snapshots are generated for each simulation. The network decision link inputs the simulation data into the collaborative decision network in real time, and the network performs state identification and control decision based on the data of the previous 9 time steps. If the network determines that the system is in an unstable state or the instability probability exceeds the threshold, an emergency control strategy is output, including the list of generators or loads that need to be cut off.

[0067] C2: Store the new sample data in the experience pool, which includes full-quantity samples covering all operating states and key samples mainly in unstable scenarios.

[0068] Specifically, the design of the experience pool needs to meet three core requirements: first, support for quickly inserting new samples, as the interaction process continuously generates data; second, support for retrieving samples by category, by time, by importance, and other ways to meet the needs of different training strategies; third, support for sample elimination and update to avoid the storage and computing burden caused by unlimited data growth. Based on these requirements, the experience pool adopts a dual-buffer plus index architecture design.

[0069] Specifically, the dual-buffer includes a full-sample buffer and a key-sample buffer. The full-sample buffer stores all samples generated during the interaction process without any screening to ensure data integrity and diversity. Full samples are stored in chronological order, and each sample contains input data (time series feature tensor, adjacency matrix), output data (state label, control action), and meta information (scenario parameters, generation time). Full samples are used to train the state recognition path, as state recognition needs to learn the state features of all categories, including safe, alert, stable, and unstable.

[0070] Specifically, the key-sample buffer specifically stores important samples, mainly including two categories: one is unstable state samples, which are rare in actual operation but crucial to safety; the other is control success case samples, recording the control strategy and its effect. The key-sample buffer uses a priority queue data structure, and each sample is assigned a priority score, with a higher score indicating greater importance. The priority of unstable samples is automatically set to the highest, and the priority of control success samples is set according to the control cost, with a high priority for successful cases with low cost, as these cases demonstrate efficient control strategies.

[0071] The calculation rule for sample priority score is: wherein, is the control cost, is the maximum cost.

[0072] C3: Continuously optimize the state recognition path using full samples in the experience pool, and specifically train the emergency control path using key samples in the experience pool.

[0073] Specifically, although the state recognition path and the emergency control path share the underlying feature extraction network, they differ in training data, optimization objectives, and training frequency. The state recognition path uses data from the full-sample buffer, and the training goal is to minimize state classification error, with a training frequency of once per iteration. The emergency control path uses data from the key-sample buffer, and the training goal is to maximize control success rate while minimizing control cost, with a training frequency of once every time enough unstable samples (e.g., 1000) are accumulated.

[0074] Specifically, the training process of the state identification path includes five steps of data loading, forward propagation, loss calculation, back propagation and parameter updating. A batch of samples is randomly extracted from the full sample buffer in the data loading stage, and the batch size is set to 64. Since the full sample contains all state categories and is unevenly distributed, a weighted sampling method is used, and the sampling probability of the unstable sample is set to 10 times its true proportion, ensuring that there are enough unstable samples in each batch. In the forward propagation stage, the batch data is input into the collaborative decision network, and after graph convolution, time series modeling, spatio-temporal fusion and classification layer, the state prediction result is output. In the loss calculation stage, the prediction result is compared with the true label, and the cross-entropy loss is calculated. The cross-entropy loss is multiplied by the class weight coefficient, and the weight coefficient of the unstable class is 10 and the weight coefficient of the other classes is 1, so that the contribution of the unstable sample to the loss is greater.

[0075] Specifically, the training of the emergency control path uses unstable samples and control success cases in the key sample buffer. The training data is constructed in the following way: for each unstable sample, record its corresponding operation data, the control strategy output by the network, and the execution effect of the strategy (success or failure). The training goal is to let the network learn to output successful control strategies through supervised learning and avoid outputting failed control strategies. The loss function is designed as a multi-objective optimization, including three components: control success reward item, control cost penalty item and smoothness regularization item. The control success reward item is a binary variable, which is positive when successful and negative when failed. Maximizing this item in the training process is equivalent to improving the control success rate. The control cost penalty item is the total amount of cut generators or loads, and minimizing this item in the training process is equivalent to reducing the control cost. The smoothness regularization item restricts the difference between adjacent decision variables to prevent outputting control strategies with severe oscillation. The three items are combined through weight coefficients, and the weight ratio is success reward:cost penalty:smoothness regularization=10:1:0.1.

[0076] The complete control loss function is: where the first term is the success reward, the second term is the cost penalty, and the third term is the smoothness regularization item, .

[0077] In an optional implementation, model training can also introduce a curriculum learning strategy. Curriculum learning simulates the human learning process, gradually transitioning from simple tasks to complex tasks. In state identification training, only obviously unstable and obviously stable samples are used at the beginning, critical samples are introduced in the middle, and noise samples and abnormal samples are added in the later period.

[0078] In another optional implementation, the model training in step C3 can also introduce a meta-learning technique. The goal of meta-learning is to enable the model to quickly adapt to new tasks or new scenarios. Specifically, different types of instability scenarios are regarded as different tasks, and the performance is evaluated after a small amount of training on each task, and then the network parameters are optimized to maximize the performance after a small amount of training. Meta-learning enables the model to have rapid adaptation capability, when encountering a new type of instability scenario not seen in training, only a small amount of samples are needed to adjust the parameters to cope with the new scenario.

[0079] In an optional implementation, the model training in step C3 can also introduce a curriculum learning strategy. Curriculum learning simulates the human learning process, gradually transitioning from simple tasks to complex tasks. In state recognition training, only obviously unstable and obviously stable samples are used at the beginning, critical samples are introduced in the middle, and noise samples and abnormal samples are added at the end. This easy-to-difficult training sequence enables the model to first establish basic discrimination ability, and then gradually improve the processing ability of difficult samples, avoiding the training difficulty caused by directly learning complex samples at the beginning.

[0080] In another optional implementation, the model training in step C3 can also introduce a meta-learning technique. The goal of meta-learning is to enable the model to quickly adapt to new tasks or new scenarios. Specifically, different types of instability scenarios are regarded as different tasks, and the performance is evaluated after a small amount of training on each task, and then the network parameters are optimized to maximize the performance after a small amount of training. Meta-learning enables the model to have rapid adaptation capability, when encountering a new type of instability scenario not seen in training, only a small amount of samples are needed to adjust the parameters to cope with the new scenario.

[0081] In this embodiment, the corresponding emergency control strategy in step S400 includes the following steps D1-D2: D1: When the state recognition path outputs an alert state or an instability state, trigger the emergency control path.

[0082] Specifically, after the collaborative decision network is trained and deployed to the actual application environment, the system enters real-time running mode. In real-time running mode, the network continuously receives running data from the monitoring system, and performs state recognition and control decision according to a fixed period. The running period is consistent with the data sampling period, which is set to 1 second, i.e. a complete state evaluation is performed once every second. In each period, the system collects running data of the past 9 seconds (corresponding to 9 sampling points), constructs an input tensor, inputs the state recognition path of the collaborative decision network, and outputs the state probability distribution at the current time.

[0083] Specifically, the triggering mechanism is based on state probability threshold judgment. The state recognition path outputs four state probabilities: safe, alert, stable, and unstable, each probability value is between 0 and 1, and the sum of the four probabilities is 2 (the safe domain and the stable domain are each normalized). The trigger condition is designed in two levels: the first level trigger condition is that the unstable probability exceeds the high threshold, and the high threshold is set to 0.8, which means that the system is extremely likely to be in an unstable state and needs to take control measures immediately. The second level trigger condition is that the alert probability or the unstable probability exceeds the medium threshold, and the medium threshold is set to 0.6, which means that the system has a high risk and starts the warning process and prepares control measures. The first level trigger directly executes emergency control, and the second level trigger first notifies the operator and prepares control strategies, and if the state further deteriorates in the next cycle, the control is automatically executed.

[0084] The trigger condition determination rule is: wherein, is the unstable probability, is the alert probability.

[0085] D2: According to the current system operating condition and the predicted state evolution trend, determine the target area that needs to take control measures, Specifically, the core task of determining the control target area is to identify the source and propagation path of instability, and to accurately act on the key area with control measures to avoid excessive intervention caused by blind cutting. The determination of the target area is based on two types of information: one is the node-level operation state analysis to identify abnormal nodes; the other is the network-level structure analysis to identify the key channels of instability propagation.

[0086] Specifically, further filter the candidate nodes according to the deviation degree of the node operating parameters. For generator nodes, calculate the deviation of its power angle relative to the average power angle of the system, and the generator whose deviation exceeds the threshold (e.g. 50 degrees) is the key object of instability. For load nodes, calculate the degree of deviation of its voltage from the rated value, and the load node with serious deviation may be the source of voltage instability. Nodes with high attention score and large parameter deviation are selected as the final target nodes, and the calculation formula of the node attention score is: wherein, the first term is the total attention of node i to its neighbors, and the second term is the total attention of the neighbors to node i, is the neighbor set of node i.

[0087] These nodes are both in key positions in the network structure and show abnormal in the operating state, and are the optimal control objects.

[0088] Specifically, the network-level analysis identifies the key branches for the propagation of instability. Instability propagates through branches in the power system, and branches with larger power flow propagate faster. By analyzing the time-series change of branch power flow, branches with rapidly increasing or oscillating power flow are identified. For the nodes connected to these key branches, even if the abnormality of a single node is not obvious, since it is on the instability propagation path, it should also be considered for control. The results of node-level analysis and network-level analysis are integrated to form a complete definition of the target area, which includes the instability source node and the propagation path node.

[0089] Specifically, after the target area is determined, the control type needs to be further decided. For instability dominated by generator nodes, generator tripping control is preferred; for instability dominated by load nodes, load shedding control is preferred; for nodes on the propagation path, the instability propagation can be blocked by cutting off the branch (disconnecting the line). The selection of control type is based on the judgment of instability type, angle instability corresponds to generator tripping, voltage instability corresponds to load shedding, and frequency instability corresponds to generator tripping or load shedding. The state identification path can output the judgment of instability type (by analyzing which parameter deviates most seriously) to provide the basis for control type selection.

[0090] D3: Output the generator tripping control amount, load shedding control amount, or energy storage scheduling control amount for the target area to restore the power system to a safe and stable operating state.

[0091] Specifically, the generation of control strategy is completed by the emergency control path. The emergency control path receives the node list of the target area and the operating state characteristics of the current system, and calculates the control action of each node through the decision network. For generator tripping control, the decision network outputs the tripping probability of each generator in the target area, and the generator with a probability greater than the threshold value of 0.5 is marked as the tripping object. The output result is a list of generator numbers, such as tripping generator 3, generator 7, and generator 12. For load shedding control, the output is the tripping probability of each load node in the target area, and the tripping object is also determined according to the threshold value, and the result is a list of load node numbers. For energy storage scheduling control, the output is the power regulation amount of each energy storage station, a positive value indicates discharging, and a negative value indicates charging, and the result is a list of energy storage station numbers and corresponding power values.

[0092] In an optional implementation, the determination of the control target area can also use the sensitivity analysis method. Sensitivity analysis calculates the improvement of system stability by tripping each candidate node, and quantifies the influence of each node on stability margin by simulating the system response after tripping different nodes in the simulation environment. For each candidate node, the critical tripping time or power angle stability index of the system after tripping is quickly calculated, and the node with the most obvious improvement effect is selected as the control target.

[0093] In another alternative embodiment, the determination of the control target region can also be combined with clustering analysis of the electrical distance. The system nodes are clustered and divided according to the electrical distance, and a node group with close electrical connection is identified. Instability often occurs and spreads within a node group with close electrical distance. By analyzing the power angle swing amplitude and voltage fluctuation degree of each node group, the boundary between the instability region and the stability region is identified.

[0094] In summary, in the data acquisition and feature construction stage, an adjacency matrix is generated by identifying the direct electrical connection relationship between nodes, and the branch power flow features are embedded in the weights to construct a spatial topology code that can reflect the electrical coupling strength. The graph attention is used to adaptively learn the attention degree of the node to the neighbor, and the key topological correlation is captured. The continuous operation data is sliced by the sliding window method, the node operation parameters of each time window are organized into a standard matrix format and stacked into a three-dimensional tensor, and the integrity and causality of the time series are preserved. The selective state update of the Mamba time series modeling module enables the model to dynamically adjust the state update method according to the input content, and automatically identify the key time sequence information such as the fault time. In the collaborative decision-making network establishment stage, the graph convolution module extracts the topological correlation features between nodes through neighborhood aggregation, the time series modeling module captures the dynamic evolution law of the parameters, and the spatio-temporal fusion module deeply integrates the two types of features to generate a comprehensive representation. The dual-output head classifier design realizes the unified identification of the safety domain and the stability domain, and the multi-time output design provides the combination of current state identification and future state prediction. The emergency control path ensures consistency with state identification by sharing the underlying feature extraction network, and realizes nonlinear mapping from the feature space to the control action space through the multi-layer perception structure. Differentiated output layers are designed for different control types such as machine tripping, load shedding, and energy storage scheduling. The present application realizes the deep integration of spatial topology features and time dynamic features, establishes an integrated decision of state identification and emergency control, improves the adaptability of the model to complex operation scenarios through interactive training, and provides intelligent protection for the safe and stable operation of the power system.

[0095] Embodiment 3 The above is a schematic scheme of a power system spatio-temporal collaborative state identification and emergency control method. It should be noted that the technical scheme of the power system spatio-temporal collaborative state identification and emergency control system belongs to the same concept as the technical scheme of the power system spatio-temporal collaborative state identification and emergency control method described above. The technical scheme of the power system spatio-temporal collaborative state identification and emergency control system in this embodiment is not described in detail, and can be referred to the description of the technical scheme of the power system spatio-temporal collaborative state identification and emergency control method.

[0096] The present embodiment also provides a power system spatio-temporal collaborative state identification and emergency control system, comprising: The data collection module collects power system operation data and constructs a spatial topology code for representing the electrical connection relationship and mutual influence mode between nodes; the operation data is subjected to time sequence slicing processing to extract time sequence evolution features within a continuous time window for describing the dynamic change law of the system operation state. The collaborative decision-making module establishes a collaborative decision-making network including a state identification path and an emergency control path; the state identification path fuses the spatial topology code and the time sequence evolution features to uniformly identify and predict the future of the safe domain state and the stable domain state of the power system; the emergency control path automatically generates an emergency control strategy when an unstable state is detected based on the prediction result of the state identification path; The interactive training module trains the collaborative decision-making network through interactive samples; the interactive samples deploy the collaborative decision-making network in a simulation environment for multiple rounds of interaction to obtain interactive data, and use the interactive data to optimize the state identification path and the emergency control path respectively; The output module outputs the operation state identification result of the power system and the corresponding emergency control strategy in real time based on the trained collaborative decision-making network.

[0097] The embodiment also provides an electronic device suitable for power system space-time collaborative state identification and emergency control, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the power system space-time collaborative state identification and emergency control method proposed in the above embodiment.

[0098] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to realize the power system space-time collaborative state identification and emergency control method proposed in the above embodiment.

[0099] The storage medium proposed in the embodiment and the power system space-time collaborative state identification and emergency control method proposed in the above embodiment belong to the same inventive concept, and the technical details not described in detail in the embodiment can be referred to the above embodiment, and the embodiment and the above embodiment have the same beneficial effects.

[0100] Those skilled in the art can clearly understand the present application through the above description of the embodiments, and the present application can be realized by software and necessary general hardware, and of course, can also be realized by hardware. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH memory, a hard disk, or an optical disc, and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods of various embodiments of the present application.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. A method for state identification and emergency control of a power system in space-time coordination, characterized in that: The method comprises: collecting power system operation data and constructing a spatial topology code for representing electrical connection relationships and mutual influence modes between nodes; The operation data is subjected to time series slicing processing to extract time series evolution features within a continuous time window for describing dynamic change rules of system operation states; A collaborative decision-making network is established, which comprises a state identification path and an emergency control path; the state identification path fuses the spatial topology code and the time series evolution features to uniformly identify and predict future states of safe and stable domains of the power system; The emergency control path automatically generates an emergency control strategy when an unstable state is detected based on the prediction results of the state identification path; The collaborative decision-making network is trained through interactive samples; the interactive samples are deployed in a simulation environment to obtain interactive data through multiple rounds of interaction, and the interactive data are used to optimize the state identification path and the emergency control path respectively; Based on the trained collaborative decision-making network, operation state identification results and corresponding emergency control strategies of the power system are output in real time.

2. The power system space-time coordination state identification and emergency control method of claim 1, wherein: The method of collecting power system operation data and constructing a spatial topology code comprises: identifying direct electrical connection relationships between nodes in the power system to generate an adjacency matrix, and embedding branch flow characteristics into weights of the adjacency matrix to obtain node representation vectors fused with topology relationships.

3. The power system space-time coordination state identification and emergency control method of claim 2, wherein: The time series slicing processing of the operation data comprises: setting a time window length, and cutting data segments containing multiple continuous time steps from the operation data sequence; The voltage amplitude, voltage phase angle, active power and reactive power of each node in each data segment are arranged in matrix form to generate a node feature matrix of a single time window; Multiple time window node feature matrices are stacked along the time dimension to construct a three-dimensional time series feature tensor.

4. The power system space-time coordination state identification and emergency control method of claim 3, wherein: The extraction of time series evolution features within a continuous time window comprises: The continuously collected operation data is subjected to fixed-length time window slicing, and each time window contains operation parameters of multiple continuous time steps; The node operation parameters within each time window are constructed into a time series feature tensor, and trend features and periodic features of the operation state changing over time are extracted from the time series feature tensor.

5. The power system space-time coordination state identification and emergency control method of claim 4, wherein: The establishment of the collaborative decision-making network comprising the state identification path and the emergency control path comprises: the state identification path outputs safe and stable domain states at the current time and future time through a classification layer; The emergency control path outputs an emergency control strategy through a decision layer when an instability risk is identified, and the emergency control strategy comprises a generator tripping control strategy, a load shedding control strategy or an energy storage scheduling control strategy.

6. The power system space-time coordination state identification and emergency control method of claim 5, wherein: The training of the collaborative decision-making network through interactive samples comprises: deploying the initially trained collaborative decision-making network in a power system simulation environment, generating new sample data containing state labels and control actions through multiple simulation interactions; The new sample data are stored in an experience pool, and the experience pool comprises full-quantity samples covering all operation states and key samples mainly in unstable scenarios; The state recognition path is continuously optimized using all samples in the experience pool, and the emergency control path is specially trained using key samples in the experience pool.

7. The power system space-time coordination state identification and emergency control method of claim 6, wherein: The corresponding emergency control strategy includes triggering the emergency control path when the state recognition path outputs an alert state or an unstable state. According to the current system operating condition and the predicted state evolution trend, a target region requiring control measures is determined. A generator tripping control amount, a load shedding control amount, or an energy storage scheduling control amount is output for the target region to restore the power system to a safe and stable operating state.

8. A power system spatiotemporal coordination state identification and emergency control system based on the power system spatiotemporal coordination state identification and emergency control method of any one of claims 1-7, characterized in that: The data acquisition module acquires power system operating data and constructs a spatial topology code for representing the electrical connection relationship and mutual influence mode between nodes; the operating data is time-sliced to extract time series evolution features within a continuous time window for describing the dynamic change law of the system operating state. The collaborative decision-making module establishes a collaborative decision-making network including a state recognition path and an emergency control path; the state recognition path fuses the spatial topology code and the time series evolution features to uniformly recognize and predict the future of the safe domain state and the stable domain state of the power system. The emergency control path automatically generates an emergency control strategy when an unstable state is detected based on the prediction result of the state recognition path. The interactive training module trains the collaborative decision-making network through interactive samples; the interactive samples deploy the collaborative decision-making network in a simulation environment for multiple rounds of interaction to obtain interaction data, and use the interaction data to optimize the state recognition path and the emergency control path, respectively. The output module outputs the operating state recognition result of the power system and the corresponding emergency control strategy in real time based on the trained collaborative decision-making network. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to implement the steps of the power system space-time collaborative state recognition and emergency control method of any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the power system space-time collaborative state recognition and emergency control method of any one of claims 1-7.

Images