A power grid dispatching risk early warning method based on multi-agent cooperation

Abstract

The application discloses a power grid dispatching risk early warning method based on multi-agent cooperation, comprising the following steps: collecting power grid dispatching node load, voltage, current and switch state data and normalizing and mapping to a topological coordinate system to construct a multi-agent state vector set; performing neighborhood weighted graph propagation iteration along the power grid topology to generate a risk potential field vector; performing timestamp correction on the non-synchronous node state and mapping to a Hamilton potential energy space to calculate a risk evolution sequence; performing quantization and cross-scale fusion on the risk sequence to generate a deviation vector field; performing bidirectional mapping on the deviation vector field after frequency domain transformation and potential field constraint; inversely transforming the frequency domain result to the time space domain to generate a cooperative risk trajectory; and generating a node risk level sequence and a risk trajectory set based on a sliding window statistics. The application realizes power grid multi-node cooperative risk dynamic modeling and evolution prediction, and improves risk identification precision and space-time consistency analysis capability.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, and in particular to a power grid dispatching risk early warning method based on multi-agent collaboration. Background Technology

[0002] During power system operation, grid dispatching requires real-time monitoring and comprehensive analysis of the load, voltage, current, and switch status of a large number of distributed nodes to ensure stable system operation under complex conditions. Existing technologies often employ rule-based threshold judgment or single time series analysis methods for early warning of dispatching risks. While these methods are simple to implement, they primarily rely on local static characteristics, making it difficult to characterize the topological relationships between multiple nodes in the power grid or reflect the collaborative coupling characteristics between different dispatching nodes. Furthermore, they are prone to misjudgment or missed judgments under conditions of load fluctuations, topological changes, or superimposed multi-source disturbances.

[0003] In recent years, some studies have introduced graph models or deep learning methods to model power grid topology to enhance the information propagation capability between nodes. However, existing methods typically employ fixed adjacency relationships or a single propagation mechanism, making it difficult to distinguish the impact of different types of electrical relationships on state evolution. Furthermore, over-smoothing can easily occur during multi-round neighborhood propagation, weakening the distinctive features of key anomalous nodes. In addition, existing methods are insufficient in handling the temporal asynchrony of node states. When time offsets exist in multi-source sampling, effective drift correction and consistency alignment are difficult to achieve, thus affecting the overall accuracy of risk assessment.

[0004] Meanwhile, existing power grid risk early warning methods typically lack the ability to continuously model the joint evolution of risks over time and topological space when dealing with dynamic evolution processes. This makes it difficult to effectively characterize the diffusion path of risks within the power grid structure, and also lacks constraints and feedback mechanisms for changes in risk potential energy. Consequently, risk indicators often remain at the static scoring level, lacking the ability to sensitively respond to changes in risk trends. Therefore, how to provide a power grid dispatch risk early warning method based on multi-agent collaboration to achieve multi-node collaborative modeling, dynamic evolution analysis, and continuous characterization of risk trajectories is a technical problem that urgently needs to be solved by those skilled in the art.

[0005] Therefore, how to provide a power grid dispatch risk early warning method based on multi-agent collaboration is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a power grid dispatch risk early warning method based on multi-agent collaboration. This invention integrates graph propagation iteration, Hamiltonian potential mapping, cross-scale deviation vector field construction, and frequency-time domain bidirectional transformation mechanism to collaboratively model and dynamically analyze the state of multiple dispatch nodes in the power grid, thereby achieving continuous characterization of risk trajectories and level assessment. It has the advantages of high risk identification accuracy, strong spatiotemporal consistency, and good early warning stability.

[0007] A power grid dispatch risk early warning method based on multi-agent cooperation according to an embodiment of the present invention includes the following steps: Collect status data from each dispatch node of the power grid, normalize the collected data according to a unified timestamp and map it to the topological coordinate system, and construct a set of multi-agent state vectors composed of node voltage, load and switch status as the initial state input; Graph propagation iteration is performed on the set of state vectors of multiple agents along the power grid topology. Each node sums the states of its neighbors by weight and calculates the risk gradient. During the iteration process, local short-term fluctuations and the long-term trend of the entire network are integrated to generate a set of risk potential vectors for each node. The asynchronous node states of multi-agent systems are corrected according to the acquisition timestamp, drift compensation is performed, the corrected node states are mapped to the Hamiltonian function potential energy space, the node risk evolution sequence is iteratively calculated along the time series, and the potential energy gradient of each node risk changes with time is recorded. The risk evolution sequence is quantized along the feature dimension, and the node risk potential field gradient is fused across time and topology space to generate a deviation vector field. The risk trajectory and multi-agent node state are iteratively updated along the time-space domain. The deviation vector field is transformed into the frequency domain feature space, and a risk potential field constraint is applied to the risk sequence of the frequency domain node. Frequency domain-spatial domain iterative mapping is performed along the topological nodes, and multi-scale risk evolution information is fused across frequency domain feature dimensions. The frequency domain mapping result is inversely transformed back to the time-space domain to generate a collaborative coupling risk trajectory. The node state is iteratively updated along the multi-agent topology, and the evolution of the risk trajectory and the change of collaborative constraints of each node are recorded. Sliding time window statistics are performed on the collaborative coupling risk trajectory, and the risk level sequence and risk trajectory record set of each node are calculated based on the risk threshold and risk evolution gradient.

[0008] Optionally, the node status acquisition step includes: The load data collected by each scheduling node is arranged in chronological order, outliers are removed, normalization is performed, and node load status vectors are generated. Based on the voltage data collected by each scheduling node, the data is aligned according to the timestamp, and a normalized mapping is performed to generate a node voltage state vector. The current data collected by each scheduling node is normalized and concatenated with the node load state vector to form a node current load composite vector. The switch status data of each scheduling node are synchronously encoded into discrete status vectors according to the acquisition time, and combined with the node current load composite vector to generate the node current load switch status vector. Arrange the node voltage state vectors and node current load switch state vectors of all scheduling nodes according to their positions in the power grid topology coordinate system to form a multi-agent state vector set; During the construction of the multi-agent state vector set, the node order is determined based on the neighbor relationship of the nodes in the topology to maintain the integrity of the topology information mapping. The set of multi-agent state vectors is arranged in the order of topological node indices to form the initial state input, which serves as the input for the next graph propagation iteration.

[0009] Optionally, the graph propagation iteration steps include: Input the initial state input and the list of adjacency information of the power grid nodes into the graph propagation structure, perform the neighborhood aggregation operation in the order of node index, collect the state vectors of the direct neighbor nodes of each node and sum them according to the adjacency weight to generate the first neighborhood aggregation result; Apply a risk potential field amplification operation to the first neighborhood aggregation result, and amplify the neighborhood value of high-risk nodes according to the deviation score between node load and switching state to obtain the stage one propagation representation; The propagation representation of stage one is subjected to nonlinear decay iteration along the time series. After calculating the amplitude of state change of each node, a dynamic decay vector is generated. The neighborhood aggregation result is adjusted element by element according to the dynamic decay vector to obtain the propagation representation of stage two. The propagation representation of stage two and the propagation representation of stage one are subjected to cross-stage residual calculation to generate a cross-stage residual vector. During the neighborhood aggregation process, only stable neighboring edges that appear in both propagation paths are retained to obtain the propagation representation of stage three. The propagation representation of stage three is normalized according to the node index, and the node state vector is linearly weighted and fused with the propagation representation of stage two to generate the final set of multi-agent state vectors after iteration. During the iteration process, the weighted contribution value of each adjacent edge in the neighborhood aggregation is recorded, and edges with contribution values ​​exceeding the threshold are selected to form a propagation path record set for use in the next step of risk trajectory evolution.

[0010] Optionally, the risk trajectory deviation evolution step includes: The propagation path record set output along the power grid topology and the multi-agent state vector set is sampled according to the time series of each node state vector to generate a node trajectory sequence; Hamiltonian dynamics iteration is performed on the node trajectory sequence along the time axis to calculate the deviation evolution sequence of the state vector of each node, and the deviation direction and magnitude are recorded along the topological coordinate system; In the deviation evolution sequence, the local deviation vector is calculated based on the combined characteristics of node load, current, voltage and switch state. Spatial normalization is performed on the deviation vector, and cross-scale weighted fusion is performed along the feature dimension to generate a node dynamic deviation representation. The dynamic deviation representation of a node is calculated by neighborhood difference along the index of neighboring nodes to generate a local deviation change vector. The local deviation change vector is then superimposed element-wise with the cross-stage residual vector to obtain the corrected deviation evolution sequence. In the corrected deviation evolution sequence, abnormal amplitude detection is performed along the time series, and the state changes of nodes whose deviation exceeds the set threshold and their adjacent nodes are recorded to form a deviation anomaly set. The corrected deviation evolution sequence is arranged in the order of node index to form the final risk trajectory deviation vector set, which serves as the input for the next step of risk potential field migration and early warning calculation.

[0011] Optionally, the risk potential field shift and continuous indicator generation steps include: Along the set of deviation vectors of the final risk trajectory, a local risk potential energy field is generated for the state vector of each node, and a node potential energy matrix is ​​constructed according to the node index and topological coordinate distribution. Perform space-time migration iteration on the node potential energy matrix to map the local deviation vector to the potential energy of adjacent nodes along the topological coordinate axis, generating a potential energy transfer sequence between nodes. Energy conservation constraints are applied to the potential energy transfer sequence between nodes, and cumulative superposition operations are performed along the time series to generate a staged node potential field evolution representation. The staged node potential field evolution representation is subjected to weighted averaging and normalization along the neighborhood index to obtain the continuous risk potential field matrix; Calculate node risk indicators based on the continuous risk potential field matrix, form a set of continuous risk indicator curves along the time series, and record the risk change magnitude and evolution direction of each node; Nodes exceeding a set threshold and their associated paths are selected from the set of continuous risk indicator curves to generate a set of high-risk nodes and a set of risk migration path records, which serve as input for subsequent early warning analysis. The set of continuous risk indicator curves and the set of high-risk nodes are arranged in the order of node index to form the final output set of multi-agent risk warning states.

[0012] Optionally, the risk indicator analysis and abnormal node output steps include: For the set of continuous risk indicator curves, a sliding time window statistical analysis is performed according to the node index, and the mean and standard deviation of each node within the window are calculated. The node risk index is standardized with the window mean and standard deviation to generate a node standardized deviation score vector. Nodes exceeding a set threshold are selected based on the deviation score vector, and the node index and its associated risk migration path are recorded to form an abnormal node set and an abnormal path record set. In the set of abnormal nodes, the state vectors of abnormal nodes are arranged in the order of node index and matched with the corresponding positions of the set of continuous risk indicator curves to form a set of node risk evolution output vectors. Normalization is performed on the set of node risk evolution output vectors to map the values ​​to a fixed interval and generate a final risk state representation set for use in early warning analysis or subsequent scheduling decisions. The abnormal node set, abnormal path record set, and node risk evolution output vector set are arranged uniformly according to the node index to form a complete abnormal node and risk indicator output set.

[0013] Optionally, the multi-agent risk aggregation and state fusion step includes: The set of abnormal nodes is matched with the set of node risk evolution output vectors according to the node index to generate a node risk state mapping matrix; The risk state mapping matrix of a node is aggregated along the power grid topology coordinate system by performing neighborhood weighted aggregation, collecting the risk state vectors of each node's direct neighbors, summing them according to the adjacency weights, and generating local risk fusion results. The results of local risk fusion are superimposed element by element with the set of continuous risk indicator curves to form a set of multi-agent risk fusion vectors in stages; A sliding window smoothing process is performed on the phased multi-agent risk fusion vector set along the time series, and the mean and standard deviation of the risk vector of each node are calculated to generate the final smoothed risk state vector set. The final smoothed risk state vector set is arranged according to the node index, and the multi-agent risk warning state set and the associated abnormal path record set are output to complete the entire risk warning state fusion process.

[0014] Optionally, the steps for statistical analysis of continuous risk indicators and output of anomaly scores include: A sliding window statistical analysis is performed on the risk warning state set of multiple agents along the time series to calculate the mean, standard deviation and rate of change of the risk state vector of each node within the window. Based on the node risk status vector and window statistics, a standardized anomaly score is calculated, and the magnitude of the node risk status deviation from the mean is mapped to anomaly score values ​​to form a set of node anomaly score vectors. The set of node anomaly score vectors is sorted according to the node index, and nodes with anomaly scores exceeding a set threshold are filtered to generate a set of high-risk nodes and a set of related risk migration path records. For each node in the high-risk node set, a weighted aggregation is performed along the neighborhood index to accumulate the anomaly scores of neighboring nodes according to the adjacency weight, generating a local anomaly enhancement vector. The local anomaly enhancement vector and the node anomaly scoring vector are superimposed element by element to generate the final anomaly scoring vector set. The final anomaly score vector set is normalized to map the values ​​to a fixed interval, and the final anomaly score vector set and the corresponding high-risk node set are output for subsequent early warning analysis or decision-making.

[0015] Optionally, the further analysis and risk state smoothing steps for the abnormal nodes include: The final set of anomaly score vectors is arranged according to the node index and the multi-agent risk warning state set to generate a node anomaly mapping matrix. The anomaly scores are weighted and averaged along the topological neighborhood of the node anomaly mapping matrix to calculate the neighborhood anomaly smoothing vector for each node. The neighborhood anomaly smoothing vector and the node anomaly score vector are superimposed element by element to form a set of corrected anomaly score vectors; The set of corrected anomaly score vectors is normalized by moving average and standard deviation along the time series to generate a set of smoothed risk state vectors. For nodes in the smoothed risk state vector set that exceed a set threshold, record their node index and corresponding risk migration path to generate the final set of abnormal nodes and risk paths.

[0016] Optionally, the risk status integration and output step includes: The multi-agent risk warning state set, the final anomaly score vector set, and the final anomaly node set are arranged in a unified manner according to the node index to generate a multi-dimensional risk state matrix; Normalization mapping is performed on the multidimensional risk state matrix along the node index and time series, mapping the values ​​to a fixed interval to form a standardized risk state matrix; A set of continuous risk curves is generated based on a standardized risk state matrix, recording the risk evolution trend and risk peak point of each node; The maximum and average values ​​of the continuous risk curve set are extracted along the time series to generate a set of risk fluctuation statistical vectors. The standardized risk state matrix, the set of continuous risk curves, and the set of risk fluctuation statistical vectors are used as the final multi-agent risk early warning output set to complete the closed loop of the entire power grid dispatch risk early warning process.

[0017] The beneficial effects of this invention are: (1) This invention constructs a set of state vectors of multiple agents and introduces a graph propagation iteration mechanism to perform topological collaborative modeling of the load, voltage, current and switch status information of the power grid dispatching nodes, and integrates local fluctuations and global trends to form a risk potential field expression, so that the coupling relationship between multiple nodes can be effectively characterized, and the expressive ability and modeling integrity of the risk propagation process in complex power grid environment can be improved.

[0018] (2) This invention introduces Hamiltonian potential energy space mapping and asynchronous timestamp correction mechanism to perform consistent alignment of multi-source asynchronous sampling data of the power grid, and combines cross-scale deviation vector field construction and frequency domain-time domain bidirectional transformation process to realize continuous characterization and multi-scale fusion analysis of risk evolution process, enhance the ability to capture dynamic changes and mutation processes of risk, and improve the stability and accuracy of risk evolution analysis.

[0019] (3) By constructing a collaboratively coupled risk trajectory and sliding window statistical mechanism, this invention dynamically evaluates and continuously updates the risk level sequence and trajectory set of power grid nodes, realizes joint tracking and hierarchical output of risks in the topological space and time dimension, and improves the real-time performance, interpretability and overall reliability of power grid dispatch risk warning. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a power grid dispatch risk early warning method based on multi-agent cooperation proposed in this invention; Figure 2 This is a schematic diagram of graph propagation and risk potential field generation for a power grid dispatch risk early warning method based on multi-agent cooperation proposed in this invention. Figure 3 This is a schematic diagram illustrating the risk evolution and frequency domain mapping of a power grid dispatch risk early warning method based on multi-agent collaboration proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-3 A power grid dispatch risk early warning method based on multi-agent cooperation includes the following steps: Collect status data from each dispatch node of the power grid, normalize the collected data according to a unified timestamp and map it to the topological coordinate system, and construct a set of multi-agent state vectors composed of node voltage, load and switch status as the initial state input; Graph propagation iteration is performed on the set of state vectors of multiple agents along the power grid topology. Each node sums the states of its neighbors by weight and calculates the risk gradient. During the iteration process, local short-term fluctuations and the long-term trend of the entire network are integrated to generate a set of risk potential vectors for each node. The asynchronous node states of multi-agent systems are corrected according to the acquisition timestamp, drift compensation is performed, the corrected node states are mapped to the Hamiltonian function potential energy space, the node risk evolution sequence is iteratively calculated along the time series, and the potential energy gradient of each node risk changes with time is recorded. The risk evolution sequence is quantized along the feature dimension, and the node risk potential field gradient is fused across time and topology space to generate a deviation vector field. The risk trajectory and multi-agent node state are iteratively updated along the time-space domain. The deviation vector field is transformed into the frequency domain feature space, and a risk potential field constraint is applied to the risk sequence of the frequency domain node. Frequency domain-spatial domain iterative mapping is performed along the topological nodes, and multi-scale risk evolution information is fused across frequency domain feature dimensions. The frequency domain mapping result is inversely transformed back to the time-space domain to generate a collaborative coupling risk trajectory. The node state is iteratively updated along the multi-agent topology, and the evolution of the risk trajectory and the change of collaborative constraints of each node are recorded. Sliding time window statistics are performed on the collaborative coupling risk trajectory, and the risk level sequence and risk trajectory record set of each node are calculated based on the risk threshold and risk evolution gradient.

[0023] In this embodiment, the node status acquisition step includes: Collect voltage, current, load and switch status data of each dispatch node in the power grid, and acquire data using smart sensors, SCADA system and remote measurement and control terminal; The data acquisition process is carried out according to a uniform sampling period to ensure that the data of all nodes are time-aligned, and the acquisition end is attached with timestamps, spatial coordinates and node identification information. The acquired data undergoes noise reduction processing, including filtering to handle voltage spikes and current noise, as well as outlier detection and interpolation to handle missing data points; After preprocessing, the data of each node is normalized and mapped to the power grid topology coordinate system to form the state vector of each node. The state vector consists of voltage amplitude, load value, and switch state code. The set of state vectors of all nodes constitutes the multi-agent state vector set, which provides the initial input for subsequent graph propagation iteration.

[0024] In this embodiment, the graph propagation iteration step includes: Multi-agent graph propagation is performed along the power grid topology, and each node performs a weighted summation based on the state vectors of its neighbors to form a local risk gradient. During the propagation process, each node calculates the average load deviation and voltage fluctuation in its neighborhood, integrates them with the historical trend of the entire network, and updates the node risk potential vector using an iterative formula; The weighting coefficients are set based on node connectivity, load sensitivity, and topology importance to ensure that the risks of critical nodes are amplified, while the states of low-risk nodes remain stable. The propagation iteration is executed in stages. The initial stage focuses on enhancing the characteristics of abnormal nodes. In the middle stage, the risk transmission is smoothed through the decay factor. In the later stage, a retention mechanism is used to record the stable propagation path, forming a set of staged risk potential field representations. After each iteration, the node state is compared with the neighbor state for gain calculation, local deviation is calculated and normalization constraints are applied to ensure that the risk potential field is numerically stable and can be compared across nodes. The risk potential field vector set of each node is output and the adjacent path record set is generated to provide input data for the risk evolution sequence.

[0025] In this embodiment, the risk trajectory deviation evolution step includes: Perform timestamp correction and drift compensation on asynchronous node status to correct errors caused by acquisition delay and data jitter. The corrected node states are mapped to the Hamiltonian function potential space, and the risk evolution sequence of each node is iteratively calculated along the time series, recording the changes in node potential gradient. The evolutionary sequence is quantified along the feature dimension to divide it into high, medium and low risk intervals, and then fused across scales in the time-space domain to form a bias vector field. During the iteration process, the risk vector of each node is weighted and fused with the vectors of its neighboring nodes, and the deviation is corrected by combining the global risk trend to form a continuous evolution trajectory.

[0026] In this embodiment, the steps of risk potential field migration and continuous index generation include: The deviation vector field is transformed into the frequency domain feature space, and a risk potential field constraint is applied to the risk sequence of frequency domain nodes to enhance the risk characteristics of key frequency bands. Frequency-space iterative mapping is performed along the topology nodes to fuse frequency domain features with the original time-space domain risk information, thereby achieving the integration of risk evolution information across multiple scales and frequency bands. In the process of cross-frequency domain feature dimension fusion, convolutional coding and neighborhood attention mechanisms are used to enhance the spatiotemporal coupling response of abnormal nodes; The frequency domain mapping result is inversely transformed back to the time-space domain to generate a collaboratively coupled risk trajectory. At the same time, the node state is updated iteratively along the topology, and the risk trajectory evolution and collaborative constraint change information of each node are recorded, providing a data foundation for the statistical analysis of continuous risk indicators.

[0027] In this embodiment, the risk indicator analysis and abnormal node output steps include: Sliding time window statistics are performed on the collaborative coupling risk trajectory to calculate the risk mean, variance and evolution gradient of each node. Based on the set risk threshold and risk gradient changes, a node risk level sequence is generated, and nodes with risk values ​​exceeding the limit are marked as abnormal nodes. The frequency of abnormal nodes and the amount of risk accumulation are statistically analyzed to form a set of abnormal node records. At the same time, the risk trajectories of the nodes are classified and sorted to provide basic data for risk aggregation of multiple agents. In this embodiment, the multi-agent risk aggregation and state fusion steps include: The risk level sequence, risk trajectory, and abnormal node set of each node are summarized to construct a multi-agent risk state matrix for the entire network. The matrix contains information on risk associations, topological dependencies, and evolutionary trends among nodes. Weighted fusion is performed on the matrix data, with the fusion weights determined by the importance of nodes and the risk gradient, highlighting key nodes and potential abnormal paths. After fusion, a set of risk state vectors for the entire network is obtained, while recording information on the collaborative changes of multiple agents, providing input for continuous risk indicator statistical analysis and anomaly scoring output; In this embodiment, the steps of continuous risk indicator statistical analysis and anomaly scoring output include: Sliding window statistics are performed on the entire network risk state vector set to calculate continuous risk indicators, including node risk mean, variance, maximum value, minimum value and evolution rate; By combining the risk level sequence and the set of abnormal node records, an anomaly scoring vector is generated to quantify the node risk level and potential threat level. The scoring vector can be used for trend analysis, risk ranking, and selection of priority processing nodes. Simultaneously, continuous risk indicators and anomaly scores are recorded in the database, enabling historical risk data tracking and subsequent trend modeling. The output includes a sequence of node risk levels, a set of abnormal nodes, a set of risk trajectory records, and a vector of continuous risk indicators, providing visual analysis and quantitative reference for scheduling decisions.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a risk monitoring scenario involving the coordinated operation of multiple nodes in a power grid dispatching system. This scenario involves the coordinated operation of various types of nodes, including generation units, substation nodes, transmission nodes, and load nodes, and aims to improve the power grid's risk identification and early warning capabilities under load fluctuations and topology disturbances. In actual operation, the voltage, current, load, and switch status of each node exhibit dynamic fluctuation characteristics due to dispatching commands and changes in electricity consumption. Different nodes form a strong coupling relationship through the power grid topology. When a disturbance occurs at a local node, the risk easily spreads outward along the topology. Traditional methods based on single-node thresholds are insufficient to characterize this type of cross-node propagation process, resulting in incomplete risk identification and unobservable propagation paths.

[0029] In this application, voltage, current, load, and switch status data of each scheduling node are collected and processed for time synchronization and normalization to ensure comparability of data from different nodes at a unified time scale. Simultaneously, the node operating status is mapped to a set of multi-agent state vectors based on the power grid topology to characterize the comprehensive operating status of each node. Subsequently, neighborhood-weighted propagation calculations are performed based on the power grid topology. Each node obtains state information from its neighboring nodes and performs weighted fusion, incorporating local fluctuations and global trend information to form a spatially propagated risk representation, enabling the node state to reflect its correlation and influence within the topological network. Furthermore, a joint temporal and spatial analysis of the risk evolution sequence is conducted, and a collaborative risk trajectory is generated by combining frequency domain transformation and potential field constraints, achieving multi-scale risk evolution modeling.

[0030] To evaluate the technical effectiveness of this invention, comparative experiments were conducted under various typical operating conditions, including load surges, load declines, and extreme fluctuation scenarios. Under the same data and operating conditions, traditional methods and the method of this invention were compared and analyzed. Indicators such as risk identification accuracy, response speed, and observability of propagation paths were comprehensively evaluated. The results show that this invention can more accurately identify risk nodes and effectively characterize the propagation path of risks in the power grid topology, thereby improving the overall dispatch risk early warning capability and system stability.

[0031] The following are quantitative evaluation results from multiple experiments under different power grid dispatching and operation conditions, verifying that the present invention has good stability and applicability under various typical risk propagation and load fluctuation conditions.

[0032] Table 1: Performance Comparison Results of Different Methods in Power Grid Multi-Node Risk Early Warning Scenarios

[0033] As can be seen from the results in Table 1, under different operating conditions, the traditional method based on static thresholds has an overall risk identification accuracy of less than 80%, and the false alarm rate and false negative rate both increase significantly under load surge and extreme fluctuation scenarios. This indicates that its adaptability to multi-node coupled disturbances is weak, and it cannot effectively characterize the propagation path of risks in the power grid topology. Therefore, the observability of the propagation path is "none".

[0034] While the comparative method introduces certain graph structure modeling capabilities and improves the accuracy of risk identification and response time compared to traditional methods, it still suffers from insufficient ability to express propagation paths. It can only achieve weak observability analysis and cannot clearly reflect the continuous diffusion relationship of risks between nodes.

[0035] The method of this invention exhibits superior performance under various operating conditions, with a risk identification accuracy consistently maintained above 93%, and significantly reduced false alarm and false negative rates. Furthermore, the average response time is significantly better than comparative and traditional methods, demonstrating that this invention effectively improves the real-time performance and stability of risk identification in multi-node collaborative risk propagation modeling. In addition, this invention can output the complete risk propagation path, making the propagation process highly observable and interpretable, thereby achieving an accurate characterization of the risk source node and its diffusion chain.

[0036] The above comparison results further demonstrate that this invention, through multi-agent collaborative modeling, neighborhood weighted propagation, and risk potential field evolution mechanism, achieves unified expression and dynamic tracking of power grid dispatch risk in the spatiotemporal dimension. Compared with existing technologies, it has stronger risk identification and topology propagation characterization capabilities, and can effectively improve the safety and stability of the power grid dispatch system under complex disturbance conditions.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A power grid dispatch risk early warning method based on multi-agent cooperation, characterized in that, Includes the following steps: Collect status data from each dispatch node of the power grid, normalize the collected data according to a unified timestamp and map it to the topological coordinate system, and construct a set of multi-agent state vectors composed of node voltage, load and switch status as the initial state input; Graph propagation iteration is performed on the set of state vectors of multiple agents along the power grid topology. Each node sums the states of its neighbors by weight and calculates the risk gradient. During the iteration process, local short-term fluctuations and the long-term trend of the entire network are integrated to generate a set of risk potential vectors for each node. The asynchronous node states of multi-agent systems are corrected according to the acquisition timestamp, drift compensation is performed, the corrected node states are mapped to the Hamiltonian function potential energy space, the node risk evolution sequence is iteratively calculated along the time series, and the potential energy gradient of each node risk changes with time is recorded. The risk evolution sequence is quantized along the feature dimension, and the node risk potential field gradient is fused across time and topology space to generate a deviation vector field. The risk trajectory and multi-agent node state are iteratively updated along the time-space domain. The deviation vector field is transformed into the frequency domain feature space, and a risk potential field constraint is applied to the risk sequence of the frequency domain node. Frequency domain-spatial domain iterative mapping is performed along the topological nodes, and multi-scale risk evolution information is fused across frequency domain feature dimensions. The frequency domain mapping result is inversely transformed back to the time-space domain to generate a collaborative coupling risk trajectory. The node state is iteratively updated along the multi-agent topology, and the evolution of the risk trajectory and the change of collaborative constraints of each node are recorded. Sliding time window statistics are performed on the collaborative coupling risk trajectory, and the risk level sequence and risk trajectory record set of each node are calculated based on the risk threshold and risk evolution gradient.

2. The power grid dispatch risk early warning method based on multi-agent cooperation according to claim 1, characterized in that, The node status acquisition steps include: The load data collected by each scheduling node is arranged in chronological order, outliers are removed, normalization is performed, and node load status vectors are generated. Based on the voltage data collected by each scheduling node, the data is aligned according to the timestamp, and a normalized mapping is performed to generate a node voltage state vector. The current data collected by each scheduling node is normalized and concatenated with the node load state vector to form a node current load composite vector. The switch status data of each scheduling node are synchronously encoded into discrete status vectors according to the acquisition time, and combined with the node current load composite vector to generate the node current load switch status vector. Arrange the node voltage state vectors and node current load switch state vectors of all scheduling nodes according to their positions in the power grid topology coordinate system to form a multi-agent state vector set; During the construction of the multi-agent state vector set, the node order is determined based on the neighbor relationship of the nodes in the topology to maintain the integrity of the topology information mapping. The set of multi-agent state vectors is arranged in the order of topological node indices to form the initial state input, which serves as the input for the next graph propagation iteration.

3. The power grid dispatch risk early warning method based on multi-agent cooperation according to claim 2, characterized in that, The graph propagation iteration steps include: Input the initial state input and the list of adjacency information of the power grid nodes into the graph propagation structure, perform the neighborhood aggregation operation in the order of node index, collect the state vectors of the direct neighbor nodes of each node and sum them according to the adjacency weight to generate the first neighborhood aggregation result; Apply a risk potential field amplification operation to the first neighborhood aggregation result, and amplify the neighborhood value of high-risk nodes according to the deviation score between node load and switching state to obtain the stage one propagation representation; The propagation representation of stage one is subjected to nonlinear decay iteration along the time series. After calculating the amplitude of state change of each node, a dynamic decay vector is generated. The neighborhood aggregation result is adjusted element by element according to the dynamic decay vector to obtain the propagation representation of stage two. The propagation representation of stage two and the propagation representation of stage one are subjected to cross-stage residual calculation to generate a cross-stage residual vector. During the neighborhood aggregation process, only stable neighboring edges that appear in both propagation paths are retained to obtain the propagation representation of stage three. The propagation representation of stage three is normalized according to the node index, and the node state vector is linearly weighted and fused with the propagation representation of stage two to generate the final set of multi-agent state vectors after iteration. During the iteration process, the weighted contribution value of each adjacent edge in the neighborhood aggregation is recorded, and edges with contribution values ​​exceeding the threshold are selected to form a propagation path record set for use in the next step of risk trajectory evolution.

4. The power grid dispatch risk early warning method based on multi-agent cooperation according to claim 3, characterized in that, The risk trajectory deviation evolution steps include: The propagation path record set output along the power grid topology and the multi-agent state vector set is sampled according to the time series of each node state vector to generate a node trajectory sequence; Hamiltonian dynamics iteration is performed on the node trajectory sequence along the time axis to calculate the deviation evolution sequence of the state vector of each node, and the deviation direction and magnitude are recorded along the topological coordinate system; In the deviation evolution sequence, the local deviation vector is calculated based on the combined characteristics of node load, current, voltage and switch state. Spatial normalization is performed on the deviation vector, and cross-scale weighted fusion is performed along the feature dimension to generate a node dynamic deviation representation. The dynamic deviation representation of a node is calculated by neighborhood difference along the index of neighboring nodes to generate a local deviation change vector. The local deviation change vector is then superimposed element-wise with the cross-stage residual vector to obtain the corrected deviation evolution sequence. In the corrected deviation evolution sequence, abnormal amplitude detection is performed along the time series, and the state changes of nodes whose deviation exceeds the set threshold and their adjacent nodes are recorded to form a deviation anomaly set. The corrected deviation evolution sequence is arranged in the order of node index to form the final risk trajectory deviation vector set, which serves as the input for the next step of risk potential field migration and early warning calculation.

5. A power grid dispatch risk early warning method based on multi-agent cooperation according to claim 4, characterized in that, The steps for risk potential field migration and continuous indicator generation include: Along the set of deviation vectors of the final risk trajectory, a local risk potential energy field is generated for the state vector of each node, and a node potential energy matrix is ​​constructed according to the node index and topological coordinate distribution. Perform space-time migration iteration on the node potential energy matrix to map the local deviation vector to the potential energy of adjacent nodes along the topological coordinate axis, generating a potential energy transfer sequence between nodes. Energy conservation constraints are applied to the potential energy transfer sequence between nodes, and cumulative superposition operations are performed along the time series to generate a staged node potential field evolution representation. The staged node potential field evolution representation is subjected to weighted averaging and normalization along the neighborhood index to obtain the continuous risk potential field matrix; Calculate node risk indicators based on the continuous risk potential field matrix, form a set of continuous risk indicator curves along the time series, and record the risk change magnitude and evolution direction of each node; Nodes exceeding a set threshold and their associated paths are selected from the set of continuous risk indicator curves to generate a set of high-risk nodes and a set of risk migration path records, which serve as input for subsequent early warning analysis. The set of continuous risk indicator curves and the set of high-risk nodes are arranged in the order of node index to form the final output set of multi-agent risk warning states.

6. The power grid dispatch risk early warning method based on multi-agent cooperation according to claim 5, characterized in that, The risk indicator analysis and abnormal node output steps include: For the set of continuous risk indicator curves, a sliding time window statistical analysis is performed according to the node index, and the mean and standard deviation of each node within the window are calculated. The node risk index is standardized with the window mean and standard deviation to generate a node standardized deviation score vector. Nodes exceeding a set threshold are selected based on the deviation score vector, and the node index and its associated risk migration path are recorded to form an abnormal node set and an abnormal path record set. In the set of abnormal nodes, the state vectors of abnormal nodes are arranged in the order of node index and matched with the corresponding positions of the set of continuous risk indicator curves to form a set of node risk evolution output vectors. Normalization is performed on the set of node risk evolution output vectors to map the values ​​to a fixed interval and generate a final risk state representation set for use in early warning analysis or subsequent scheduling decisions. The abnormal node set, abnormal path record set, and node risk evolution output vector set are arranged uniformly according to the node index to form a complete abnormal node and risk indicator output set.

7. A power grid dispatch risk early warning method based on multi-agent cooperation as described in claim 6, characterized in that, The multi-agent risk aggregation and state fusion steps include: The set of abnormal nodes is matched with the set of node risk evolution output vectors according to the node index to generate a node risk state mapping matrix; The risk state mapping matrix of a node is aggregated along the power grid topology coordinate system by performing neighborhood weighted aggregation, collecting the risk state vectors of each node's direct neighbors, summing them according to the adjacency weights, and generating local risk fusion results. The results of local risk fusion are superimposed element by element with the set of continuous risk indicator curves to form a set of multi-agent risk fusion vectors in stages; A sliding window smoothing process is performed on the phased multi-agent risk fusion vector set along the time series, and the mean and standard deviation of the risk vector of each node are calculated to generate the final smoothed risk state vector set. The final smoothed risk state vector set is arranged according to the node index, and the multi-agent risk warning state set and the associated abnormal path record set are output to complete the entire risk warning state fusion process.

8. A power grid dispatch risk early warning method based on multi-agent cooperation according to claim 7, characterized in that, The steps for statistical analysis and anomaly scoring output of continuous risk indicators include: A sliding window statistical analysis is performed on the risk warning state set of multiple agents along the time series to calculate the mean, standard deviation and rate of change of the risk state vector of each node within the window. Based on the node risk status vector and window statistics, a standardized anomaly score is calculated, and the magnitude of the node risk status deviation from the mean is mapped to anomaly score values ​​to form a set of node anomaly score vectors. The set of node anomaly score vectors is sorted according to the node index, and nodes with anomaly scores exceeding a set threshold are filtered to generate a set of high-risk nodes and a set of related risk migration path records. For each node in the high-risk node set, a weighted aggregation is performed along the neighborhood index to accumulate the anomaly scores of neighboring nodes according to the adjacency weight, generating a local anomaly enhancement vector. The local anomaly enhancement vector and the node anomaly scoring vector are superimposed element by element to generate the final anomaly scoring vector set. The final anomaly score vector set is normalized to map the values ​​to a fixed interval, and the final anomaly score vector set and the corresponding high-risk node set are output for subsequent early warning analysis or decision-making.

9. A power grid dispatch risk early warning method based on multi-agent cooperation as described in claim 8, characterized in that, The further analysis and risk status smoothing steps for the abnormal nodes include: The final set of anomaly score vectors is arranged according to the node index and the multi-agent risk warning state set to generate a node anomaly mapping matrix. The anomaly scores are weighted and averaged along the topological neighborhood of the node anomaly mapping matrix to calculate the neighborhood anomaly smoothing vector for each node. The neighborhood anomaly smoothing vector and the node anomaly score vector are superimposed element by element to form a set of corrected anomaly score vectors; The set of corrected anomaly score vectors is normalized by moving average and standard deviation along the time series to generate a set of smoothed risk state vectors. For nodes in the smoothed risk state vector set that exceed a set threshold, record their node index and corresponding risk migration path to generate the final set of abnormal nodes and risk paths.

10. A power grid dispatch risk early warning method based on multi-agent cooperation as described in claim 9, characterized in that, The risk status integration and output steps include: The multi-agent risk warning state set, the final anomaly score vector set, and the final anomaly node set are arranged in a unified manner according to the node index to generate a multi-dimensional risk state matrix; Normalization mapping is performed on the multidimensional risk state matrix along the node index and time series, mapping the values ​​to a fixed interval to form a standardized risk state matrix; A set of continuous risk curves is generated based on a standardized risk state matrix, recording the risk evolution trend and risk peak point of each node; The maximum and average values ​​of the continuous risk curve set are extracted along the time series to generate a set of risk fluctuation statistical vectors. The standardized risk state matrix, the set of continuous risk curves, and the set of risk fluctuation statistical vectors are used as the final multi-agent risk early warning output set to complete the closed loop of the entire power grid dispatch risk early warning process.

Images