Power distribution network operation risk control method based on Bayesian optimization

Abstract

The invention discloses a Bayesian optimization-based power distribution network operation risk control method. The method comprises the following steps of S1, establishing a node-side double-space coupling dynamic graph by using multi-source operation data of a power distribution network; s2, on the basis of the dynamic graph, generating a space-time embedded vector through a self-supervision task of multilevel space-time interaction between nodes and node neighborhoods; s3, constructing a risk sensitive network with multi-scale dynamic clustering, and predicting a risk propagation trajectory; s4, forming a cross-period continuous dynamic risk evolution link by using the risk propagation trajectory, and dividing and distributing a power grid risk control sub-domain; s5, evaluating sub-domain risk dynamic evolution, and constructing a joint optimization state space; s6, constructing a distributed bidirectional information cooperation link between neighborhoods; and S7, the decision convergence condition is monitored in real time, and path switching is carried out. The real-time performance and accuracy of risk control of the power distribution network are improved, and safe and stable operation of the power distribution network is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and in particular to a distribution network operation risk control method based on Bayesian optimization. Background Art

[0002] With the continuous expansion of new energy access, the continued increase in electricity load, and the increasing complexity of distribution network structures, operational risk management of distribution networks faces significant challenges. Traditional distribution network risk control methods are primarily based on expert experience and static models, assessing and controlling risk states through preset rules or fixed parameters. Typical technical approaches include fault tree analysis, risk matrix assessment, and fuzzy reasoning. However, the parameters of these methods often rely on manual expert setting and lack the ability to adaptively adjust to real-time operating conditions. Faced with complex operating environments such as rapid changes in grid load, fluctuations in distributed generation output, and sudden extreme weather events, risk assessment accuracy and real-time performance are significantly insufficient, making it difficult to respond efficiently and make quick decisions.

[0003] In recent years, with the development of artificial intelligence and machine learning technologies, some scholars have proposed using deep learning algorithms, such as recurrent neural networks and graph convolutional networks, to automatically extract temporal and spatial characteristics of distribution networks and identify and predict operational risks. However, due to their reliance on conventional supervised learning, these methods typically require large amounts of labeled data for training. In practice, labeled data is difficult to obtain in large quantities, and the training models lack generalization capabilities, significantly limiting their adaptability and generalization to new abnormal operating conditions.

[0004] Furthermore, while existing reinforcement learning algorithms can dynamically optimize decision-making strategies, they often have slow initial convergence rates, making them ineffective in addressing the rapidly evolving risks of distribution networks. In particular, in complex risk-based, multi-node, cross-regional collaborative control scenarios, traditional reinforcement learning methods fail to effectively implement information exchange and policy coordination across nodes or subdomains, making it difficult to achieve rapid and effective dynamic convergence of the overall system control strategy, presenting significant deficiencies in practical applications.

[0005] Therefore, how to provide a distribution network operation risk control method based on Bayesian optimization is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a distribution network operation risk control method based on Bayesian optimization. The present invention effectively improves the real-time and accuracy of distribution network operation risk assessment, enhances the adaptability to complex dynamic operating conditions and the risk control robustness of the overall system.

[0007] A distribution network operation risk control method based on Bayesian optimization according to an embodiment of the present invention includes the following steps:

[0008] S1. Build a node-edge dual-space coupling dynamic graph using multi-source operation data of the distribution network;

[0009] S2. Based on the dynamic graph, through the self-supervision task of multi-level spatiotemporal interaction between nodes and their neighborhoods, the associated feature weights of nodes in different operating states are dynamically adjusted node by node to generate spatiotemporal embedding vectors;

[0010] S3. Based on spatiotemporal embedding vectors, a risk-sensitive network with multi-scale dynamic clustering is constructed to automatically extract the risk diffusion path under the operating state and predict the risk propagation trajectory;

[0011] S4. Utilize risk propagation trajectories to automatically form a continuous and dynamic risk evolution link across time periods within the distribution network, and dynamically divide the grid risk control subdomains based on the spatiotemporal embedding trend of the link;

[0012] S5. For the dynamically formed risk control subdomain, evaluate the dynamic evolution of the operational risk within the subdomain, construct a joint optimization state space of risk state, decision history, and spatiotemporal characteristics, and update the risk control instruction sequence based on the evaluation results;

[0013] S6. During the dynamic execution of the risk control instruction sequence, a distributed bidirectional information collaboration link is constructed between neighboring domains to synchronize the risk status between risk control subdomains and to self-coordinate the adjustment of cross-regional risk control instructions;

[0014] S7. Monitor the dynamic convergence of decisions in each risk control subdomain. When it is predicted that the local risk convergence state has degraded, automatically start the pre-built redundant control path to achieve seamless switching of risk decision paths.

[0015] Optionally, the S1 specifically includes:

[0016] S11. Collect the flow data, voltage data, and load data of all nodes in the distribution network, calculate the data change rate and change trend of each node, and use the node data change rate to build a dynamic node feature hierarchical system;

[0017] S12. Determine the dynamic association relationship between nodes based on the node feature dynamic hierarchical system, calculate the data association strength and association direction between nodes, and form an updated node association strength matrix;

[0018] S13. Dynamically select the distribution lines directly connected to each node based on the node association strength matrix, calculate the line load level and fluctuation trend line by line, and establish a dynamic association mapping between the distribution network nodes and the corresponding lines;

[0019] S14. Based on the dynamic association mapping, a node-edge bidirectional state coupling network is constructed that simultaneously expresses the electrical connection relationship between nodes and the dynamic change trend of data, and the coupling network topology is updated;

[0020] S15, detecting a topology change of the node-edge bidirectional state coupling network, and adaptively updating a dynamic coupling weight between the node and the corresponding line according to the frequency of topology change, the number of changed nodes, and the severity of the coupling state;

[0021] S16. Convert the node-edge bidirectional state coupling network into a dynamic graph according to the updated dynamic coupling weights and topology structure.

[0022] Optionally, the S2 specifically includes:

[0023] S21. Based on the topological structure of the dynamic graph and changes in node feature data, identify adjacent nodes in the node neighborhood whose data interaction frequency is higher than a preset threshold, and determine the direction of data interaction between the nodes to form a node neighborhood interaction relationship graph;

[0024] S22. Based on the node neighborhood interaction relationship diagram, the node neighborhood is divided into direct neighborhood, secondary neighborhood, and edge neighborhood to form a multi-level structure of the node neighborhood, and the feature interaction priority of different neighborhood levels is determined;

[0025] S23, determining a characteristic sensitivity parameter based on the amplitude and frequency of changes in the characteristic data of the node itself, and screening the characteristic data in different levels of neighborhood of the node according to the characteristic sensitivity parameter;

[0026] S24, calculating the data interaction sensitivity coefficient between the node and the neighboring nodes at each level based on the neighborhood feature data, and dynamically adjusting the feature association weights of the nodes in each neighborhood level node by node based on the sensitivity coefficient;

[0027] S25. Determine the optimal feature aggregation mode based on the dynamically adjusted feature association weights, and use it to aggregate the filtered feature information of the node itself and its neighborhood.

[0028] S26. Using the feature aggregation mode, the node's own features are integrated with the multi-level neighborhood feature data to generate a spatiotemporal embedding vector for each node.

[0029] Optionally, the S3 specifically includes:

[0030] S31. Based on the spatiotemporal embedding vectors of nodes, a multi-dimensional dynamic perception network of feature interactions between nodes is constructed to identify implicit feature interaction relationships between nodes one by one.

[0031] S32, determining multi-scale cluster centers according to implicit feature interaction relationships, and generating dynamic clustering topology between nodes using cluster centers at each scale;

[0032] S33. Based on the change trend of the node embedding vector in the dynamic clustering topology, detect the boundary area where the feature changes between the node sets are abnormal, and form a risk-sensitive area with the boundary area as the core;

[0033] S34. Track the node sequence with the most drastic fluctuations in node embedding vectors within the risk-sensitive area and extract the dominant path of risk diffusion;

[0034] S35. Analyze the temporal change direction and change intensity of the node embedding vector node by node along the risk diffusion dominant path to form a continuous risk diffusion state chain;

[0035] S36. Dynamically evaluate the state transitions between nodes in the risk diffusion state chain to generate a multi-step continuous evolution model of risk diffusion;

[0036] S37. Based on the multi-step continuous evolution model of risk diffusion, the risk diffusion development direction of the risk diffusion path is deduced, and the risk propagation trajectory under the operation status of the distribution network is predicted.

[0037] Optionally, the S31 specifically includes:

[0038] S311, calculating the change rate and direction of the node feature vector based on the node spatiotemporal embedding vector, and generating a multi-dimensional feature space of the node feature change trend node by node;

[0039] S312, mapping the features of each node in the multi-dimensional feature space, and determining the spatial position relationship between each node and the feature change trend of other nodes;

[0040] S313, identifying characteristic conduction paths between nodes based on the spatial position relationship of the node feature change trend, and calibrating the conduction direction on the path;

[0041] S314, calculating the contribution coefficient of each node in the characteristic conduction path to the overall characteristic conduction strength of the path node by node, and determining the key interaction nodes whose contribution coefficient exceeds a preset threshold;

[0042] S315, monitoring the synchronization of feature change trends between key interaction nodes, calculating a synchronization index and using the index to characterize the implicit feature interaction strength between the nodes;

[0043] S316. According to the implicit feature interaction strength, the implicit feature interaction relationship between nodes is divided into three levels: high intensity, medium intensity and low intensity, forming an implicit feature interaction grading system.

[0044] Optionally, the S4 specifically includes:

[0045] S41. Based on the changing trend of the spatiotemporal embedding vectors of nodes within the risk propagation trajectory, the continuous risk transmission coupling paths between nodes across time periods are automatically extracted node by node, and the characteristic change trend vector of each node in the path is calculated;

[0046] S42. Calculate the change rate and change acceleration of the characteristic change trend vector of each node on the risk conduction coupling path node by node, and determine the node whose acceleration exceeds a preset threshold as a risk state turning point node;

[0047] S43. Divide the risk transmission coupling path into multiple independent risk propagation links according to the risk state turning points, and determine the starting node, ending node, and spatial boundary of each link.

[0048] S44, tracking the changes in the spatiotemporal embedding vectors of all nodes in each risk propagation link, establishing a spatiotemporal feature difference matrix of nodes link by link, and calculating the feature difference index within the matrix;

[0049] S45. Analyze the similarities and differences between any two adjacent risk propagation links based on the characteristic difference index to determine the correlation level between the links;

[0050] S46. Adaptively aggregate the risk propagation links based on the correlation levels between the links, and dynamically divide them into multiple risk control subdomains;

[0051] S47. Monitor the characteristic change trends and risk propagation intensity of nodes within the risk control subdomain boundary, and dynamically adjust the subdomain boundary subdomain by subdomain.

[0052] Optionally, the S5 specifically includes:

[0053] S51. Calculate the change trend of the node risk status indicator for each node in the risk control subdomain, identify the node set with the same change trend of the risk status indicator node by node, and determine the spatial distribution characteristics of the node risk evolution in the subdomain;

[0054] S52, tracking the risk control instruction sequence historically executed by each node within the spatial distribution feature, determining the dependencies between the historical risk control instructions node by node, and forming a historical decision interaction network between nodes;

[0055] S53. Calculate the dynamic impact factor of historical risk control instructions on the current node risk state node by node based on the historical decision interaction network between nodes, and determine the specific effect of historical decisions on the current risk state;

[0056] S54, extracting node risk status indicators, historical decision dynamic influencing factors, and node spatiotemporal embedding vectors, constructing a joint state data set of risk status, decision history, and spatiotemporal features for each node, and fusing them to generate a three-dimensional joint optimization state space;

[0057] S55. Calculate the sensitivity coefficient of the node risk state change to the risk control instruction update requirement node by node according to the three-dimensional joint optimization state space, and dynamically determine the priority of the node risk control instruction update;

[0058] S56. Determine the update interval of the risk control instruction sequence in the subdomain based on the node update priority, and dynamically adjust the execution order of the risk control instructions of each node;

[0059] S57. Monitor the fluctuation amplitude and frequency of the overall risk status of the risk control subdomain, automatically trigger an updated risk control instruction sequence, and execute the risk control instructions node by node.

[0060] Optionally, the S6 specifically includes:

[0061] S61. When a risk control instruction sequence is executed within a risk control subdomain, the subdomain boundary node monitors the change in its spatiotemporal embedding vector and risk score relative to the last instruction execution. When any threshold change reaches a preset asynchronous trigger condition, the boundary node automatically generates a boundary node feature snapshot containing the node spatiotemporal embedding vector, risk quantification value, timestamp, and node unique identifier.

[0062] S62. Based on the self-organizing mesh topology, encapsulate the boundary node feature snapshot into a broadcast data packet with a time slot identifier, and periodically broadcast it to the subdomains directly physically adjacent to the current subdomain via adjacent links in a multi-path parallel manner.

[0063] S63. After receiving the boundary node feature snapshots broadcast by adjacent subdomains, each subdomain parses the spatiotemporal embedding vectors and risk quantification values ​​of the boundary nodes in the snapshots and sorts them by timestamp to construct a cross-period boundary coupling graph containing boundary node information for the current period and the previous period. The nodes of the boundary coupling graph represent boundary nodes, and the edges represent risk associations between nodes in adjacent periods.

[0064] S64. For each edge in the cross-period boundary coupling graph, calculate the difference in spatiotemporal embedding vectors and the volatility of risk quantification values ​​of boundary nodes in adjacent time periods, compare the combined indicators of the difference and volatility with a preset mutation threshold, automatically identify the risk volatility mutation edges that meet the condition of being above the threshold, and generate a dynamic negotiation trigger signal using the mutation edge as the trigger source;

[0065] S65: The dynamic negotiation trigger signal is broadcast to all boundary nodes via multiple backup communication channels within the subdomain in a load-balanced manner. After receiving the trigger signal, the boundary node performs a weighted fusion calculation based on the local spatiotemporal embedding vector and the neighborhood snapshot information to adjust the risk quantification value calibration factor of the boundary node.

[0066] S66. Each subdomain self-organizes and generates a micro-transition risk control instruction mapping sequence containing micro-transition increments and time sequence based on the calibrated boundary node risk rating values ​​and multi-subdomain interaction feedback information. Each micro-transition increment in the mapping sequence is combined with the risk gradient difference of neighboring subdomains to drive the evolutionary self-coordination of cross-subdomain risk control instructions.

[0067] S67. Each subdomain sends the updated risk control instructions to each node within the subdomain according to the mapping sequence, and continuously monitors the changes in the spatiotemporal embedding vectors of the boundary nodes and the calibrated risk quantification values. When the spatiotemporal embedding vector change rate or the risk quantification value volatility of any boundary node meets the asynchronous trigger condition again, a new boundary node feature snapshot is automatically generated to form an adaptive boundary suppression closed loop and iteratively update the state of the two-way information collaboration link.

[0068] Optionally, the S7 specifically includes:

[0069] S71. For each risk control subdomain, asynchronously capture the spatiotemporal embedding vectors of all nodes in the subdomain and the corresponding risk control instruction execution feedback states, and generate a distributed adaptive error spectrum through node state transition accumulation, where the distributed adaptive error spectrum represents the error distribution characteristics of the node risk state in the temporal and spatial domains;

[0070] S72. Compare the distributed adaptive error spectrum with the historical multi-scale error spectrum in parallel, and use the spectrum resonance detection mechanism to compare the spectrum energy level distribution. When the error energy level aggregation across time periods appears in the error spectrum and the spectrum resonance coefficient exceeds the preset stability threshold, identify the high-energy-level error node cluster, thereby determining that the local risk convergence state has degraded.

[0071] S73. After determining that the local risk convergence state has degraded, trigger the subdomain internal path regeneration mechanism and autonomously query the redundant control path library for multiple redundant path candidate sets corresponding to the high-energy error node group. The redundant control path library is pre-classified and stored according to node spatiotemporal embedding matching degree, path redundancy, and fault tolerance level.

[0072] S74. For the candidate redundant path set, sequentially calculate the spatiotemporal embedding vector spectrum compatibility index of the nodes in each path, the link oscillation amplitude index of the path relay node, and the node load jump tolerance index, and comprehensively adopt a parallel adaptation priority calculation strategy to select the redundant control path with the highest adaptation priority as the switching target;

[0073] S75. Based on the information of the starting node, relay node, and terminating node of the selected redundant control path, generate a token command sequence according to the self-organizing token ring protocol. The token command sequence includes control instructions and token transfer rules for each node in the redundant path, and seamlessly switch to the token command sequence in the current control time slot.

[0074] S76: After switching to the redundant control path, the asynchronous hedging mechanism is started in parallel within the subdomain to collect the convergence rate difference and risk control resonance strength of each node on the redundant path. When the resonance strength is lower than the preset resonance stability threshold, the rapid rollback mechanism is triggered and step S73 is re-entered to loop search and switch to the next redundant control path.

[0075] S77. When the switched redundant control path reaches a state of convergence and risk resonance equilibrium, the adaptive priority weight of the corresponding path in the redundant control path library is updated through the dynamic memory interaction mechanism, and the asynchronous capture and monitoring of the conventional distributed adaptive error spectrum in the subdomain is restored.

[0076] The beneficial effects of the present invention are:

[0077] (1) The present invention constructs a node-edge dual-space coupling dynamic graph and calculates the node feature change trend in real time to generate a multi-dimensional feature space, thereby achieving high-precision real-time mapping between distribution network nodes and lines, effectively improving the accuracy of risk assessment and the dynamic response speed, and enhancing the adaptability to complex operating conditions.

[0078] (2) The present invention automatically identifies the dominant path of risk diffusion and divides the risk control subdomains in real time through dynamic clustering and implicit feature interactive perception network, which significantly improves the accuracy of risk status prediction and the efficiency of risk propagation path identification, and shows better real-time adaptability in distribution network risk outbreak and propagation scenarios.

[0079] (3) In terms of cross-regional risk collaborative control, the present invention effectively solves the problem of the lack of real-time self-coordination capability of traditional methods by constructing a distributed two-way information collaboration link and an asynchronous event triggering mechanism between neighborhoods. It breaks through the bottleneck of the existing technology that risk control instructions are difficult to quickly and synchronously update in complex node interaction scenarios, and realizes efficient collaboration and adaptive optimization of risk control instructions, thereby effectively improving the overall risk control stability of the distribution network.

[0080] (4) The present invention establishes a redundant control path library and a self-organizing token ring mechanism to monitor and quickly determine the degradation or failure of the risk convergence state in real time, automatically start the redundant path to achieve seamless switching, significantly reduce the delay and error when switching the risk decision path, and effectively ensure the real-time and robustness of the distribution network operation risk decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0082] Figure 1This is a schematic diagram of the node-edge dual-space coupling dynamic graph structure of a distribution network operation risk control method based on Bayesian optimization proposed in the present invention;

[0083] Figure 2 Schematic diagram of the risk propagation trajectory and risk control subdomain division of a distribution network operation risk control method based on Bayesian optimization proposed in the present invention;

[0084] Figure 3 This is a schematic diagram of the distributed bidirectional information cooperation link structure between neighborhoods for the distribution network operation risk control method based on Bayesian optimization proposed in the present invention. DETAILED DESCRIPTION

[0085] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0086] refer to Figure 1-Figure 3 , a distribution network operation risk control method based on Bayesian optimization, comprising the following steps:

[0087] S1. Build a node-edge dual-space coupling dynamic graph using multi-source operation data of the distribution network;

[0088] S2. Based on the dynamic graph, through the self-supervision task of multi-level spatiotemporal interaction between nodes and their neighborhoods, the associated feature weights of nodes in different operating states are dynamically adjusted node by node to generate spatiotemporal embedding vectors;

[0089] S3. Based on spatiotemporal embedding vectors, a risk-sensitive network with multi-scale dynamic clustering is constructed to automatically extract the risk diffusion path under the operating state and predict the risk propagation trajectory;

[0090] S4. Utilize risk propagation trajectories to automatically form a continuous and dynamic risk evolution link across time periods within the distribution network, and dynamically divide the grid risk control subdomains based on the spatiotemporal embedding trend of the link;

[0091] S5. For the dynamically formed risk control subdomain, evaluate the dynamic evolution of the operational risk within the subdomain, construct a joint optimization state space of risk state, decision history, and spatiotemporal characteristics, and update the risk control instruction sequence based on the evaluation results;

[0092] S6. During the dynamic execution of the risk control instruction sequence, a distributed bidirectional information collaboration link is constructed between neighboring domains to synchronize the risk status between risk control subdomains and to self-coordinate the adjustment of cross-regional risk control instructions;

[0093] S7. Monitor the dynamic convergence of decisions in each risk control subdomain. When it is predicted that the local risk convergence state has degraded, automatically start the pre-built redundant control path to achieve seamless switching of risk decision paths.

[0094] By constructing a node-edge dual-space coupling dynamic graph, refined dynamic modeling of multi-source data in the distribution network is achieved, and spatiotemporal embedding vectors are generated using multi-level self-supervision tasks, thereby improving the accuracy of risk status assessment. A risk-sensitive network combined with multi-scale dynamic clustering is used to accurately predict risk propagation paths and dynamically divide risk control subdomains, thereby improving the real-time and targeted nature of risk management. By constructing a distributed bidirectional information collaboration link between the state space and neighborhoods that jointly optimizes risk status, decision history, and spatiotemporal characteristics, real-time self-coordination adjustment of cross-regional risk control instructions is effectively achieved. At the same time, redundant control paths are established to ensure seamless switching of risk decision paths, thereby improving the robustness and reliability of the overall system.

[0095] In this embodiment, S1 specifically includes:

[0096] S11. Collect the flow data, voltage data, and load data of all nodes in the distribution network, calculate the data change rate and change trend of each node, and use the node data change rate to build a dynamic node feature hierarchical system;

[0097] S12. Determine the dynamic association relationship between nodes based on the node feature dynamic hierarchical system, calculate the data association strength and association direction between nodes, and form an updated node association strength matrix;

[0098] S13. Dynamically select the distribution lines directly connected to each node based on the node association strength matrix, calculate the line load level and fluctuation trend line by line, and establish a dynamic association mapping between the distribution network nodes and the corresponding lines;

[0099] S14. Based on the dynamic association mapping, a node-edge bidirectional state coupling network is constructed that simultaneously expresses the electrical connection relationship between nodes and the dynamic change trend of data, and the coupling network topology is updated;

[0100] S15, detecting a topology change of the node-edge bidirectional state coupling network, and adaptively updating a dynamic coupling weight between the node and the corresponding line according to the frequency of topology change, the number of changed nodes, and the severity of the coupling state;

[0101] S16. Convert the node-edge bidirectional state coupling network into a dynamic graph according to the updated dynamic coupling weights and topology structure.

[0102] By collecting flow, voltage and load data in real time, calculating the change rate and trend of node data, establishing a dynamic hierarchical system of node characteristics, and determining the data correlation strength and direction between nodes in real time, a dynamic correlation strength matrix and a node-edge bidirectional state coupling network are constructed, effectively realizing real-time and refined mapping of node and line states; by dynamically detecting changes in network topology and adaptively updating the coupling weights between nodes and lines, the real-time performance and accuracy of the distribution network dynamic graph model are improved, and the responsiveness and accuracy of distribution network operation risk assessment and control are further optimized.

[0103] In this embodiment, S2 specifically includes:

[0104] S21. Based on the topological structure of the dynamic graph and changes in node feature data, identify adjacent nodes in the node neighborhood whose data interaction frequency is higher than a preset threshold, and determine the direction of data interaction between the nodes to form a node neighborhood interaction relationship graph;

[0105] S22. Based on the node neighborhood interaction relationship diagram, the node neighborhood is divided into direct neighborhood, secondary neighborhood, and edge neighborhood to form a multi-level structure of the node neighborhood, and the feature interaction priority of different neighborhood levels is determined;

[0106] S23, determining a characteristic sensitivity parameter based on the amplitude and frequency of changes in the characteristic data of the node itself, and screening the characteristic data in different levels of neighborhood of the node according to the characteristic sensitivity parameter;

[0107] S24, calculating the data interaction sensitivity coefficient between the node and the neighboring nodes at each level based on the neighborhood feature data, and dynamically adjusting the feature association weights of the nodes in each neighborhood level node by node based on the sensitivity coefficient;

[0108] S25. Determine the optimal feature aggregation mode based on the dynamically adjusted feature association weights, and use it to aggregate the filtered feature information of the node itself and its neighborhood.

[0109] S26. Using the feature aggregation mode, the node's own features are integrated with the multi-level neighborhood feature data to generate a spatiotemporal embedding vector for each node.

[0110] By real-time identification of adjacent nodes with high data interaction frequency in the node neighborhood, a node neighborhood interaction relationship diagram is formed, and the real-time optimal feature aggregation mode is determined based on the data interaction sensitivity parameters and feature association weights between the node and the neighborhood; then, the node's own characteristics and the neighborhood feature data are integrated to generate a spatiotemporal embedding vector, realizing adaptive dynamic adjustment of node characteristics, effectively improving the accuracy and real-time performance of distribution network risk status assessment, and enhancing the system's adaptability to complex dynamic operating environments.

[0111] In this embodiment, S3 specifically includes:

[0112] S31. Based on the spatiotemporal embedding vectors of nodes, a multi-dimensional dynamic perception network of feature interactions between nodes is constructed to identify implicit feature interaction relationships between nodes one by one.

[0113] S32, determining multi-scale cluster centers according to implicit feature interaction relationships, and generating dynamic clustering topology between nodes using cluster centers at each scale;

[0114] S33. Based on the change trend of the node embedding vector in the dynamic clustering topology, detect the boundary area where the feature changes between the node sets are abnormal, and form a risk-sensitive area with the boundary area as the core;

[0115] S34. Track the node sequence with the most drastic fluctuations in node embedding vectors within the risk-sensitive area and extract the dominant path of risk diffusion;

[0116] S35. Analyze the temporal change direction and change intensity of the node embedding vector node by node along the risk diffusion dominant path to form a continuous risk diffusion state chain;

[0117] S36. Dynamically evaluate the state transitions between nodes in the risk diffusion state chain to generate a multi-step continuous evolution model of risk diffusion;

[0118] S37. Based on the multi-step continuous evolution model of risk diffusion, the risk diffusion development direction of the risk diffusion path is deduced, and the risk propagation trajectory under the operation status of the distribution network is predicted.

[0119] By constructing a multi-dimensional dynamic perception network for feature interaction between nodes, real-time identification of implicit feature interaction relationships between nodes is achieved, and a multi-scale clustering topology structure is determined to further accurately divide risk-sensitive areas. By tracking node sequences with drastic fluctuations in node embedding vectors, the dominant path of risk diffusion is extracted, and the node state transition characteristics are analyzed in real time to generate a multi-step continuous evolution model of risk diffusion, thereby accurately deducing the subsequent development direction of the risk diffusion path, effectively improving the accuracy and real-time performance of the distribution network risk propagation trajectory prediction, and enhancing the perception of the dynamic evolution of risks.

[0120] In this embodiment, the S31 specifically includes:

[0121] S311, calculating the change rate and direction of the node feature vector based on the node spatiotemporal embedding vector, and generating a multi-dimensional feature space of the node feature change trend node by node;

[0122] S312, mapping the features of each node in the multi-dimensional feature space, and determining the spatial position relationship between each node and the feature change trend of other nodes;

[0123] S313, identifying characteristic conduction paths between nodes based on the spatial position relationship of the node feature change trend, and calibrating the conduction direction on the path;

[0124] S314, calculating the contribution coefficient of each node in the characteristic conduction path to the overall characteristic conduction strength of the path node by node, and determining the key interaction nodes whose contribution coefficient exceeds a preset threshold;

[0125] S315, monitoring the synchronization of feature change trends between key interaction nodes, calculating a synchronization index and using the index to characterize the implicit feature interaction strength between the nodes;

[0126] S316. According to the implicit feature interaction strength, the implicit feature interaction relationship between nodes is divided into three levels: high intensity, medium intensity and low intensity, forming an implicit feature interaction grading system.

[0127] By calculating the rate and direction of change of node feature vectors, a multi-dimensional feature space is generated, and the spatial position relationship of the feature change trend between nodes is determined by real-time mapping, and the feature conduction path between nodes is identified and calibrated; further, the contribution coefficient of the node to the overall feature conduction strength of the path is calculated, the key interaction nodes are determined and the synchronization of their feature change trends is monitored, so as to quantify the implicit feature interaction strength between nodes in real time and establish a clear implicit feature interaction grading system; therefore, the present invention can more accurately describe the implicit feature interaction relationship between nodes, effectively improving the precision and reliability of risk feature propagation analysis.

[0128] In this embodiment, the S4 specifically includes:

[0129] S41. Based on the changing trend of the spatiotemporal embedding vectors of nodes within the risk propagation trajectory, the continuous risk transmission coupling paths between nodes across time periods are automatically extracted node by node, and the characteristic change trend vector of each node in the path is calculated;

[0130] S42. Calculate the change rate and change acceleration of the characteristic change trend vector of each node on the risk conduction coupling path node by node, and determine the node whose acceleration exceeds a preset threshold as a risk state turning point node;

[0131] S43. Divide the risk transmission coupling path into multiple independent risk propagation links according to the risk state turning points, and determine the starting node, ending node, and spatial boundary of each link.

[0132] S44, tracking the changes in the spatiotemporal embedding vectors of all nodes in each risk propagation link, establishing a spatiotemporal feature difference matrix of nodes link by link, and calculating the feature difference index within the matrix;

[0133] S45. Analyze the similarities and differences between any two adjacent risk propagation links based on the characteristic difference index to determine the correlation level between the links;

[0134] S46. Adaptively aggregate the risk propagation links based on the correlation levels between the links, and dynamically divide them into multiple risk control subdomains;

[0135] S47. Monitor the characteristic change trends and risk propagation intensity of nodes within the risk control subdomain boundary, and dynamically adjust the subdomain boundary subdomain by subdomain.

[0136] By calculating the changing trends and accelerations of node characteristics on the risk transmission path in real time, the risk state turning points are automatically identified and the risk transmission links are accurately divided. The feature difference matrix is ​​established and the difference index is calculated using the node spatiotemporal embedding vector to determine the correlation level between links, and the risk transmission links are dynamically aggregated or separated to form multiple risk control subdomains in real time. By dynamically monitoring the node feature trends and risk transmission intensity within the subdomain boundaries, the subdomain boundaries are adjusted in real time, thereby improving the accuracy and flexibility of the risk control area division and effectively improving the overall performance of distribution network risk management.

[0137] In this embodiment, the S5 specifically includes:

[0138] S51. Calculate the change trend of the node risk status indicator for each node in the risk control subdomain, identify the node set with the same change trend of the risk status indicator node by node, and determine the spatial distribution characteristics of the node risk evolution in the subdomain;

[0139] S52, tracking the risk control instruction sequence historically executed by each node within the spatial distribution feature, determining the dependencies between the historical risk control instructions node by node, and forming a historical decision interaction network between nodes;

[0140] S53. Calculate the dynamic impact factor of historical risk control instructions on the current node risk state node by node based on the historical decision interaction network between nodes, and determine the specific effect of historical decisions on the current risk state;

[0141] S54, extracting node risk status indicators, historical decision dynamic influencing factors, and node spatiotemporal embedding vectors, constructing a joint state data set of risk status, decision history, and spatiotemporal features for each node, and fusing them to generate a three-dimensional joint optimization state space;

[0142] S55. Calculate the sensitivity coefficient of the node risk state change to the risk control instruction update requirement node by node according to the three-dimensional joint optimization state space, and dynamically determine the priority of the node risk control instruction update;

[0143] S56. Determine the update interval of the risk control instruction sequence in the subdomain based on the node update priority, and dynamically adjust the execution order of the risk control instructions of each node;

[0144] S57. Monitor the fluctuation amplitude and frequency of the overall risk status of the risk control subdomain, automatically trigger an updated risk control instruction sequence, and execute the risk control instructions node by node.

[0145] By calculating the changing trends of node risk status indicators in real time and identifying the node risk evolution characteristics, a real-time interactive network of node historical decisions is formed, and the dynamic impact of historical risk decisions on the current risk status is quantitatively analyzed. A three-dimensional joint optimization state space of risk status-decision history-spatiotemporal characteristics is constructed, and the sensitivity coefficient and priority of risk control instruction updates are determined in real time to dynamically adjust the execution order and update frequency of node risk control instructions, thereby improving the pertinence and real-time nature of the risk control strategy and significantly enhancing the dynamic adaptability and response speed of distribution network risk decision-making.

[0146] In this embodiment, S6 specifically includes:

[0147] S61. When a risk control instruction sequence is executed within a risk control subdomain, the subdomain boundary node monitors the change in its spatiotemporal embedding vector and risk score relative to the last instruction execution. When any threshold change reaches a preset asynchronous trigger condition, the boundary node automatically generates a boundary node feature snapshot containing the node spatiotemporal embedding vector, risk quantification value, timestamp, and node unique identifier.

[0148] S62. Based on the self-organizing mesh topology, encapsulate the boundary node feature snapshot into a broadcast data packet with a time slot identifier, and periodically broadcast it to the subdomains directly physically adjacent to the current subdomain via adjacent links in a multi-path parallel manner.

[0149] S63. After receiving the boundary node feature snapshots broadcast by adjacent subdomains, each subdomain parses the spatiotemporal embedding vectors and risk quantification values ​​of the boundary nodes in the snapshots and sorts them by timestamp to construct a cross-period boundary coupling graph containing boundary node information for the current period and the previous period. The nodes of the boundary coupling graph represent boundary nodes, and the edges represent risk associations between nodes in adjacent periods.

[0150] S64. For each edge in the cross-period boundary coupling graph, calculate the difference in spatiotemporal embedding vectors and the volatility of risk quantification values ​​of boundary nodes in adjacent time periods, compare the combined indicators of the difference and volatility with a preset mutation threshold, automatically identify the risk volatility mutation edges that meet the condition of being above the threshold, and generate a dynamic negotiation trigger signal using the mutation edge as the trigger source;

[0151] S65: The dynamic negotiation trigger signal is broadcast to all boundary nodes via multiple backup communication channels within the subdomain in a load-balanced manner. After receiving the trigger signal, the boundary node performs a weighted fusion calculation based on the local spatiotemporal embedding vector and the neighborhood snapshot information to adjust the risk quantification value calibration factor of the boundary node.

[0152] S66. Each subdomain self-organizes and generates a micro-transition risk control instruction mapping sequence containing micro-transition increments and time sequence based on the calibrated boundary node risk rating values ​​and multi-subdomain interaction feedback information. Each micro-transition increment in the mapping sequence is combined with the risk gradient difference of neighboring subdomains to drive the evolutionary self-coordination of cross-subdomain risk control instructions.

[0153] S67. Each subdomain sends the updated risk control instructions to each node within the subdomain according to the mapping sequence, and continuously monitors the changes in the spatiotemporal embedding vectors of the boundary nodes and the calibrated risk quantification values. When the spatiotemporal embedding vector change rate or the risk quantification value volatility of any boundary node meets the asynchronous trigger condition again, a new boundary node feature snapshot is automatically generated to form an adaptive boundary suppression closed loop and iteratively update the state of the two-way information collaboration link.

[0154] Through the asynchronous trigger mechanism of boundary nodes, node feature changes are monitored in real time, boundary node feature snapshots are generated, and multi-path parallel broadcasting is performed using the self-organizing grid topology to realize the construction of a cross-time boundary coupling diagram; by automatically identifying the spatiotemporal embedding differences and risk quantification volatility of boundary nodes, negotiation trigger signals and micro-transition risk control instruction mapping sequences are dynamically generated, realizing real-time coordination of information between risk control subdomains and dynamic self-coordination adjustment of risk control instructions, effectively improving the real-time response capability and cross-regional collaboration accuracy of distribution network risk control, and enhancing the overall robustness of the risk control strategy.

[0155] In this embodiment, the S7 specifically includes:

[0156] S71. For each risk control subdomain, asynchronously capture the spatiotemporal embedding vectors of all nodes in the subdomain and the corresponding risk control instruction execution feedback states, and generate a distributed adaptive error spectrum through node state transition accumulation, where the distributed adaptive error spectrum represents the error distribution characteristics of the node risk state in the temporal and spatial domains;

[0157] S72. Compare the distributed adaptive error spectrum with the historical multi-scale error spectrum in parallel, and use the spectrum resonance detection mechanism to compare the spectrum energy level distribution. When the error energy level aggregation across time periods appears in the error spectrum and the spectrum resonance coefficient exceeds the preset stability threshold, identify the high-energy-level error node cluster, thereby determining that the local risk convergence state has degraded.

[0158] S73. After determining that the local risk convergence state has degraded, trigger the subdomain internal path regeneration mechanism and autonomously query the redundant control path library for multiple redundant path candidate sets corresponding to the high-energy error node group. The redundant control path library is pre-classified and stored according to node spatiotemporal embedding matching degree, path redundancy, and fault tolerance level.

[0159] S74. For the candidate redundant path set, sequentially calculate the spatiotemporal embedding vector spectrum compatibility index of the nodes in each path, the link oscillation amplitude index of the path relay node, and the node load jump tolerance index, and comprehensively adopt a parallel adaptation priority calculation strategy to select the redundant control path with the highest adaptation priority as the switching target;

[0160] S75. Based on the information of the starting node, relay node, and terminating node of the selected redundant control path, generate a token command sequence according to the self-organizing token ring protocol. The token command sequence includes control instructions and token transfer rules for each node in the redundant path, and seamlessly switch to the token command sequence in the current control time slot.

[0161] S76: After switching to the redundant control path, the asynchronous hedging mechanism is started in parallel within the subdomain to collect the convergence rate difference and risk control resonance strength of each node on the redundant path. When the resonance strength is lower than the preset resonance stability threshold, the rapid rollback mechanism is triggered and step S73 is re-entered to loop search and switch to the next redundant control path.

[0162] S77. When the switched redundant control path reaches a state of convergence and risk resonance equilibrium, the adaptive priority weight of the corresponding path in the redundant control path library is updated through the dynamic memory interaction mechanism, and the asynchronous capture and monitoring of the conventional distributed adaptive error spectrum in the subdomain is restored.

[0163] By asynchronously capturing node state transition data, a distributed adaptive error spectrum is constructed and the spectrum resonance mechanism is used to identify local risk convergence degradation states. The redundant path regeneration mechanism is triggered, the pre-stored path library is autonomously queried, and path selection is performed based on spatiotemporal embedding compatibility, link oscillation amplitude, and fault tolerance indicators. The self-organizing token ring protocol is used to achieve seamless switching of risk control paths, monitor the convergence rate and risk resonance intensity in real time, and update the path priority through a dynamic memory mechanism, thereby effectively improving the ability to quickly identify and respond to risk convergence degradation states, and significantly improving the real-time and reliability of risk decision-making path switching.

[0164] Example 1:

[0165] In order to verify the feasibility of the present invention in practice, the present invention was applied to the real-time control and management of operational risks in a regional power grid under a certain power grid company. Specifically, it involved risk status monitoring, risk propagation path prediction, and control strategy optimization tasks for multiple substations and distributed renewable energy power generation nodes within the regional power grid. The regional distribution network has a high proportion of renewable energy, rapid load changes, and complex interactions between nodes. Traditional methods rely heavily on manual experience to preset static risk models and parameters, and are unable to accurately assess and effectively control risks in real time. This is especially true under sudden abnormal load fluctuations and extreme weather conditions. Traditional models have poor prediction accuracy and significant response lag, making it difficult to quickly formulate effective risk control strategies, posing a significant challenge to the safe and stable operation of the power grid.

[0166] In actual application, first, the flow data, voltage data and load data of all nodes in the distribution network are collected in real time, and the data change rate and change trend of each node are calculated in real time to generate the dynamic feature space of the node, and the real-time dynamic association relationship between the nodes is determined; then, according to the dynamic association relationship of the nodes, the lines directly connected to the nodes are selected, and the real-time load level and fluctuation trend of the lines are calculated line by line, and a real-time node-line status association mapping diagram is established; on this basis, a node-edge dual-space dynamic coupling network is constructed, and the network topology changes are detected in real time and the network weights are updated to form a dynamic graph model of the distribution network.

[0167] Next, the system performs self-supervised learning on the dynamic graph of the distribution network. Through a multi-level node neighborhood feature interaction mechanism, it dynamically adjusts the node feature association weights to obtain the node's spatiotemporal embedding vector. Then, based on the node embedding vector, it constructs a multi-scale dynamic clustering risk-sensitive network. This network automatically identifies the dominant risk diffusion paths in real time, forms risk propagation trajectories, and predicts risk diffusion trends. Based on these real-time risk propagation trajectories, the system then divides the distribution network into multiple risk control subdomains to accurately monitor and manage risk propagation.

[0168] For each divided risk control subdomain, the system evaluates the dynamic evolution of risk, integrating node risk status indicators, historical risk control decisions, and node spatiotemporal embedding features in real time to establish a three-dimensional joint optimization state space of risk status, decision history, and spatiotemporal features. This automatically updates the risk control instruction sequence for nodes within the subdomain. Furthermore, by establishing distributed bidirectional information collaboration links between neighboring subdomains, subdomains share risk status information on boundary nodes in real time, enabling self-coordinated optimization of cross-regional risk control strategies.

[0169] During the execution of risk control instructions, the system monitors changes in the spatiotemporal embedding vectors and risk scores of boundary nodes in real time. Once a node's feature changes meet a preset asynchronous trigger condition, the boundary node immediately generates a feature snapshot and broadcasts it to adjacent subdomains, enabling real-time synchronization of risk status across subdomains. Furthermore, by building an asynchronous monitoring mechanism and a distributed error spectrum resonance analysis method, the system captures and determines subdomain risk convergence degradation in real time. Once significant risk convergence degradation is detected, a preset redundant path switching strategy is immediately triggered.

[0170] To verify the practical effects of the present invention, 30 key nodes within the regional power grid were selected for risk status prediction and control effect evaluation. The performance of the traditional manual static risk assessment method and the present invention's method in predicting risk status indicators, risk propagation paths, and risk convergence effects was compared. During the specific experimental process, the system used a Bayesian optimization algorithm for automatic optimization and adaptive adjustment of model parameters. After 50 rounds of Bayesian optimization iterations, the final selection was a risk feature embedding dimension of 64, a risk clustering scale of 4, a risk-sensitive network dynamic clustering radius threshold of 0.05, a risk control instruction update sensitivity coefficient threshold of 0.08, and a boundary node asynchronous trigger threshold of 0.03.

[0171] The following table shows the comparison data between the risk status prediction results and the measured values ​​of some key nodes:

[0172] Table 1 Comparison of node risk status prediction and measured performance

[0173] Node number Measured value of risk status Risk status prediction value Absolute value of error A-01 0.58 0.56 0.02 A-05 0.73 0.71 0.02 A-12 0.65 0.66 0.01 B-07 0.41 0.40 0.01 B-15 0.85 0.83 0.02

[0174] As can be clearly seen from the data in Table 1, the absolute error in node risk status prediction using the present invention is consistently within 0.02, demonstrating the high accuracy and reliability of risk status prediction. Taking node A-05 as an example, the traditional static risk model has a prediction error of approximately 0.11, while the proposed method has a prediction error of only 0.02, an 81.8% reduction compared to the traditional method, demonstrating a significant improvement in the accuracy of the proposed risk assessment model.

[0175] To further analyze the performance of risk transmission path prediction, five leading risk transmission paths were randomly selected and compared with the subsequent measured paths. The results are shown in the following table:

[0176] Table 2 Comparison of predicted and measured risk transmission paths

[0177] Propagation path number Measured path node sequence Predicting path node sequence Path accuracy P-1 A-01→A-05→A-12 A-01→A-05→A-12 100％ P-2 B-03→B-07→B-15 B-03→B-07→B-15 100％ P-3 C-08→C-11→C-14 C-08→C-10→C-14 66.7％ P-4 D-02→D-05→D-09 D-02→D-05→D-09 100％ P-5 E-01→E-04→E-08 E-01→E-04→E-07 66.7％

[0178] As shown in Table 2, the prediction accuracy of the risk propagation path of the present invention is generally high, and the accuracy of most paths reaches 100%, which is significantly better than the average accuracy of 60% of the traditional static model, reflecting the accuracy and reliability of the present invention in capturing risk diffusion trends.

[0179] In terms of risk convergence effect evaluation, the subdomain risk decision convergence period is used as the evaluation indicator. The average convergence time of the traditional method is 18 minutes, while after asynchronous monitoring and redundant path switching optimization, the average convergence time of the method of the present invention is reduced to 4.2 minutes, and the convergence speed is increased by about 76.7%, which greatly improves the risk management efficiency of the distribution network.

[0180] In summary, this embodiment fully verifies that the present invention effectively realizes node-line dynamic coupling modeling, accurate risk status prediction and efficient risk propagation path identification, real-time risk decision-making dynamic coordination and redundant path rapid switching optimization through the deep integration of the self-supervised spatiotemporal graph transformer and the meta-reinforcement learning optimization algorithm. In practical applications, it significantly improves the accuracy, real-time and robustness of distribution network risk prediction and control, can provide reliable technical support for real-time risk management in complex dynamic distribution network environments, and has broad application prospects.

[0181] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A distribution network operation risk control method based on Bayesian optimization, characterized in that: The steps include: S1. Build a node-edge dual-space coupling dynamic graph using multi-source operation data of the distribution network; S2. Based on the dynamic graph, through the self-supervision task of multi-level spatiotemporal interaction between nodes and their neighborhoods, the associated feature weights of nodes in different operating states are dynamically adjusted node by node to generate spatiotemporal embedding vectors; S3. Based on spatiotemporal embedding vectors, a risk-sensitive network with multi-scale dynamic clustering is constructed to automatically extract the risk diffusion path under the operating state and predict the risk propagation trajectory; S4. Utilize risk propagation trajectories to automatically form a continuous and dynamic risk evolution link across time periods within the distribution network, and dynamically divide the grid risk control subdomains based on the spatiotemporal embedding trend of the link; S5. For the dynamically formed risk control subdomain, evaluate the dynamic evolution of the operational risk within the subdomain, construct a joint optimization state space of risk state, decision history, and spatiotemporal characteristics, and update the risk control instruction sequence based on the evaluation results; S6. During the dynamic execution of the risk control instruction sequence, a distributed bidirectional information collaboration link is constructed between neighboring domains to synchronize the risk status between risk control subdomains and to self-coordinate the adjustment of cross-regional risk control instructions; S7. Monitor the dynamic convergence of decisions in each risk control subdomain. When it is predicted that the local risk convergence state has degraded, automatically start the pre-built redundant control path to achieve seamless switching of risk decision paths.

2. A distribution network operation risk control method based on Bayesian optimization according to claim 1, characterized in that: Said S1 specifically includes: S11. Collect the flow data, voltage data, and load data of all nodes in the distribution network, calculate the data change rate and change trend of each node, and use the node data change rate to build a dynamic node feature hierarchical system; S12. Determine the dynamic association relationship between nodes based on the node feature dynamic hierarchical system, calculate the data association strength and association direction between nodes, and form an updated node association strength matrix; S13. Dynamically select the distribution lines directly connected to each node based on the node association strength matrix, calculate the line load level and fluctuation trend line by line, and establish a dynamic association mapping between the distribution network nodes and the corresponding lines; S14. Based on the dynamic association mapping, a node-edge bidirectional state coupling network is constructed that simultaneously expresses the electrical connection relationship between nodes and the dynamic change trend of data, and the coupling network topology is updated; S15, detecting a topology change of the node-edge bidirectional state coupling network, and adaptively updating a dynamic coupling weight between the node and the corresponding line according to the frequency of topology change, the number of changed nodes, and the severity of the coupling state; S16. Convert the node-edge bidirectional state coupling network into a dynamic graph according to the updated dynamic coupling weights and topology structure.

3. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S2 specifically includes: S21. Based on the topological structure of the dynamic graph and changes in node feature data, identify adjacent nodes in the node neighborhood whose data interaction frequency is higher than a preset threshold, and determine the direction of data interaction between the nodes to form a node neighborhood interaction relationship graph; S22. Based on the node neighborhood interaction relationship diagram, the node neighborhood is divided into direct neighborhood, secondary neighborhood, and edge neighborhood to form a multi-level structure of the node neighborhood, and the feature interaction priority of different neighborhood levels is determined; S23, determining a characteristic sensitivity parameter based on the amplitude and frequency of changes in the characteristic data of the node itself, and screening the characteristic data in different levels of neighborhood of the node according to the characteristic sensitivity parameter; S24, calculating the data interaction sensitivity coefficient between the node and the neighboring nodes at each level based on the neighborhood feature data, and dynamically adjusting the feature association weights of the nodes in each neighborhood level node by node based on the sensitivity coefficient; S25. Determine the optimal feature aggregation mode based on the dynamically adjusted feature association weights, and use it to aggregate the filtered feature information of the node itself and its neighborhood. S26. Using the feature aggregation mode, the node's own features are integrated with the multi-level neighborhood feature data to generate a spatiotemporal embedding vector for each node.

4. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S3 specifically includes: S31. Based on the spatiotemporal embedding vectors of nodes, a multi-dimensional dynamic perception network of feature interactions between nodes is constructed to identify implicit feature interaction relationships between nodes one by one. S32, determining multi-scale cluster centers according to implicit feature interaction relationships, and generating dynamic clustering topology between nodes using cluster centers at each scale; S33. Based on the change trend of the node embedding vector in the dynamic clustering topology, detect the boundary area where the feature changes between the node sets are abnormal, and form a risk-sensitive area with the boundary area as the core; S34. Track the node sequence with the most drastic fluctuations in node embedding vectors within the risk-sensitive area and extract the dominant path of risk diffusion; S35. Analyze the temporal change direction and change intensity of the node embedding vector node by node along the risk diffusion dominant path to form a continuous risk diffusion state chain; S36. Dynamically evaluate the state transitions between nodes in the risk diffusion state chain to generate a multi-step continuous evolution model of risk diffusion; S37. Based on the multi-step continuous evolution model of risk diffusion, the risk diffusion development direction of the risk diffusion path is deduced, and the risk propagation trajectory under the operation status of the distribution network is predicted.

5. A distribution network operation risk control method based on Bayesian optimization according to claim 4, characterized in that: The S31 specifically includes: S311, calculating the change rate and direction of the node feature vector based on the node spatiotemporal embedding vector, and generating a multi-dimensional feature space of the node feature change trend node by node; S312, mapping the features of each node in the multi-dimensional feature space, and determining the spatial position relationship between each node and the feature change trend of other nodes; S313, identifying characteristic conduction paths between nodes based on the spatial position relationship of the node feature change trend, and calibrating the conduction direction on the path; S314, calculating the contribution coefficient of each node in the characteristic conduction path to the overall characteristic conduction strength of the path node by node, and determining the key interaction nodes whose contribution coefficient exceeds a preset threshold; S315, monitoring the synchronization of feature change trends between key interaction nodes, calculating a synchronization index and using the index to characterize the implicit feature interaction strength between the nodes; S316. According to the implicit feature interaction strength, the implicit feature interaction relationship between nodes is divided into three levels: high intensity, medium intensity and low intensity, forming an implicit feature interaction grading system.

6. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the changing trend of the spatiotemporal embedding vectors of nodes within the risk propagation trajectory, the continuous risk transmission coupling paths between nodes across time periods are automatically extracted node by node, and the characteristic change trend vector of each node in the path is calculated; S42. Calculate the change rate and change acceleration of the characteristic change trend vector of each node on the risk conduction coupling path node by node, and determine the node whose acceleration exceeds a preset threshold as a risk state turning point node; S43. Divide the risk transmission coupling path into multiple independent risk propagation links according to the risk state turning points, and determine the starting node, ending node, and spatial boundary of each link. S44, tracking the changes in the spatiotemporal embedding vectors of all nodes in each risk propagation link, establishing a spatiotemporal feature difference matrix of nodes link by link, and calculating the feature difference index within the matrix; S45. Analyze the similarities and differences between any two adjacent risk propagation links based on the characteristic difference index to determine the correlation level between the links; S46. Adaptively aggregate the risk propagation links based on the correlation levels between the links, and dynamically divide them into multiple risk control subdomains; S47. Monitor the characteristic change trends and risk propagation intensity of nodes within the risk control subdomain boundary, and dynamically adjust the subdomain boundary subdomain by subdomain.

7. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S5 specifically includes: S51. Calculate the change trend of the node risk status indicator for each node in the risk control subdomain, identify the node set with the same change trend of the risk status indicator node by node, and determine the spatial distribution characteristics of the node risk evolution in the subdomain; S52, tracking the risk control instruction sequence historically executed by each node within the spatial distribution feature, determining the dependencies between the historical risk control instructions node by node, and forming a historical decision interaction network between nodes; S53. Calculate the dynamic impact factor of the historical risk control instructions on the current node risk state node by node based on the historical decision interaction network between nodes, and determine the specific effect of the historical decisions on the current risk state; S54, extracting node risk status indicators, historical decision dynamic influencing factors, and node spatiotemporal embedding vectors, constructing a joint state data set of risk status, decision history, and spatiotemporal features for each node, and fusing them to generate a three-dimensional joint optimization state space; S55. Calculate the sensitivity coefficient of the node risk state change to the risk control instruction update requirement node by node according to the three-dimensional joint optimization state space, and dynamically determine the priority of the node risk control instruction update; S56. Determine the update interval of the risk control instruction sequence in the subdomain based on the node update priority, and dynamically adjust the execution order of the risk control instructions of each node; S57. Monitor the fluctuation amplitude and frequency of the overall risk status of the risk control subdomain, automatically trigger an updated risk control instruction sequence, and execute the risk control instructions node by node.

8. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S6 specifically includes: S61. When a risk control instruction sequence is executed within a risk control subdomain, the subdomain boundary node monitors the change in its spatiotemporal embedding vector and risk score relative to the last instruction execution. When any threshold change reaches a preset asynchronous trigger condition, the boundary node automatically generates a boundary node feature snapshot containing the node spatiotemporal embedding vector, risk quantification value, timestamp, and node unique identifier. S62. Based on the self-organizing mesh topology, encapsulate the boundary node feature snapshot into a broadcast data packet with a time slot identifier, and periodically broadcast it to the subdomains directly physically adjacent to the current subdomain via adjacent links in a multi-path parallel manner. S63. After receiving the boundary node feature snapshots broadcast by adjacent subdomains, each subdomain parses the spatiotemporal embedding vectors and risk quantification values ​​of the boundary nodes in the snapshots and sorts them by timestamp to construct a cross-period boundary coupling graph containing boundary node information for the current period and the previous period. The nodes of the boundary coupling graph represent boundary nodes, and the edges represent risk associations between nodes in adjacent periods. S64. For each edge in the cross-period boundary coupling graph, calculate the difference in spatiotemporal embedding vectors and the volatility of risk quantification values ​​of boundary nodes in adjacent time periods, compare the combined indicators of the difference and volatility with a preset mutation threshold, automatically identify the risk volatility mutation edges that meet the condition of being above the threshold, and generate a dynamic negotiation trigger signal using the mutation edge as the trigger source; S65: The dynamic negotiation trigger signal is broadcast to all boundary nodes via multiple backup communication channels within the subdomain in a load-balanced manner. After receiving the trigger signal, the boundary node performs a weighted fusion calculation based on the local spatiotemporal embedding vector and the neighborhood snapshot information to adjust the risk quantification value calibration factor of the boundary node. S66. Each subdomain self-organizes and generates a micro-transition risk control instruction mapping sequence containing micro-transition increments and time sequence based on the calibrated boundary node risk rating values ​​and multi-subdomain interaction feedback information. Each micro-transition increment in the mapping sequence is combined with the risk gradient difference of neighboring subdomains to drive the evolutionary self-coordination of cross-subdomain risk control instructions. S67. Each subdomain sends the updated risk control instructions to each node within the subdomain according to the mapping sequence, and continuously monitors the changes in the spatiotemporal embedding vectors of the boundary nodes and the calibrated risk quantification values. When the spatiotemporal embedding vector change rate or the risk quantification value volatility of any boundary node meets the asynchronous trigger condition again, a new boundary node feature snapshot is automatically generated to form an adaptive boundary suppression closed loop and iteratively update the state of the two-way information collaboration link.

9. The distribution network operation risk control method based on Bayesian optimization according to claim 1 is characterized in that: The S7 specifically includes: S71. For each risk control subdomain, asynchronously capture the spatiotemporal embedding vectors of all nodes in the subdomain and the corresponding risk control instruction execution feedback states, and generate a distributed adaptive error spectrum through node state transition accumulation, where the distributed adaptive error spectrum represents the error distribution characteristics of the node risk state in the temporal and spatial domains; S72. Compare the distributed adaptive error spectrum with the historical multi-scale error spectrum in parallel, and use the spectrum resonance detection mechanism to compare the spectrum energy level distribution. When the error energy level aggregation across time periods appears in the error spectrum and the spectrum resonance coefficient exceeds the preset stability threshold, identify the high-energy-level error node cluster, thereby determining that the local risk convergence state has degraded. S73. After determining that the local risk convergence state has degraded, trigger the subdomain internal path regeneration mechanism and autonomously query the redundant control path library for multiple redundant path candidate sets corresponding to the high-energy error node group. The redundant control path library is pre-classified and stored according to node spatiotemporal embedding matching degree, path redundancy, and fault tolerance level. S74. For the candidate redundant path set, sequentially calculate the spatiotemporal embedding vector spectrum compatibility index of the nodes in each path, the link oscillation amplitude index of the path relay node, and the node load jump tolerance index, and comprehensively adopt a parallel adaptation priority calculation strategy to select the redundant control path with the highest adaptation priority as the switching target; S75. Based on the information of the starting node, relay node, and terminating node of the selected redundant control path, generate a token command sequence according to the self-organizing token ring protocol. The token command sequence includes control instructions and token transfer rules for each node in the redundant path, and seamlessly switch to the token command sequence in the current control time slot. S76: After switching to the redundant control path, the asynchronous hedging mechanism is started in parallel within the subdomain to collect the convergence rate difference and risk control resonance strength of each node on the redundant path. When the resonance strength is lower than the preset resonance stability threshold, the rapid rollback mechanism is triggered and step S73 is re-entered to loop search and switch to the next redundant control path. S77. When the switched redundant control path reaches a state of convergence and risk resonance equilibrium, the adaptive priority weight of the corresponding path in the redundant control path library is updated through the dynamic memory interaction mechanism, and the asynchronous capture and monitoring of the conventional distributed adaptive error spectrum in the subdomain is restored.

Images