Power distribution network stability control method based on multiple agents

Abstract

The invention discloses a power distribution network stability control method based on multiple agents. The method comprises the following steps of deploying the agents, collecting data and constructing graph data; establishing an improved graph echo state network and forming a structured input sequence; inputting an input sequence to carry out time sequence evolution, and forming a dynamic echo state vector; exchanging dynamic echo state vectors with a neighborhood agent, and splicing to form comprehensive time sequence state information; each agent analyzes the comprehensive time sequence state information to generate a local stability control instruction; when disturbance occurs, disturbance updating comprehensive time sequence state information is constructed, and a disturbance updating local stability control instruction is generated; and periodically and continuously executing dynamic echo state vector updating, comprehensive time sequence state information construction and local stability control instruction generation in a normal operation period. According to the invention, the improved graph echo network is adopted to realize stable control of the power distribution network, and the method has the advantages of strong real-time performance and fast disturbance response.

Description

Technical Field

[0001] This invention relates to the field of power distribution network operation control technology, and in particular to a power distribution network stability control method based on multi-agent systems. Background Technology

[0002] With the high proportion of distributed renewable energy integrated into the distribution network, node voltage fluctuations are frequent and power flow directions change rapidly. Traditional distribution network stability control methods centered on central controllers are gradually revealing limitations in terms of real-time performance and scalability. Existing technologies mostly employ centralized control or distributed control strategies based on local measurements. However, these methods struggle to uniformly characterize the topological relationships between nodes, cannot consistently model the temporal operating states of each node, have insufficient response capabilities to rapid disturbances, and are prone to control bottlenecks under communication delays or local faults.

[0003] Existing research attempts to introduce multi-agent cooperative control concepts and graph-based machine learning methods, but generally lacks a unified modeling framework that can simultaneously reflect power grid topology constraints and node temporal dynamic characteristics. The information exchanged between agents is loosely structured, failing to support stable collaborative decision-making across nodes. Furthermore, existing graph-based models have inadequate state update mechanisms under disturbance scenarios, making it difficult to effectively decouple periodic operation control from disturbance response control. Therefore, there is an urgent need for a distribution network stability control method with distributed, high real-time performance, and strong disturbance adaptability under graph topology constraints.

[0004] Therefore, how to provide a multi-agent-based distribution network stability control method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-agent-based distribution network stability control method. This invention uses an improved graph echo network to achieve distribution network stability control, which has the advantages of strong real-time performance and fast disturbance response.

[0006] A distribution network stability control method based on multi-agent technology according to an embodiment of the present invention includes the following steps:

[0007] Deploy intelligent agents at each operating node of the distribution network to collect data on the voltage amplitude, phase angle, active power, reactive power, and connection relationships between nodes, and construct graph data.

[0008] An improved graph echo state network is established for each agent, and input weights, reservoir weights and state update parameters are set to form a structured input sequence.

[0009] The input sequence is fed into the improved graph echo state network of the corresponding agent, and time-series evolution is performed to obtain the current reservoir state and form a dynamic echo state vector.

[0010] Each agent sends its dynamic echo state vector to its neighboring agents, receives the dynamic echo state vectors sent by its neighboring agents, and splices them together to form a comprehensive temporal state information.

[0011] Each intelligent agent analyzes the voltage, active power, reactive power, and correlation characteristics of the node and its neighboring nodes based on the comprehensive time-series state information, and generates local stability control commands.

[0012] When a disturbance occurs in the distribution network, the disturbance time-series operation data is acquired, the disturbance update dynamic echo state vector is generated, the disturbance update comprehensive time-series state information is constructed, and the disturbance update local stability control command is generated.

[0013] Each agent periodically acquires periodic time-series operational data during its normal operation cycle, and continuously performs dynamic echo state vector updates, constructs comprehensive time-series state information, and generates local stability control commands.

[0014] Optionally, the generation of the graph data includes:

[0015] Each operating node in the distribution network is assigned a unique node index, and all node indices are arranged in a fixed order to form a fixed node index sequence.

[0016] Based on a fixed node index sequence, the connection relationship between any two nodes is mapped to the adjacency matrix according to the correspondence of node indices;

[0017] Based on the fixed node index sequence, all node pairs with line connections are recorded as edge sets in node index order;

[0018] The fixed node index sequence, adjacency matrix, and edge set are organized in a unified manner, and the normalized voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence of each node are associated with the corresponding node index to form graph data.

[0019] Optionally, the generation of the structured input sequence includes:

[0020] The voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence are normalized.

[0021] Based on the fixed node index sequence in the graph data, the normalized voltage amplitude sequence, phase angle sequence, active power sequence and reactive power sequence of each node are arranged in the order of the fixed node index to form the time sequence operation input structure of the corresponding node.

[0022] Based on a fixed node index sequence, a fixed mapping relationship is established between the input weights and the temporal execution amount of the nodes according to the node index order;

[0023] The weights of the reserve pool are constructed based on the adjacency matrix, so that any position in the reserve pool is connected to the neighboring nodes of its corresponding node by weight only.

[0024] The state update parameters are set according to the preset internal attenuation coefficient, input mapping coefficient, normalization coefficient and maximum change limit coefficient, so that the state update process performs temporal evolution under topological constraints within the reservoir.

[0025] The voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence arranged in chronological order are sequentially input into the improved graph echo state network of the corresponding intelligent agent according to the node index order, forming the structured input sequence of the improved graph echo state network.

[0026] Optionally, the generation of the dynamic echo state vector includes:

[0027] The structured input sequence is input into the improved graph echo state network of the corresponding agent according to the fixed node index order;

[0028] Based on the input weights, the structured input sequence at the current time is mapped to the corresponding position in the reserve pool in a fixed node index order, so that the corresponding position obtains the input driving quantity at the current time;

[0029] Based on the reserve pool weight, the previous reserve pool state is decayed according to the internal decay coefficient. Then, based on the connection relationship represented by the adjacency matrix, the decayed reserve pool state is superimposed with the previous reserve pool states of all neighboring nodes that have a connection relationship with it according to the connection weight, to form the intermediate state of the current time.

[0030] Based on the state update parameters, the input driving quantity and the intermediate state are summed in a fixed ratio to obtain the final state of the reservoir at the current time. The final state is then arranged in a fixed node index order to form a dynamic echo state vector.

[0031] Optionally, the generation of the integrated time-series state information includes:

[0032] After each agent completes the generation of the dynamic echo state vector at the corresponding node, it organizes the dynamic echo state vector of its node obtained at the current time and sends the organized dynamic echo state vector of its node to all neighboring agents with which it has a connection relationship.

[0033] Each agent receives the neighborhood dynamic echo state vector generated at the current time from all neighboring agents with which it has a connection relationship, and performs preliminary collection of the received neighborhood dynamic echo state vectors. According to the fixed node index order, the preliminary collection of neighborhood dynamic echo state vectors is uniformly arranged according to the pre-set index order.

[0034] Each agent performs a structure check on the arranged neighborhood dynamic echo state vectors to ensure that each neighborhood dynamic echo state vector is consistent with the dynamic echo state vector of its own node in terms of length, order, and node order identifier.

[0035] Each agent concatenates its own node's dynamic echo state vector with all the neighboring dynamic echo state vectors after they have been arranged and checked, according to a fixed combination format.

[0036] Each agent performs consistency maintenance on the spliced ​​data structure, sets the consolidated spliced ​​structure as the comprehensive time-series state information of the current time, and stores it in the internal time-series information cache.

[0037] Optionally, the generation of the local stability control command includes:

[0038] Each agent acquires the comprehensive temporal state information at the current time, and performs structural analysis on the local node's dynamic echo state vector and the neighborhood dynamic echo state vector in the comprehensive temporal state information to obtain the comprehensive temporal state information after structural analysis.

[0039] Based on the comprehensive time-series state information after structural analysis, each agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node correlation change characteristics of its current node at the current time, and combines the change characteristics in a fixed order to form the time-series operation change set of its current node.

[0040] Based on the comprehensive temporal state information after structural analysis, each intelligent agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of all neighboring nodes at the current time, and arranges the change characteristics into a neighborhood temporal operation change set according to a fixed node index order;

[0041] Each agent merges its own temporal operation change set with the neighborhood temporal operation change set according to a fixed combination format, so that the merged temporal operation change structure can simultaneously reflect the local topological structure relationship and synchronous temporal operation state of the local node and the neighborhood nodes.

[0042] Each agent generates node-level adjustment quantities based on the fused temporal operational change structure;

[0043] Each agent sets the node-level adjustment to the local stability control command for the current time and executes the local stability control command through the execution device corresponding to its node.

[0044] Optionally, the generation of the disturbance update local stability control command includes:

[0045] When each intelligent agent detects fluctuations in new energy output, load changes, or line disturbances, it acquires the disturbance time-series operation data of its own node and organizes the disturbance time-series operation data into a disturbance update input sequence according to the fixed node index order.

[0046] The perturbation update input sequence is fed into the improved graph echo state network to form the perturbation update reservoir state;

[0047] The state of the disturbance update reserve pool is arranged in the order of fixed node index to form the dynamic echo state vector of the disturbance update of this node, and sent to all neighboring agents that have a connection with it. At the same time, the dynamic echo state vector of the neighborhood disturbance update generated by all neighboring nodes at the same disturbance time is received.

[0048] All neighborhood perturbations are updated to the dynamic echo state vector and arranged in a fixed node index order.

[0049] The dynamic echo state vector of the disturbance update of this node is concatenated with the arranged dynamic echo state vector of the disturbance update of the neighboring area according to a fixed combination format to form the comprehensive temporal state information of the disturbance update.

[0050] Based on the comprehensive time-series state information of the disturbance update, the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of the local node and neighboring nodes at the time of the disturbance are extracted, and the change characteristics are organized into a disturbance update operation quantity set.

[0051] Node-level adjustment quantities are generated based on the set of disturbance update operation quantities. These node-level adjustment quantities are then used as disturbance update local stability control commands, which are executed by the corresponding execution devices.

[0052] Optionally, the generation of the local stability control command includes:

[0053] When the power distribution network is in normal operation, each intelligent agent periodically acquires the periodic time-series operation data of its own node and organizes it into a periodic update input sequence according to a fixed node index order;

[0054] Each agent inputs the periodic update input sequence into the improved graph echo state network and forms a periodic update reserve pool state. The periodic update reserve pool states are arranged in order of fixed node index to form the periodic update dynamic echo state vector of this node.

[0055] Each agent sends its periodically updated dynamic echo state vector to all neighboring agents with which it has a connection, and receives the periodically updated dynamic echo state vectors from all neighboring nodes.

[0056] Each agent arranges all neighborhood periodically updated dynamic echo state vectors in a fixed node index order, and then concatenates the arranged neighborhood periodically updated dynamic echo state vectors with the current node's periodically updated dynamic echo state vectors in a fixed combination format.

[0057] Each agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of its own node and neighboring nodes based on the periodically updated comprehensive time-series state information, and organizes the extracted change characteristics into a periodically updated set of operational quantities;

[0058] Each intelligent agent generates a periodically updated node-level adjustment quantity based on the periodically updated operational quantity set, uses the periodically updated node-level adjustment quantity as a periodically updated local stability control instruction, and executes the periodically updated local stability control instruction through the corresponding execution device.

[0059] After each agent completes the execution of the periodic update local stability control command, it distinguishes the periodic update input sequence, the periodic update reserve pool state, the periodic update dynamic echo state vector, the periodic update comprehensive time-series state information, and the corresponding content generated during the periodic update node-level adjustment and disturbance update process. This ensures that the periodic update control chain and the disturbance update control chain remain structurally independent and can work together in a dual-track manner to maintain overall stable operation during the operation of the distribution network.

[0060] The beneficial effects of this invention are:

[0061] This invention deploys intelligent agents at each node of the distribution network and uses graph data constructed from fixed node indices, adjacency matrices, and edge sets as a foundation. This establishes a strict one-to-one correspondence between the timing-series operational quantities of nodes, such as voltage, phase angle, active power, and reactive power, and the distribution network topology, achieving a structured power grid modeling effect that traditional techniques cannot reach. Building upon this, the invention introduces an improved graph echo state network, directly embedding topological constraints into the reserve pool weights. This ensures that the state evolution within the reserve pool strictly follows the physical connections of the power grid, forming a graph-structured timing evolution chain of node timing states. This continuously captures the dynamic coupling relationships between nodes, significantly enhancing the timing representation capability of complex electrical quantity changes. Through the exchange of dynamic echo state vectors between intelligent agents, this invention constructs comprehensive timing state information containing collaborative information between the current node and neighboring nodes. This enables each intelligent agent to perceive voltage and power flow trends within a region locally, thus avoiding the shortcomings of traditional centralized control structures, such as reliance on a central node, high communication pressure, and susceptibility to single-point failures.

[0062] Furthermore, this invention establishes a dual-track operation mechanism of periodic update control chain and disturbance update control chain, enabling the intelligent agent to continuously update the grid operating state during normal operation cycles. When fluctuations in renewable energy output, sudden load changes, or line disturbances occur, it can rapidly generate disturbance update dynamic echo state vectors and comprehensive temporal state information locally, achieving disturbance response at the second or even millisecond level. This mechanism solves the problems of high coupling between periodic control and disturbance control, slow disturbance response, and insufficient dynamic robustness in existing methods. Through structured analysis, feature extraction, and fusion of comprehensive temporal state information, this invention can generate local stability control commands with physical meaning for each node that can directly affect voltage, active power, reactive power, and power flow regulation. This ensures that the control strategy possesses both distributed autonomy and consistency and coordination across the entire network. In summary, this invention significantly improves the dynamic stability, disturbance recovery speed, distributed coordination capability, and real-time online adjustment capability of distribution networks under high-proportion renewable energy access conditions, possessing significant engineering application value and technological advancement significance for solving stability control problems in complex distribution network operation scenarios. Attached Figure Description

[0063] 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:

[0064] Figure 1 This is a flowchart of the overall process of a multi-agent-based distribution network stability control method proposed in this invention;

[0065] Figure 2 This is a schematic diagram of the structure of an improved graphical echo state network for a multi-agent-based distribution network stability control method proposed in this invention. Detailed Implementation

[0066] 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.

[0067] refer to Figure 1 and Figure 2 A multi-agent-based distribution network stability control method includes the following steps:

[0068] Intelligent agents are deployed at each operating node of the distribution network to collect the voltage amplitude, phase angle, active power, reactive power and connection relationship between nodes, and to construct graph data based on the connection relationship. The graph data includes fixed node index, adjacency matrix and edge set, so that the node operation volume and topology form a structured correspondence.

[0069] An improved graph echo state network is established for each agent based on the graph data. Input weights, reservoir weights, and state update parameters corresponding to the graph data are set. The reservoir weights are constructed based on the adjacency matrix, so that any position in the reservoir is only connected to the neighboring nodes of its corresponding node. The voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence of the node are arranged in time order and then input into the improved graph echo state network of the corresponding agent to form a structured input sequence of the improved graph echo state network.

[0070] The input sequence is fed into the improved graph echo state network of the corresponding agent. The state of the reservoir is evolved in time according to the topological connection relationship and update rule represented by the reservoir. The current reservoir state is obtained by the decay of the state in the previous time, the superposition of the states of adjacent nodes according to the connection weight, and the input action. The state is then arranged in the order of the nodes to form a dynamic echo state vector.

[0071] Each agent sends its dynamic echo state vector to its neighboring agents and receives the dynamic echo state vectors sent by its neighboring agents. The agents arrange the vectors according to the node index and then concatenate the dynamic echo state vectors of the current node and its neighboring nodes in a fixed format to form a comprehensive temporal state information.

[0072] Each intelligent agent analyzes the voltage, active power, reactive power, and correlation characteristics of the node and its neighboring nodes based on the comprehensive time-series state information, generates local stability control commands to adjust the operating state of its own node, and makes real-time adjustments through the corresponding actuators.

[0073] When there are fluctuations in new energy output, load changes, or line disturbances in the distribution network, each intelligent agent acquires new time-series operating data, updates the improved graph echo state network, generates a disturbance update dynamic echo state vector, constructs comprehensive time-series state information of disturbance update, and generates disturbance update local stability control commands to achieve rapid disturbance response.

[0074] Each agent periodically acquires new time-series operational data during normal operation cycles, continuously executes dynamic echo state vector updates, constructs comprehensive time-series state information, and generates local stability control commands. This enables the agents to maintain the distributed and stable operation of the distribution network based on a dual-track mechanism of periodic update control and disturbance update control without the need for a central controller.

[0075] In this embodiment, the generation of the graph data includes:

[0076] Each operating node in the distribution network is assigned a unique node index, and all node indices are arranged in a fixed order to form a fixed node index sequence.

[0077] Based on the fixed node index sequence, the connection relationship between any two nodes is mapped to the adjacency matrix according to the correspondence of the node indices, so that the row and column numbers correspond one-to-one with the fixed node index sequence;

[0078] Based on the fixed node index sequence, all node pairs with line connections are recorded as edge sets in node index order, so that each edge in the edge set is identified by the node number in the fixed node index sequence.

[0079] The fixed node index sequence, adjacency matrix, and edge set are organized in a unified manner, and the normalized voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence of each node are associated with the corresponding node index, so that the time-series operation data of the nodes and the topology form a one-to-one correspondence at the data level, thus forming graph data for describing the distribution network structure.

[0080] In this embodiment, the generation of the structured input sequence includes:

[0081] The voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence are normalized so that all time-series operating quantities are mapped to a unified range in terms of numerical value.

[0082] Based on the fixed node index sequence in the graph data, the normalized voltage amplitude sequence, phase angle sequence, active power sequence and reactive power sequence of each node are arranged in the order of the fixed node index to form the time sequence operation input structure of the corresponding node.

[0083] The generation of the time-series operational quantity input structure specifically includes: organizing the voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence collected by each node in continuous time according to time order, so that each type of time-series operational quantity is arranged consistently in the time dimension; according to the fixed node index order, matching the organized types of time-series operational quantities with the corresponding nodes one-to-one, so that the time-series operational quantity of each node occupies a stable position in the structure; combining the voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence of each node according to a fixed combination format to form a node-level time-series operational quantity structure; and arranging the node-level time-series operational quantity structures of all nodes in the fixed node index order to form the overall time-series operational quantity input structure.

[0084] Based on a fixed node index sequence, a fixed mapping relationship is established between the input weights and the time series execution quantities of the nodes according to the node index order, so that each position of the input weight corresponds one-to-one with the time series execution quantity of the corresponding node.

[0085] The weights of the reserve pool are constructed based on the adjacency matrix, so that any position in the reserve pool is connected to the neighboring nodes of its corresponding node by weight, and the connection relationship of all reserve pool weights is consistent with the structure of the adjacency matrix.

[0086] The state update parameters are set according to the preset internal attenuation coefficient, input mapping coefficient, normalization coefficient and maximum change limit coefficient, so that the state update process performs temporal evolution under topological constraints within the reservoir.

[0087] The temporal evolution specifically includes: attenuating the previous time's reservoir state according to an internal attenuation coefficient, so that the previous time's state forms the attenuated base state at the current time; based on the node connection relationships represented by the reservoir weights, weighting and superimposing the attenuated base state with the previous time's reservoir states of all neighboring nodes connected to it, so that the weighted superposition result reflects the propagation influence of nodes under graph topological constraints; on this basis, combining the input action and the weighted superposition result according to a fixed ratio, so that the combined state forms the current time's reservoir state; and using the current time's reservoir state as the basic input for the subsequent generation of dynamic echo state vectors, so that the reservoir state forms a temporal evolution process with graph structure constraints over continuous time.

[0088] The voltage amplitude sequence, phase angle sequence, active power sequence, and reactive power sequence arranged in chronological order are sequentially input into the improved graph echo state network of the corresponding intelligent agent according to the node index order, so that the input sequence establishes a structured correspondence with the reservoir structure, forming the structured input sequence of the improved graph echo state network.

[0089] In this embodiment, the generation of the dynamic echo state vector includes:

[0090] The structured input sequence is input into the improved graph echo state network of the corresponding agent according to the fixed node index order, so that each time-series operation quantity establishes a fixed correspondence with the position of the corresponding node in the reserve pool;

[0091] Based on the input weights, the structured input sequence at the current time is mapped to the corresponding position in the reserve pool in a fixed node index order, so that the corresponding position obtains the input driving quantity at the current time;

[0092] The generation of input driving quantities specifically includes: organizing the voltage amplitude, phase angle, active power, and reactive power collected by the node at the current time according to a fixed node index order, so that various time-series operating quantities of the same node occupy fixed positions in the structure; according to the input weights set in the improved graphical echo state network, proportionally applying the organized various time-series operating quantities to their corresponding input weights, so that the voltage amplitude becomes the voltage input component, the phase angle becomes the phase angle input component, the active power becomes the active power input component, and the reactive power becomes the reactive power input component; combining all input components according to a fixed combination format, so that the combined structure simultaneously contains the driving information of various time-series operating quantities of the current time on the state of the reserve pool at the data level; and setting the combined structure as the input driving quantity.

[0093] Based on the reserve pool weight, the previous reserve pool state is decayed according to the internal decay coefficient. Then, based on the connection relationship represented by the adjacency matrix, the decayed reserve pool state is superimposed with the previous reserve pool states of all neighboring nodes that have a connection relationship with it according to the connection weight, to form the intermediate state of the current time.

[0094] Based on the state update parameters, the input driving quantity and the intermediate state are summed in a fixed ratio to obtain the final state of the reservoir at the current time. The final state is then arranged in a fixed node index order to form a dynamic echo state vector that characterizes the temporal operation characteristics of the node under topological constraints.

[0095] In this embodiment, the generation of the integrated time-series state information includes:

[0096] After each agent generates the dynamic echo state vector at its corresponding node, it restructures the dynamic echo state vector of its node obtained at the current time, so that it maintains the same arrangement as the fixed node index order internally. Each agent sends the restructured dynamic echo state vector of its node to all neighboring agents that have a connection with it, so that the neighboring agents can receive the dynamic echo state vector of their node at the current time.

[0097] Each agent receives the neighborhood dynamic echo state vector generated at the current time from all neighboring agents with which it has a connection relationship, and performs preliminary collection of the received neighborhood dynamic echo state vectors. According to the fixed node index order, the preliminary collection of neighborhood dynamic echo state vectors is uniformly arranged according to the pre-set index order, so that the arranged neighborhood dynamic echo state vectors maintain a consistent index correspondence in structure.

[0098] Each agent performs a structure check on the arranged neighborhood dynamic echo state vectors to ensure that each neighborhood dynamic echo state vector is consistent with the dynamic echo state vector of its own node in terms of length, order and node order identifier, so as to ensure that they can be combined according to a unified structure during splicing.

[0099] Each agent concatenates its own dynamic echo state vector with the arranged and checked dynamic echo state vectors of all neighboring nodes according to a fixed combination format, so that the concatenated result contains, in a single data structure, the dynamic echo state vectors generated by the current node and all neighboring nodes in the current time in sequence.

[0100] Each agent performs consistency maintenance on the concatenated data structure, ensuring that the concatenated structure remains consistent with the concatenated structure of the previous time step in terms of length, corresponding node positions, sorting rules, and overall format. This ensures that the structure remains stable when the comprehensive temporal state information is used continuously at different times. The consolidated concatenated structure is then set as the comprehensive temporal state information for the current time, enabling the comprehensive temporal state information to simultaneously reflect the temporal operation characteristics and topological relationships of the current node and its neighboring nodes, and is stored in the internal temporal information cache.

[0101] In this embodiment, the generation of the local stability control command includes:

[0102] Each agent acquires the comprehensive temporal state information at the current time and performs structural analysis on the dynamic echo state vector of its own node and the dynamic echo state vector of its neighborhood in the comprehensive temporal state information. This allows each temporal operation quantity to be mapped to its respective node position according to the fixed node index order after analysis, thus obtaining the comprehensive temporal state information after structural analysis.

[0103] The generation of the integrated temporal state information after structural analysis specifically includes: identifying the positions of the local node's dynamic echo state vector and all neighboring dynamic echo state vectors in the integrated temporal state information according to a fixed node index order, thus determining the position of each dynamic echo state vector in the overall structure; extracting the local node's dynamic echo state vector from the integrated temporal state information and sequentially organizing the various temporal execution quantities contained within it, ensuring that the structured execution quantities of the local node maintain a fixed format in structure; extracting all neighboring dynamic echo state vectors one by one according to the fixed node index order and structurally organizing the temporal execution quantities within them, ensuring that the organized neighboring structured execution quantities maintain the same arrangement as the local node's structured execution quantities; placing the local node's structured execution quantities at the beginning of the combined structure and arranging all neighboring structured execution quantities according to the fixed node index order; and combining the arranged structured execution quantities according to a fixed combination format to form the integrated temporal state information after structural analysis.

[0104] Based on the comprehensive time-series state information after structural analysis, each agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node correlation change characteristics of its current node at the current time, and combines the change characteristics in a fixed order to form the time-series operation change set of its current node.

[0105] Based on the comprehensive temporal state information after structural analysis, each intelligent agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of all neighboring nodes at the current time, and arranges the change characteristics into a neighborhood temporal operation change set according to a fixed node index order;

[0106] Each agent merges its own temporal operation change set with the neighborhood temporal operation change set according to a fixed combination format, so that the merged temporal operation change structure can simultaneously reflect the local topological structure relationship and synchronous temporal operation state of the local node and the neighborhood nodes.

[0107] The generation of the fused time-series operational change structure specifically includes: using the time-series operational change set of the current node as the first set of data in the fused structure, giving it a fixed position in the overall structure; adding the time-series operational change sets of all neighboring nodes one by one according to the fixed node index order, so that the time-series operational change sets of each neighboring node are arranged sequentially in the fused structure according to a preset order; combining the voltage amplitude change, active power change, reactive power change and node association change in the time-series operational change set of the current node according to a fixed format, and combining the corresponding change features in the time-series operational change sets of all neighboring nodes according to the same format, so that the change features of the current node and neighboring nodes are arranged in a consistent manner in terms of structure; structurally splicing the above-combined change features according to a fixed combination format, so that the spliced ​​fused structure simultaneously contains the time-series operational change features of the current node and all neighboring nodes at the current time; performing a unified reorganization on the spliced ​​fused structure, so that it is consistent with the fused structure of the previous time in terms of node order, length, field arrangement and overall format, and setting the reorganized structure as the fused time-series operational change structure;

[0108] Each intelligent agent generates node-level adjustment quantities based on the fused temporal operation change structure, so that the node-level adjustment quantities can respectively correspond to the adjustment actions of the voltage, active power, reactive power or power flow distribution of the node.

[0109] Each agent sets the node-level adjustment amount as the local stability control command for the current time, and executes the local stability control command through the execution device corresponding to its node, so that the node can achieve stable adjustment based on the graph topology-time sequence joint operation characteristics.

[0110] In this embodiment, the generation of the disturbance update local stability control command includes:

[0111] When each intelligent agent detects fluctuations in new energy output, load changes, or line disturbances, it acquires the disturbance time-series operation data of its node and organizes the disturbance time-series operation data into a disturbance update input sequence according to a fixed node index order, so that the disturbance update input sequence is structurally consistent with the structured input sequence of the previous time.

[0112] The perturbation update input sequence is input into the improved graph echo state network, so that the reservoir performs a reservoir state update process independent of the periodic update based on the perturbation time series operation data, and forms the perturbation updated reservoir state.

[0113] The state of the disturbance update reserve pool is arranged in the order of fixed node index to form the dynamic echo state vector of the disturbance update of this node, and sent to all neighboring agents that have a connection with it. At the same time, the dynamic echo state vector of the neighborhood disturbance update generated by all neighboring nodes at the same disturbance time is received.

[0114] Arrange all neighborhood perturbation update dynamic echo state vectors according to a fixed node index order, so that the arranged neighborhood perturbation update dynamic echo state vectors can maintain a consistent structural format with the perturbation update dynamic echo state vector of the current node.

[0115] The dynamic echo state vector of the disturbance update of this node is concatenated with the arranged dynamic echo state vector of the disturbance update of the neighboring area according to a fixed combination format to form the comprehensive temporal state information of the disturbance update.

[0116] Based on the comprehensive time-series state information of the disturbance update, the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of the local node and neighboring nodes at the time of the disturbance are extracted, and the change characteristics are organized into a disturbance update operation quantity set.

[0117] Based on the set of disturbance update operation quantities, node-level adjustment quantities for disturbance scenarios are generated, enabling the node-level adjustment quantities to be used to adjust the voltage, active power, reactive power, or power flow distribution of the node. The node-level adjustment quantities are used as disturbance update local stability control commands, and the disturbance update local stability control commands are executed by the corresponding execution devices, so that the node can complete rapid adjustment based on graph topology and time-series operation change characteristics at the time of disturbance.

[0118] In this embodiment, the generation of the local stability control command includes:

[0119] When the power distribution network is in normal operation, each intelligent agent periodically acquires the periodic time-series operation data of its own node and organizes it into a periodic update input sequence according to a fixed node index order;

[0120] Each agent inputs the periodic update input sequence into the improved graph echo state network, so that the reserve pool performs the periodic update reserve pool state generation process according to the periodic update input sequence and forms the periodic update reserve pool state. The periodic update reserve pool state is arranged according to the fixed node index order to form the periodic update dynamic echo state vector of the node.

[0121] Each agent sends its periodically updated dynamic echo state vector to all neighboring agents with which it has a connection, and receives the periodically updated dynamic echo state vectors from all neighboring nodes.

[0122] Each agent arranges all neighborhood periodically updated dynamic echo state vectors according to a fixed node index order, and concatenates the arranged neighborhood periodically updated dynamic echo state vectors with the current node's periodically updated dynamic echo state vector according to a fixed combination format, so that the concatenated data forms periodically updated comprehensive temporal state information.

[0123] Each agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics, and node association relationship change characteristics of its own node and neighboring nodes based on the periodically updated comprehensive time-series state information, and organizes the extracted change characteristics into a periodically updated set of operational quantities;

[0124] Each intelligent agent generates a periodically updated node-level adjustment quantity based on the periodically updated operational quantity set, so that the periodically updated node-level adjustment quantity can be used to adjust the voltage, active power, reactive power or power flow distribution of this node. The periodically updated node-level adjustment quantity is used as a periodically updated local stability control command, and the periodically updated local stability control command is executed through the corresponding execution device, so that this node can complete continuous stable adjustment according to the periodically updated comprehensive timing state information during the normal operation cycle.

[0125] After each agent completes the execution of the periodic update local stability control command, it distinguishes the periodic update input sequence, the periodic update reserve pool state, the periodic update dynamic echo state vector, the periodic update comprehensive time-series state information, and the corresponding content generated during the periodic update node-level adjustment and disturbance update process. This ensures that the periodic update control chain and the disturbance update control chain remain structurally independent and can work together in a dual-track manner to maintain overall stable operation during the operation of the distribution network.

[0126] Example 1:

[0127] To verify the feasibility of this invention in practice, it was applied to a power supply area surrounding a city in a southeastern coastal region. The distribution network in this area covers coastal residential areas, sea breeze corridors, and several distributed renewable energy access points, including rooftop photovoltaics, commercial and industrial photovoltaic arrays, small-scale wind turbines, and some energy storage systems. Due to the open terrain and rapid climate change, the distribution network in this area has consistently faced operational challenges such as large voltage fluctuations, frequent power flow reversals, and rapid changes in node load. Particularly during the summer months with strong sunlight and sudden changes in sea breezes, the output fluctuations of photovoltaics and wind power are extremely pronounced, making the regulation and stability of the distribution network a long-standing problem for local maintenance personnel.

[0128] In this power distribution network, the intelligent agent collaborative control system proposed in this invention is introduced. Intelligent agent devices are deployed at each operating node, capable of independently collecting the voltage amplitude, phase angle, active power, and reactive power of that node. The intelligent agents organize this data and the connection relationships between nodes into graph data, clearly expressing the topology of the entire power distribution network at the data level using fixed node indices, adjacency matrices, and edge sets. Compared to traditional control methods that rely on single measurement data, this invention enables each node to obtain a structured representation consistent with the entire network within a digital structure, avoiding judgment biases caused by inconsistent measurement sources.

[0129] After the agent is deployed, an improved graph echo state network is constructed based on graph data, enabling the time-series operational quantities of each node to propagate within the storage pool in a topologically constrained manner. In practical applications, when the solar radiation in this distribution network rapidly increases in the morning, the photovoltaic output begins to rise, and some nodes exhibit a voltage rise trend. The agent inputs the continuously collected time-series operational quantities into the improved graph echo state network, generating a dynamic echo state vector through the time-series evolution process of the storage pool. Since the weights of the storage pool are constructed entirely based on adjacency relationships, the dynamic echo state vector not only reflects the time-series characteristics of the node itself but also implicitly contains structured information about the influence of neighboring nodes, thus effectively overcoming the deficiency of traditional local control in obtaining regional dynamic characteristics.

[0130] To achieve coordinated regulation among nodes, each agent sends its own generated dynamic echo state vector to all connected neighboring nodes, while simultaneously receiving the dynamic echo state vectors from its neighbors. The received information is arranged in a fixed node index order and formatted for concatenation to form comprehensive time-series state information. Because this structure strictly adheres to a unified format, agents can directly identify the operational trends of their own node and neighboring nodes at the same time without complex data alignment. During periods of rapid photovoltaic changes, agents can determine the direction of voltage trend changes, the transfer trend of active power flow, and the synchronous changes in reactive power by integrating the comprehensive time-series state information, thereby generating local stability control commands that better suit the current state.

[0131] In this region, strong sea breezes often occur in the afternoon, causing significant fluctuations in wind power output and easily leading to simultaneous disturbances at multiple nodes. To address this issue, the disturbance update mechanism of this invention enables the agent to immediately collect new time-series operational data upon detecting a disturbance, rapidly generating a new disturbance update dynamic echo state vector. This vector is then concatenated with the corresponding vectors from neighboring nodes to form comprehensive disturbance update time-series state information. Based on this information, the agent generates disturbance update local stability control commands, ensuring stable operation of nodes during the duration of the disturbance. Under this mechanism, voltage fluctuations are rapidly suppressed, power flow impacts are significantly mitigated, the overall distribution network becomes more stable, and cascading fluctuations caused by disturbance propagation no longer occur.

[0132] During stable operation of the distribution network, the intelligent agent periodically acquires new time-series operational data for its node and executes a periodic update mechanism. At different stages—morning load increases, midday photovoltaic boosts, evening photovoltaic declines, and nighttime energy storage charging and discharging—the agent's periodically updated comprehensive time-series state information continuously reflects the grid's operational trends, enabling the control strategy to evolve and adapt over time. Compared to traditional fixed-parameter control methods, this invention allows each node to autonomously adjust its operating strategy based on periodic data, ensuring the overall distribution network maintains higher sensitivity and adaptability over long-term operation.

[0133] Long-term operational observations in this region show that, after implementing this invention, the distribution network can maintain a more stable operating state under the influence of various variables such as weather changes, load fluctuations, and unstable output from photovoltaic and wind power. The intelligent agent, leveraging the temporal information captured by the improved graph echo state network and the comprehensive temporal state information efficiently exchanged with neighboring nodes, enables control behavior to not only possess local autonomy but also demonstrate cross-node collaborative consistency. Whether in rapid response during disturbance phases or continuous adjustment during cyclical operation, this invention ensures the entire distribution network maintains reliable and stable operating characteristics under different operating conditions, fully demonstrating its practical value and potential for widespread application in complex distribution network scenarios.

[0134] Table 1. Performance comparison test data between the method of this invention and the traditional distributed control method.

[0135] Comparison indicators Traditional distributed control Traditional robust control Improved Graph ESN Method Comparison of improvement magnitude (traditional distributed control) Comparison of improvement magnitude (traditional robust control) Average voltage deviation (%) 3.8 3.2 1.4 63.2% 56.3% Number of voltage exceedances (times / day) 14 10 3 78.6% 70.0% Trend reversal recovery time (seconds) 4.6 4.2 1.9 58.7% 54.8% Node synchronization deviation under disturbance (ms) 29 24 11 62.1% 54.2% Control command stability score (0-1) 0.63 0.71 0.89 +41.3% +25.4%

[0136] As shown in Table 1, the method of this invention significantly outperforms traditional distributed control and robust control methods in several key performance indicators. Regarding voltage deviation, this invention reduces the average deviation from 3.8% in traditional distributed control to 1.4%, a reduction of over 60%, effectively avoiding voltage instability caused by renewable energy fluctuations. In terms of voltage limit violations, the method of this invention, by introducing a neighborhood dynamic echo state vector splicing mechanism, enables the agent to more accurately identify regional voltage trend changes, thereby reducing the number of limit violations from 14 to 3, significantly improving power supply quality. In power flow reversal scenarios, traditional methods generally rely on static parameters or single-node data, resulting in slow recovery speeds. This invention, through the time-series evolution of the reserve pool under dynamic topology constraints, significantly accelerates inter-node state coordination, shortening the recovery time to less than half that of traditional distributed control. Regarding disturbance synchronization, this invention significantly reduces node deviation, comprehensively enhancing the multi-node collaborative capability. Comprehensive analysis reveals that the main reasons for the performance improvement include: the graph-based reservoir weight construction avoids the shortcomings of traditional methods in capturing regional correlations; the dynamic echo state vector enables the agent to have adaptive capabilities in the temporal dimension; and the composition of comprehensive temporal state information makes the generation of control commands no longer dependent on single-node inference, but based on the cooperative evolution of local topology, which fundamentally improves the overall stability.

[0137] 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 multi-agent based power distribution network stability control method, characterized in that, The method comprises the following steps: Deploying intelligent agents at each operating node of the power distribution network, collecting voltage amplitude, phase angle, active power, reactive power of the node and connection relationship between nodes, and constructing graph data; Establishing an improved graph echo state network for each intelligent agent, setting input weight, reservoir pool weight and state update parameters, and forming a structured input sequence; Inputting the input sequence into the improved graph echo state network of the corresponding intelligent agent, performing time evolution, obtaining the current reservoir pool state, and forming a dynamic echo state vector; Each intelligent agent sends the dynamic echo state vector to the neighboring intelligent agents and receives the dynamic echo state vector sent by the neighboring intelligent agents to splice and form comprehensive time sequence state information; Each intelligent agent analyzes the voltage, active power, reactive power and correlation relationship change characteristics of the node and the neighboring nodes according to the comprehensive time sequence state information, and generates a local stable control instruction; When a disturbance occurs in the power distribution network, the time sequence operation data of the disturbance is obtained, the disturbance update dynamic echo state vector is generated, the disturbance update comprehensive time sequence state information is constructed, and the disturbance update local stable control instruction is generated; Each intelligent agent periodically obtains the periodic time sequence operation data in the normal operation period and continuously performs dynamic echo state vector update, comprehensive time sequence state information construction and local stable control instruction generation.

2. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the graph data comprises: Assigning a unique node index to each operating node of the power distribution network, and arranging all node indexes in a fixed order to form a fixed node index sequence; According to the fixed node index sequence, the connection relationship between any two nodes is mapped to an adjacency relationship matrix according to the corresponding relationship of the node index; According to the fixed node index sequence, all node pairs connected by lines are recorded in the order of node index as an edge set; Organize the fixed node index sequence, the adjacency relationship matrix and the edge set, and associate the voltage amplitude sequence, the phase angle sequence, the active power sequence and the reactive power sequence of each node with the corresponding node index to form the graph data.

3. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the structured input sequence comprises: Normalizing the voltage amplitude sequence, the phase angle sequence, the active power sequence and the reactive power sequence; According to the fixed node index sequence in the graph data, the normalized voltage amplitude sequence, phase angle sequence, active power sequence and reactive power sequence of each node are arranged in the order of fixed node index to form the time sequence operation input structure of the corresponding node; According to the fixed node index sequence, the input weight is mapped to the time sequence operation of the node in the order of node index; According to the adjacency relationship matrix, the reservoir pool weight is constructed so that any position in the reservoir pool is only connected to the neighboring nodes of the corresponding node through weight; According to the preset internal decay coefficient, input mapping coefficient, normalization coefficient and maximum change limit coefficient, the state update parameters are set to enable the state update process to perform time evolution under topological constraints within the reservoir pool. The voltage amplitude sequence, the phase angle sequence, the active power sequence and the reactive power sequence arranged in time sequence and normalized are sequentially input to the improved graph echo state network of the corresponding agent in the order of node index to form a structured input sequence of the improved graph echo state network.

4. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the dynamic echo state vector comprises: The structured input sequence is input to the improved graph echo state network of the corresponding agent in the order of fixed node index; According to the input weight, the structured input sequence at the current time is mapped to the corresponding position in the reservoir in the order of fixed node index, so that the corresponding position obtains an input driving quantity at the current time; According to the reservoir weight, the reservoir state at the last time is attenuated according to an internal attenuation coefficient, and then the attenuated reservoir state and the reservoir state at the last time of all the neighboring nodes having a connection relationship with the attenuated reservoir state are superimposed according to the connection relationship represented by the adjacency relationship matrix to form an intermediate state at the current time; According to the state update parameter, the input driving quantity and the intermediate state are added in a fixed proportion to obtain the final state of the reservoir at the current time, and the final state is arranged in the order of fixed node index to form a dynamic echo state vector.

5. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the comprehensive time sequence state information comprises: After each agent completes the generation of the dynamic echo state vector at the corresponding node, the dynamic echo state vector of the node at the current time is structurally arranged, and each agent sends the arranged dynamic echo state vector of the node to all the neighboring agents having a connection relationship with the agent; Each agent receives the dynamic echo state vectors of the neighboring nodes generated at the current time from all the neighboring agents having a connection relationship with the agent, and preliminarily collects the received dynamic echo state vectors of the neighboring nodes, arranges the preliminarily collected dynamic echo state vectors of the neighboring nodes in a predetermined index order according to the fixed node index order, and checks the arranged dynamic echo state vectors of the neighboring nodes to make the dynamic echo state vectors of the neighboring nodes consistent with the dynamic echo state vector of the node in length, order and node order identification. Each agent arranges the dynamic echo state vector of the node and all the dynamic echo state vectors of the neighboring nodes arranged and checked in a fixed combination format. Each agent arranges the data structure after splicing to be consistent, sets the arranged splicing structure as the comprehensive time sequence state information at the current time, and stores the comprehensive time sequence state information in the time sequence information buffer. The generation of the local stability control instruction comprises:

6. The multi-agent based power distribution network stability control method of claim 1, wherein, Each agent acquires the comprehensive time sequence state information at the current time, and structurally analyzes the dynamic echo state vector of the node and the dynamic echo state vectors of the neighboring nodes in the comprehensive time sequence state information to obtain the structurally analyzed comprehensive time sequence state information; Each agent extracts the voltage amplitude change feature, the active power change feature, the reactive power change feature and the node association relationship change feature of the node at the current time according to the structurally analyzed comprehensive time sequence state information, and combines the change features in a fixed order into a time sequence running change set of the node. ​ The intelligent agent extracts voltage amplitude variation characteristics, active power variation characteristics, reactive power variation characteristics and node association relationship variation characteristics of all adjacent nodes at the current time according to the integrated time sequence state information after structure analysis, and arranges the variation characteristics into an adjacent time sequence operation variation set in the order of fixed node index; Each intelligent agent fuses the time sequence operation variation set of the node with the adjacent time sequence operation variation set according to a fixed combination format, so that the fused time sequence operation variation structure can reflect the local topological structure relationship and the synchronous time sequence operation state of the node and the adjacent nodes; Each intelligent agent generates a node-level adjustment amount according to the fused time sequence operation variation structure; Each intelligent agent sets the node-level adjustment amount as a local stability control instruction at the current time, and executes the local stability control instruction through the corresponding execution device of the node.

7. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the disturbance updated local stability control instruction includes: When the intelligent agent detects new energy output fluctuation, load change or line disturbance, the intelligent agent obtains disturbance time sequence operation data of the node, and arranges the disturbance time sequence operation data into a disturbance update input sequence in the order of fixed node index; The disturbance update input sequence is input into the improved graph echo state network, and a disturbance update reserve pool state is formed; The disturbance update reserve pool state is arranged in the order of fixed node index to form a disturbance update dynamic echo state vector of the node, which is sent to all adjacent intelligent agents having a connection relationship with the node, and all adjacent node disturbance update dynamic echo state vectors generated at the same disturbance moment are received; All adjacent node disturbance update dynamic echo state vectors are arranged in the order of fixed node index; The disturbance update dynamic echo state vector of the node is spliced with the arranged adjacent node disturbance update dynamic echo state vectors in a fixed combination format to form disturbance update integrated time sequence state information; According to the disturbance update integrated time sequence state information, voltage amplitude variation characteristics, active power variation characteristics, reactive power variation characteristics and node association relationship variation characteristics of the node and the adjacent nodes at the disturbance moment are extracted, and the variation characteristics are arranged into a disturbance update operation amount set; According to the disturbance update operation amount set, a node-level adjustment amount is generated, which is used as a disturbance update local stability control instruction, and the disturbance update local stability control instruction is executed through the corresponding execution device.

8. The multi-agent based power distribution network stability control method of claim 1, wherein, The generation of the local stability control instruction includes: When the intelligent agent detects new energy output fluctuation, load change or line disturbance, the intelligent agent obtains disturbance time sequence operation data of the node, and arranges the disturbance time sequence operation data into a disturbance update input sequence in the order of fixed node index; Each intelligent agent inputs the periodic update input sequence into the improved graph echo state network, and forms a periodic update reserve pool state, and arranges the periodic update reserve pool state in the order of fixed node index to form a periodic update dynamic echo state vector of the node; Each intelligent agent sends the periodic update dynamic echo state vector of the node to all adjacent intelligent agents having a connection relationship with the node, and receives periodic update dynamic echo state vectors of all adjacent nodes; The intelligent agent arranges all the neighborhood period update dynamic echo state vectors in order of fixed node index, and splices the arranged neighborhood period update dynamic echo state vectors and the node period update dynamic echo state vector according to a fixed combination format; The intelligent agent extracts the voltage amplitude change characteristics, active power change characteristics, reactive power change characteristics and node association relationship change characteristics of the node and the neighborhood nodes according to the period update comprehensive time sequence state information, and arranges the extracted change characteristics into a period update operation quantity set; The intelligent agent generates a period update node-level adjustment quantity according to the period update operation quantity set, takes the period update node-level adjustment quantity as a period update local stability control instruction, and executes the period update local stability control instruction through a corresponding execution device; After the execution of the period update local stability control instruction is completed, the intelligent agent distinguishes the period update input sequence, the period update reserve pool state, the period update dynamic echo state vector, the period update comprehensive time sequence state information and the period update node-level adjustment quantity from the corresponding contents generated in the disturbance update process, so that the period update control chain and the disturbance update control chain are independent in structure and can maintain overall stable operation in a double-track manner during the operation of the power distribution network.

Images