A complex network feature guided master-slave coordination system dynamic reconstruction method

Abstract

The application provides a complex network feature guided main distribution coordination system dynamic reconstruction method, including: checking connectivity and radiation constraint on candidate distribution network topology, generating a feasible topology set and sampling to obtain a labeled sample set; calculating a reconstruction period operation cost for each topology sample in the labeled sample set and extracting a corresponding complex network feature vector to fit a period operation cost prediction model based on the complex network feature vector; predicting the operation cost of each topology in the feasible topology set through the model to obtain a high-value candidate topology set after preliminary screening and clustering to generate a high-value topology set; based on the high-value topology set, training an online Q network through Markov decision process modeling and reinforcement learning; and outputting a reconstruction action by using the online Q network at each reconstruction period to realize dynamic reconstruction of the main distribution system. The application can improve the solution efficiency and decision quality of the main distribution coordination system dynamic reconstruction under the condition of high proportion of photovoltaic access.

Description

Technical Field

[0001] This invention belongs to the field of power system operation and control technology, and specifically relates to a dynamic reconfiguration method for a master-distributor coordinated system guided by complex network characteristics. Background Technology

[0002] Large-scale development and grid connection of distributed photovoltaic (PV) power has become one of the main ways to utilize clean energy. In scenarios with a high proportion of distributed PV grid connection, coordinated operation of main and distribution networks helps alleviate the problem of safe integration of renewable energy. Network reconfiguration can improve power flow distribution and enhance system operation by optimizing the switch states of the distribution network. In scenarios involving coordinated operation of main and distribution networks, network reconfiguration involves not only adjustments to internal branch switches but also the selection of power supply interfaces at the boundary between main and distribution networks, making its decision space and coupling relationships more complex. Since dynamic network reconfiguration is essentially a mixed-integer nonlinear programming problem with temporal coupling characteristics, the difficulty of solving it increases significantly with the expansion of the system scale. Therefore, improving the solution efficiency of dynamic network reconfiguration in coordinated main and distribution systems is a key issue supporting the flexible, safe, and efficient operation of distributed renewable energy.

[0003] Traditional distribution network reconfiguration methods can be broadly categorized into two types: heuristic algorithm-based methods and mixed-integer programming-based methods. Heuristic algorithm-based methods experience a rapid increase in computation time with the number of controllable variables and struggle to obtain optimal solutions. Mixed-integer programming-based methods rely on explicit mathematical models, leading to complex optimization problem modeling and low solution efficiency. Some studies employ convex optimization techniques, such as second-order cone relaxation, to simplify the reconfiguration mathematical model and improve solution efficiency. Building on this, some studies propose decomposing distribution network optimization reconfiguration into topology generation and solving a linear model based on a defined topology, reducing model complexity. Other studies propose dynamic distribution network reconfiguration methods based on decision-oriented time-series division, improving the solution speed for multi-time-series reconfiguration problems. However, as distribution network scale continues to expand, both of these methods still have limitations in adaptability to complex scenarios and efficiency in high-dimensional combinatorial search, making it difficult to meet the dynamic control requirements of main-distribution coordinated systems.

[0004] Deep reinforcement learning (DRM) is widely used in power system optimization and scheduling due to its adaptability to uncertain environments and fast online decision-making speed. Some studies employ DRM to eliminate reliance on explicit physical models and improve the efficiency of online decision-making for reconfiguration strategies. Other studies propose a structure-aware DRM dynamic reconfiguration method that optimizes the decision space of dynamic reconfiguration from three aspects: dimensionality, performance, and structure, thereby improving solution efficiency and interpretability. Some studies apply graph reinforcement learning to the collaborative optimization of active distribution network dynamic reconfiguration and reactive power compensation, utilizing graph structure information to enhance the agent's perception of the topological coupling relationships of the distribution network. Still other studies combine DRM with multi-objective optimization to coordinate multiple objectives such as network losses, voltage deviations, renewable energy consumption, and operating costs, improving the search efficiency for complex reconfiguration problems. These studies demonstrate that DRM provides an effective solution for sequential decision-making problems such as distribution network dynamic reconfiguration.

[0005] However, current research has not yet addressed the dynamic reconfiguration problem of master-distributor coordinated systems. Applying deep reinforcement learning to the dynamic reconfiguration of master-distributor coordinated systems still faces significant challenges: First, there is a strong coupling relationship between the distribution network topology, power flow distribution, boundary power interaction, and switching actions in master-distributor coordinated systems. Local reconfiguration actions can cause power flow redistribution within the distribution network, further affecting the power transmission at the master-distributor boundary and the operation mode of the master network. Existing deep reinforcement learning methods typically flatten the network state into vector inputs, making it difficult to effectively capture the correlation between topology and operational performance, resulting in weak perception of network structure by the policy network. Second, master-distributor coordinated dynamic reconfiguration simultaneously involves internal switch combinations and boundary power supply interface selection, resulting in a huge action space and drastic fluctuations in the reward function. Traditional deep reinforcement learning algorithms suffer from problems such as training instability and convergence difficulties. Complex network theory provides an effective analytical tool for revealing the characteristics and operational mechanisms of network topology. Introducing complex network features into the deep reinforcement learning framework helps enhance the agent's perception of network structural attributes, guides the policy network to focus on key switching actions, thereby improving training efficiency and decision quality. Therefore, it is necessary to embed complex network features as mechanistic guiding information into the deep reinforcement learning decision-making process to enhance the interpretability and solution performance of the dynamic reconstruction method for master-partner collaborative systems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a dynamic reconfiguration method for master-distributor coordinated systems guided by complex network characteristics. This invention extracts feasible topology samples and features from multiple types of complex networks, establishes a mapping relationship between topology features and operating costs using XGBoost, generates a high-value topology set using K-Medoids clustering, and employs graph reinforcement learning to achieve offline training and online solution of the dynamic reconfiguration strategy. This method improves the solution efficiency and interpretability of dynamic reconfiguration for master-distributor coordinated systems, enhances reinforcement learning training efficiency, and improves decision quality. In high-proportion distributed photovoltaic (PV) grid integration scenarios, distribution network reconfiguration not only improves PV absorption capacity by changing power flow distribution but also alters the power interaction at the master-distributor boundary and affects the main grid operation. In actual grid operation, this invention can quickly generate dynamic reconfiguration strategies that balance economy and security in high-proportion integration scenarios with PV penetration rates of 200%–250%, effectively reducing PV reduction rates and network losses, and improving renewable energy absorption capacity. Furthermore, this invention can adaptively respond to source-load prediction deviations and topology changes, exhibiting strong robustness and generalization ability, and is applicable to master-distributor coordinated systems of different scales and structures.

[0007] A first aspect of this invention proposes a dynamic reconfiguration method for a master-slave cooperative system guided by complex network features, comprising:

[0008] A graph search method is used to test the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; the set of feasible topologies is sampled to generate a set of labeled samples;

[0009] For each topology sample in the labeled sample set, calculate the operating cost for each reconstruction period and extract the complex network feature vector corresponding to the topology sample. Fit the model to obtain the period operating cost prediction model based on the complex network feature vector.

[0010] Based on the time period operation cost prediction model, by predicting the operation cost of each topology in the feasible topology set, a high-value candidate topology set after initial screening is selected from topologies with low operation cost potential.

[0011] Clustering is performed on the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity.

[0012] Based on the high-value topology set, Markov decision process modeling is performed, and the multi-period dynamic reconstruction problem is modeled as a reinforcement learning sequential decision problem. After reinforcement learning training is completed, an online Q-network for generating multi-period dynamic reconstruction strategies is obtained.

[0013] When each reconfiguration period arrives, the online Q network is used to output a reconfiguration action to achieve dynamic reconfiguration of the main and auxiliary systems.

[0014] In one specific embodiment of the present invention, generating the labeled sample set includes:

[0015] 1) Calculate the photovoltaic center-oriented extension cost of the distribution network branch;

[0016] Among them, branch roads The cost of the photovoltaic center-oriented expansion is:

[0017] (1)

[0018] In the formula, For distribution network branch edge weights, For distribution network Net photovoltaic injection center, Indicates distribution network node Distance to the net photovoltaic center As a weighting factor;

[0019] The calculation expression for the net photovoltaic injection center is as follows:

[0020] (2)

[0021] In the formula, To reconstruct the time period set, For distribution network The set of nodes, For distribution network node During the period Available photovoltaic power output below For distribution network node During the participation period The active load below, For distribution network node Topological location, Indicates taking the positive part;

[0022] 2) Based on the results of step 1), generate candidate topologies;

[0023] Among them, distribution network Using the boundary power supply node as the starting node, the Prim algorithm is used to progressively expand the spanning tree. Each time, the branch with the lowest photovoltaic center-guided expansion cost is selected from the candidate expansion branch set formed by all connecting branches between the currently added node set and the node set not yet added to the spanning tree; when the resulting topology covers the distribution network... When all nodes are found, a candidate topology that satisfies connectivity and radial constraints is obtained;

[0024] 3) By changing the combination of the main and distribution network boundary nodes, and by repeatedly executing steps 1)-2) based on different weight factors, multiple candidate topologies are generated; duplicate topologies are removed from all candidate topologies, and finally a set of feasible topologies oriented towards photovoltaic center is obtained;

[0025] 4) Based on the results of step 3), sample topological samples without replacement uniformly from the feasible topology set to form a labeled sample set.

[0026] In a specific embodiment of the present invention, calculating the operating cost for each reconstruction period for each topological sample in the labeled sample set includes:

[0027] Annotated sample set Each topological sample in In each refactoring period Operating costs The network reconfiguration optimization model for a high-proportion distributed photovoltaic main-distribution coordinated system is obtained by solving the network reconfiguration optimization model under this topological sample;

[0028] The objective function expression of the network reconfiguration optimization model for the high-proportion distributed photovoltaic main-distribution coordinated system is as follows:

[0029] (3)

[0030] in,

[0031] (4)

[0032] (5)

[0033] (6)

[0034] (7)

[0035] (8)

[0036] In the formula, To reconstruct the time period set; The length of a single time period; Total operating cost, , , , Time periods The costs of generating electricity from the units, the costs of photovoltaic reduction, the costs of grid losses, and the costs of switching penalties; For time period node The generator has active power output; , , They are nodes The quadratic coefficient, linear coefficient, and constant term of the generator's power generation cost; , and These are the photovoltaic reduction penalty coefficient, the grid loss cost coefficient, and the switching action penalty coefficient, respectively. , Separate time periods node Available photovoltaic power output and actual power consumption; For time period Reconstructed branch The switch status is 1, indicating that the branch is closed, and 0 indicates that the branch is open; A set of photovoltaic access nodes; Represents the set of branches in a distribution network; For time period branch road Active power loss, branch road electrical conductivity, For nodes During the period voltage amplitude, For nodes and nodes During the period The voltage phase angle difference;

[0037] The constraints of the optimization model include:

[0038] Parent node constraints;

[0039] (9)

[0040] (10)

[0041] In the formula, For the set of distribution network nodes, Represents nodes The set of directly connected adjacent nodes. For nodes and nodes During the period The parent node is associated with the variable, if the node It is a node During the period The parent node, then Otherwise, it is 0;

[0042] Constraints on the relationship between branch switch state variables and parent / child node variables;

[0043] (11)

[0044] (12)

[0045] (13)

[0046] Constraints on the main and distribution boundary power supply branches;

[0047] (14)

[0048] In the formula, For distribution network aggregation; For distribution network The set of candidate main and distribution boundary power supply branches; Indicates time period Branch closure, Indicates time period The branch circuit is disconnected;

[0049] Closed branch constraints;

[0050] (15)

[0051] Constraint on the number of switch switching operations;

[0052] (16)

[0053] In the formula, This represents the number of switching actions compared to the previous time period. This represents the maximum number of switching actions allowed in a single time segment.

[0054] Photovoltaic power output constraints;

[0055] (17)

[0056] In the formula, This is the minimum power factor allowed for a photovoltaic inverter. Its corresponding maximum power factor angle;

[0057] Trend constraints;

[0058] (18)

[0059] (19)

[0060] In the formula, , The points are nodes. and nodes The per-unit value of voltage amplitude, , These are the resistance and reactance of the branch circuit, respectively. It is a positive number used to relax the constraint when the branch is disconnected;

[0061] Cybersecurity constraints:

[0062] (20)

[0063] In the formula, This is the lower limit of the per-unit value of the voltage amplitude. This represents the per-unit upper limit of the voltage amplitude.

[0064] In a specific embodiment of the present invention, the step of extracting the complex network feature vector corresponding to the topology sample includes:

[0065] 1) Calculate the seepage threshold;

[0066] Among them, for topological samples ,set up Preserve the probability for the edge; given Randomly retain topological samples Some branches in the middle are repeated. After the Monte Carlo test, the magnetic susceptibility of the network was calculated:

[0067] (twenty one)

[0068] In the formula, This represents the proportion of the largest connected component to the total number of nodes. Indicates Monte Carlo average;

[0069] Then topology The seepage threshold is:

[0070] (twenty two)

[0071] 2) Calculate the fractal dimension;

[0072] For topological samples The box-covering algorithm is used to calculate the fractal dimension; where the box diameter is . The minimum number of boxes required to cover all nodes is ,but and It approximately satisfies a power-law relationship:

[0073] (twenty three)

[0074] The fractal dimension is then obtained from the slope of the linear fit in double logarithmic coordinates:

[0075] (twenty four)

[0076] 3) Calculate the network diameter;

[0077] For topological samples The network diameter is represented as:

[0078] (25)

[0079] In the formula, For a set of nodes, Representing topology Next node With nodes The number of branches traversed by the shortest path between them, i.e., the topological hop count;

[0080] 4) Calculate the photovoltaic weighted topology depth;

[0081] For topological samples Time period The photovoltaic weighted topology depth is represented as follows:

[0082] (26)

[0083] In the formula, 0 represents the root node of the substation. For topology From the next root node to the photovoltaic node Topological hop count; weight From time period The available photovoltaic power is normalized to obtain:

[0084] (27)

[0085] 5) Calculate the photovoltaic weighted path resistance;

[0086] For topological samples Time period The photovoltaic weighted path resistance is expressed as:

[0087] (28)

[0088] In the formula, For topology From the next root node to the photovoltaic node The only path branch road The resistance;

[0089] 6) Calculate the transmission capacity at the primary / secondary boundary;

[0090] For topological samples The primary-secondary boundary transmission capacity is expressed as:

[0091] (29)

[0092] In the formula, For topology The number of the main distribution boundary power supply node connected to the downstream distribution network;

[0093] 7) Based on the results of steps 1)-6), the feature vectors of the complex network corresponding to the topology sample;

[0094] Among them, time period Lower topological samples The feature vector of a complex network is represented as:

[0095] (30).

[0096] In a specific embodiment of the present invention, the fitting of the time-period running cost prediction model based on complex network feature vectors includes:

[0097] 1) For each topological sample Based on each time period Operating costs and feature vectors of complex networks Construct the XGBoost training dataset:

[0098] (31)

[0099] In the formula, For time period The following XGBoost training dataset;

[0100] 2) Based on the XGBoost training dataset obtained in step 1), fit a time-period running cost prediction model based on complex network feature vectors;

[0101] Among them, the feature vector of complex network As input, the time-period operating cost To produce the output, a time-period operating cost prediction model is established:

[0102] (32)

[0103] In the formula, For the predicted topology During the period The operating costs under the specified conditions; this prediction model is for a specific time period. The corresponding XGBoost regression model;

[0104] Then topology The predicted total operating cost over the entire dynamic reconfiguration cycle is expressed as:

[0105] (33).

[0106] In a specific embodiment of the present invention, the process of constructing the high-value candidate topology set after initial screening is as follows:

[0107] Based on the time-period operating cost prediction model, from the feasible topology set In this process, the operating costs are predicted according to the topology corresponding to each time period. Sort by size from smallest to largest and select the first few. These topologies constitute a time-segmented set of high-value topologies:

[0108] (34)

[0109] In the formula, This indicates sorting by the indicators in parentheses from smallest to largest and selecting the top ones. One element;

[0110] From feasible topology set In the middle, the predicted total operating cost over the entire reconstruction cycle corresponding to each topology is calculated. Sort by size from smallest to largest and select the first few. These topologies constitute a multi-time-phase integrated high-value topology set:

[0111] (35)

[0112] Will and Taking the union of the sets yields the set of high-value candidate topologies after the initial screening by XGBoost:

[0113] (36).

[0114] In a specific embodiment of the present invention, the clustering of the high-value candidate topology set after initial screening includes:

[0115] K-Medoids clustering was used to remove redundancy from the high-value candidate topology set after initial screening;

[0116] Among them, Each topology Represented as length is Branch switch state vector:

[0117] (37)

[0118] in, Representing topology The Middle The switch status of each branch circuit; This indicates the total number of operable switch branches in the distribution network;

[0119] Then, Any two topologies and The structural distance between them is represented by Hamming distance as follows:

[0120] (38)

[0121] Hamming distance is used as a similarity metric for K-Medoids clustering to measure the degree of structural difference in the switching states between two topologies; the number of clusters is set. ,right Perform K-Medoids clustering; after clustering, select the topology with the lowest total XGBoost prediction cost within each cluster as the representative topology of that cluster, and the final high-value topology set is composed of the representative topologies of each cluster. .

[0122] In one specific embodiment of the present invention, obtaining the online Q-network for generating a multi-time-period dynamic reconstruction strategy includes:

[0123] 1) Construct a Markov decision process model;

[0124] Among them, the agent in the first Each decision step observes the current state. Select Action The environment updates the topology and calculates operating costs based on actions, providing immediate rewards. and the next state This process can be represented as a Markov decision process:

[0125] (39)

[0126] In the formula, For state space; For action space; The state transition probability; For the reward function; Discount factor;

[0127] The state space is as follows:

[0128] (40)

[0129] In the formula, For the current topology The corresponding graph structure includes the set of nodes, the set of closed branches, and their adjacency relationships; The node feature matrix; The characteristic matrix of the branch; For the current topology in the th The time period corresponding to each decision step The complex network feature vectors are the time-period running costs predicted by XGBoost; and These are the load level coefficient and photovoltaic output coefficient after considering disturbances, respectively;

[0130] The node feature matrix is ​​represented as follows:

[0131] (41)

[0132] In the formula, and They are nodes During the period Active and reactive loads; For nodes During the period Available photovoltaic power; For photovoltaic node marking; This serves as the root node identifier; For nodes In topology The degree of the nodes;

[0133] The branch feature matrix is ​​represented as follows:

[0134] (42)

[0135] In the formula, For the first Each decision-making branch The on / off state; and Branch roads Resistance and reactance; Indicates an operable switch;

[0136] Introducing source-load perturbation variables into the state space, the perturbed load level coefficient and photovoltaic output coefficient are respectively:

[0137] (43)

[0138] (44)

[0139] In the formula, and The first The baseline load level and photovoltaic output factor for each decision step; and These are respectively the load and photovoltaic output disturbance terms; and These are the standard deviations of load and photovoltaic disturbance, respectively.

[0140] Actions are taken from high-value topology sets Select a candidate topology from the options, the first... The action of each decision step is defined as:

[0141] (45)

[0142] action Indicates that the agent is in the first... Each decision step selects a high-value topology set. The Middle Each candidate topology is used as a reconstruction action; after the action is executed, the topology is changed from... Updated to ;

[0143] The set of possible actions after filtering by action mask is:

[0144] (46)

[0145] In the formula, Indicates the state The set of possible actions after filtering by action mask; Number the candidate actions; To start from the current topology Switch to candidate topology The required number of switching actions;

[0146] During state transition, when the agent is in state Select action Afterwards, the environment will update the topology as follows:

[0147] (47)

[0148] Environment based on the next period The source load level is recalculated to determine node characteristics, branch characteristics, complex network characteristics, and XGBoost prediction running cost, generating the next state:

[0149] (48)

[0150] In the formula, This is the state transition function;

[0151] No. The immediate reward for each decision step is:

[0152] (49)

[0153] In the formula, For time period Operating costs of the baseline topology; For agent state Next action The reconstructed topology obtained later in the time period Operating costs below; The penalty coefficient for the infeasibility of exchanging optimal trendlines;

[0154] Operating costs consist of photovoltaic reduction costs, grid loss costs, and switching operation costs.

[0155] (50)

[0156] In the formula, To reduce costs for photovoltaics; For network loss costs; For the first The number of switching actions per decision step relative to the reference topology;

[0157] The cumulative reward for each training round is:

[0158] (51)

[0159] In the formula, The number of decision steps included in each training round;

[0160] 2) Fusion of graph structure and complex network features into state encoding;

[0161] Specifically, a graph convolutional network is used to extract the structural embedding of the current topology, and the complex network features and the XGBoost prediction running cost are jointly input into the value network as global prior information.

[0162] The output of the graph structure encoder is represented as follows:

[0163] (52)

[0164] In the formula, It is a GCN graph structure encoder. For the first Graph structure embedding for each decision step;

[0165] By concatenating the GCN graph structure embedding, complex network features, XGBoost prediction runtime cost, and source load perturbation variables, the fused state embedding of DDQN is obtained:

[0166] (53)

[0167] In the formula, Embedding the fusion state of DDQN;

[0168] 3) Dual-Depth Q-Network Training and Decision Making;

[0169] Among them, in the first In a reinforcement learning decision-making step, the agent, based on... The -greedy strategy selects an action from the set of available actions:

[0170] (54)

[0171] In the formula, This indicates uniform random sampling from the set of possible actions; For online Q network; For online network parameters;

[0172] The agent performs actions Afterwards, immediate rewards for environmental feedback. and the next state To form interactive samples:

[0173] (55)

[0174] In the formula, This is the end marker; when hour, ;otherwise, Interactive samples are stored in the experience replay pool. During training, small batches of samples are randomly sampled from the experience replay pool to update the online Q-network;

[0175] The temporal difference objective value of a dual deep Q-network is:

[0176] (56)

[0177] In the formula, For the target network parameters, Used to assess the value of the action; This is a termination marker;

[0178] The loss function of a dual-depth Q-network is:

[0179] (57)

[0180] In the formula, This represents the training loss of the action value network;

[0181] Online Q-networks approximate the action value function given by the DDQN objective value by minimizing the loss function.

[0182] In a specific embodiment of the present invention, the implementation of dynamic reconfiguration of the main and auxiliary systems includes:

[0183] 1) Collect current operating status information of the power distribution network;

[0184] At each decision step, the current operating status information of the distribution network is collected, including: the active power load of each node. reactive load Available photovoltaic power Photovoltaic node markers Root node marker In topology The degree of the nodes Resistance of each branch Reactance Current switch status Operable switch indicator The connectivity and adjacency relationships of the current topology; the baseline load level coefficient for the current time period. and photovoltaic power output coefficient ;

[0185] 2) Based on the results of step 1), construct the current state;

[0186] Specifically, based on the collected operational status information (1), a graph structure of the current topology is constructed. Node feature matrix and branch feature matrix Then construct the feature vector of the current complex network. Based on the time-period operating cost prediction model, the operating cost of the current topology in the current time period is obtained. Thus, the current state is generated. ;

[0187] 3) Encode the current state obtained in step 2) into a graph structure, and then input it into the trained online Q network. Select the reconstruction action for the current decision step based on the output of the online Q network.

[0188] The online Q-network output is an action value estimate for each candidate topology in the high-value topology set. And select the candidate topology with the highest action value as the reconstruction action for the current decision step:

[0189] (58)

[0190] 4) Based on the reconfiguration action obtained in step 3), perform the corresponding opening and closing operations on each branch switch in the distribution network to complete the reconfiguration task for the current time period.

[0191] A second aspect of the present invention provides a dynamic reconfiguration device for a master-slave cooperative system guided by complex network features, comprising:

[0192] The topology sample generation module is used to use a graph search method to check the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; and to sample the set of feasible topologies to generate a set of labeled samples.

[0193] The time-period operation cost prediction model fitting module is used to calculate the operation cost of each reconstructed time period for each topology sample in the labeled sample set and extract the complex network feature vector corresponding to the topology sample, and fit the model to obtain the time-period operation cost prediction model based on the complex network feature vector.

[0194] The high-value candidate topology set initial screening module is used to select low-operating-cost potential topologies to form a high-value candidate topology set after initial screening by predicting the operating costs of each topology in the feasible topology set based on the time period operation cost prediction model.

[0195] The high-value topology set optimization module is used to cluster the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity.

[0196] The reinforcement learning module is used to perform Markov decision process modeling based on the high-value topology set, and to model the multi-period dynamic reconstruction problem as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network is obtained to generate multi-period dynamic reconstruction strategies.

[0197] The online reconfiguration module is used to output reconfiguration actions using the online Q network when each reconfiguration period arrives, thereby realizing dynamic reconfiguration of the main and auxiliary systems.

[0198] A third aspect of the present invention provides an electronic device comprising:

[0199] At least one processor; and a memory communicatively connected to said at least one processor;

[0200] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described dynamic reconfiguration method for a master-slave cooperative system guided by complex network features.

[0201] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described dynamic reconfiguration method for a master-slave cooperative system guided by complex network features.

[0202] The features and beneficial effects of this invention are as follows:

[0203] This invention addresses the challenges of existing deep reinforcement learning methods in dynamically reconstructing master-distributor coordinated systems under high-proportion distributed photovoltaic (PV) grid connection conditions. These methods struggle to effectively capture the correlation between topology and operational performance, and the massive action space leads to training instability and convergence difficulties. To address these issues, a novel dynamic reconstruction method for master-distributor coordinated systems guided by complex network features is proposed. This method achieves a deep integration of complex network theory and deep reinforcement learning by constructing a solution framework that combines complex network feature representation, high-value topology selection, and graph reinforcement learning decision-making. Specifically, this invention first extracts complex network features from three aspects: network structure morphology, photovoltaic spatial distribution and electrical accessibility, and the support capability of the main-distribution boundary. Then, XGBoost is used to establish a mapping relationship between these complex network features and multi-period operating costs, providing interpretable mechanistic guidance for network reconfiguration decisions. Based on this, a method combining XGBoost prediction screening and K-Medoids clustering is employed to construct a high-value topology set that combines low-cost potential with structural diversity, reducing the size of the reinforcement learning action space while ensuring search quality. Finally, the graph structure embedding extracted by the graph convolutional network, the complex network features, the XGBoost predicted operating costs, and the source-load states are jointly embedded into the state space. A dual-depth Q-network based on the graph convolutional network guides the agent to search for dynamic reconfiguration strategies within the high-value action space. This invention can improve the solution efficiency and decision quality of dynamic reconfiguration of main-distribution coordinated systems under high-proportion photovoltaic access conditions, and enhance the mechanistic interpretability of the reconfiguration strategy.

[0204] Specifically, the present invention has the following advantages:

[0205] 1) This invention proposes a method for characterizing complex network features of a high-proportion photovoltaic main-distribution coordinated system. It extracts complex network features from three aspects: network structure morphology, photovoltaic spatial distribution and electrical accessibility, and main-distribution boundary support capability. It also uses XGBoost to fit the mapping relationship between complex network features and multi-time period operating costs, providing an interpretable mechanism guidance for network reconfiguration decisions.

[0206] 2) This invention proposes a high-value topology set screening method that combines XGBoost prediction with K-Medoids clustering. By screening for operational cost potential and removing structural redundancy, a candidate topology set with both low cost potential and structural diversity is constructed, thereby reducing the size of the reinforcement learning action space while ensuring search quality.

[0207] 3) This invention proposes a graph reinforcement learning solution framework guided by complex network features. Based on a dual-depth Q-network of graph convolutional networks, the graph structure of the graph convolutional network, complex network features, XGBoost prediction running cost and source charge state are jointly embedded into the state space, guiding the agent to search for dynamic reconstruction strategies in the high-value action space, thereby improving the interpretability of the dynamic reconstruction decision mechanism, solution efficiency and training stability.

[0208] 4) In actual power grid operation, this invention expands the action space from 10 through high-value topology sets. 19 The computational power is compressed to around 300, reducing the online solution time for the intelligent agent to only 0.01 seconds, meeting the real-time operation requirements of the power grid. It can quickly generate dynamic reconfiguration strategies that balance economy and security in high-proportion grid access scenarios with photovoltaic penetration rates of 200% to 250%, effectively reducing photovoltaic reduction rate and network losses, improving the capacity for renewable energy absorption, and can adaptively cope with source-load prediction deviations and topology changes. It has strong robustness and generalization ability and is suitable for main-distribution coordinated systems of different scales and structures. Attached Figure Description

[0209] Figure 1 This is an overall flowchart of a dynamic reconfiguration method for a master-slave collaborative system guided by complex network features, as described in an embodiment of the present invention. Detailed Implementation

[0210] This invention proposes a dynamic reconfiguration method for a master-slave collaborative system guided by complex network characteristics, which is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0211] A method for dynamic reconfiguration of a master-slave cooperative system guided by complex network features, as proposed in the first aspect of this invention, includes:

[0212] A graph search method is used to test the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; the set of feasible topologies is sampled to generate a set of labeled samples;

[0213] For each topology sample in the labeled sample set, calculate the operating cost for each reconstruction period and extract the complex network feature vector corresponding to the topology sample. Fit the model to obtain the period operating cost prediction model based on the complex network feature vector.

[0214] Based on the time period operation cost prediction model, the operation cost of each topology in the feasible topology set is predicted, and a high-value candidate topology set after initial screening is selected from the topologies with low operation cost potential.

[0215] Clustering is performed on the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity.

[0216] Based on the high-value topology set, Markov decision process modeling is performed, and the multi-period dynamic reconstruction problem is modeled as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network for generating multi-period dynamic reconstruction strategies is obtained.

[0217] When each reconfiguration period arrives, the online Q network is used to output a reconfiguration action to achieve dynamic reconfiguration of the main and auxiliary systems.

[0218] In a specific embodiment of the present invention, the overall process of the dynamic reconfiguration method for a master-slave cooperative system guided by complex network features is as follows: Figure 1 As shown, it includes the following steps:

[0219] 1) Use graph search methods to check the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; sample the set of feasible topologies to generate a set of labeled samples.

[0220] Specifically, considering the large size of the candidate topologies for the TD408 main-distribution coordinated system, it is difficult to obtain all feasible topologies that satisfy the radial constraints through full enumeration. This embodiment proposes a photovoltaic center-oriented heuristic sampling method, which guides the spanning tree to preferentially expand towards the concentrated area of ​​net photovoltaic injection, constructing a set of feasible topologies that combine connectivity, radial constraints, and photovoltaic absorption potential, and further generates a labeled sample set through sampling; the specific steps are as follows:

[0221] 1-1) Calculate the photovoltaic center-guided expansion cost of the distribution network branch.

[0222] In this embodiment, branches are defined. The cost of the photovoltaic center-oriented expansion is:

[0223] (1)

[0224] In the formula, For distribution network branch The edge weight (here, the line impedance is taken). For distribution network Net photovoltaic injection center, Indicates distribution network node Distance to the net photovoltaic center This is the weighting factor.

[0225] The calculation expression for the net photovoltaic injection center is as follows:

[0226] (2)

[0227] In the formula, To reconstruct the time period set, For distribution network The set of nodes, For distribution network node During the period Available photovoltaic power output below For distribution network node During the participation period The active load below, For distribution network node Topological location, This indicates taking the positive part.

[0228] 1-2) Based on the results of step 1-1), generate candidate topologies.

[0229] In this embodiment, the distribution network is used. Starting with the boundary power supply node, the Prim algorithm is used to progressively expand the spanning tree. Each time, the branch with the lowest photovoltaic center-guided expansion cost is selected from the candidate expansion branch set formed by all connecting branches between the currently added node set and the node set not yet added to the spanning tree. When the resulting topology covers the distribution network... When all nodes are found, a candidate topology that satisfies connectivity and radial constraints is obtained.

[0230] 1-3) By changing the combination of the main and distribution network boundary nodes, and repeating steps 1-1)-1-2) based on different weight factors, multiple candidate topologies are generated; duplicate topologies are removed from all candidate topologies, and finally a set of feasible topologies oriented towards photovoltaic center is obtained.

[0231] In one specific embodiment of the present invention, a photovoltaic center-oriented heuristic sampling method is used in the TD408 master-distributor coordination system to generate 10,000 candidate topology samples that satisfy connectivity and radial constraints to form a feasible topology set and participate in subsequent reconstruction decisions.

[0232] 1-4) Based on the results of steps 1-3), sample topological samples without replacement uniformly from the feasible topology set to form a labeled sample set.

[0233] In one specific embodiment of the present invention, 1,000 samples are uniformly sampled without replacement from a set of 10,000 feasible topologies for subsequent complex network feature extraction and XGBoost model training.

[0234] 2) Based on the results of step 1), calculate the operating cost of each reconstruction period for each topological sample in the labeled sample set and extract the complex network feature vector corresponding to the topological sample, and fit the model to obtain the time period operating cost prediction model based on the complex network feature vector.

[0235] In this embodiment, the labeled sample set Each topological sample in By each reconstruction period Calculate operating costs And extract the corresponding complex network feature vectors. A time-period operating cost prediction model based on complex network feature vectors is obtained through fitting; the specific steps are as follows:

[0236] 2-1) Calculate the operating cost for each reconstruction period for each topological sample in the labeled sample set.

[0237] In this embodiment, each reconstruction period Operating costs The network reconfiguration optimization model of a high-proportion distributed photovoltaic main and distribution coordinated system is obtained by solving the network reconfiguration optimization model under this topological sample.

[0238] Specifically, the optimization objective of the network reconfiguration optimization model for a high-proportion distributed photovoltaic (PV) main-distribution coordinated system is to minimize the total operating cost of the main and distribution networks while ensuring the safe operation of the system. The objective function expression of this optimization model is as follows:

[0239] (3)

[0240] in,

[0241] (4)

[0242] (5)

[0243] (6)

[0244] (7)

[0245] (8)

[0246] In the formula, To reconstruct the time period set; The length of a single time period; Total operating cost, , , , They are respectively The unit's power generation cost, photovoltaic reduction cost, grid loss cost, and switch-off penalty cost during the time period; for Time period nodes The generator has active power output; , , They are nodes The quadratic coefficient, linear coefficient, and constant term of the generator's power generation cost; , and These are the photovoltaic reduction penalty coefficient, the grid loss cost coefficient, and the switching action penalty coefficient, respectively. , They are respectively Time period nodes Available photovoltaic power output and actual power consumption; For time period Reconstructed branch The switch status is 1, indicating that the branch is closed, and 0 indicates that the branch is open; A set of photovoltaic access nodes; Represents the set of branches in a distribution network; For time period branch road Active power loss, branch road electrical conductivity, For nodes During the period voltage amplitude, For nodes and nodes During the period The voltage phase angle difference.

[0247] The constraints of this optimization model include:

[0248] Parent node constraints;

[0249] Distribution networks typically operate using a "closed-loop design, open-loop operation" mode; therefore, the reconfigured network should maintain a radial structure. To ensure that every node except the root node has one and only one parent node, and to ensure that the root node has no parent node, the following conditions must be met:

[0250] (9)

[0251] (10)

[0252] In the formula, For the set of distribution network nodes, Represents nodes The set of directly connected adjacent nodes. For nodes and nodes During the period The parent node is associated with the variable, if the node It is a node During the period The parent node, then Otherwise, it is 0.

[0253] Constraints on the relationship between branch switch state variables and parent / child node variables;

[0254] The following conditions should be met between the branch switch state variables and the parent-child node associated variables:

[0255] (11)

[0256] (12)

[0257] (13)

[0258] Constraints on the main and distribution boundary power supply branches;

[0259] In a primary-distribution coordinated system, each distribution network has only one primary-distribution boundary power supply branch in a closed state at any given time, that is:

[0260] (14)

[0261] In the formula, For distribution network aggregation; For distribution network The set of candidate primary and secondary boundary power supply branches; Indicates time period Branch closure, Indicates time period The branch line is disconnected.

[0262] Closed branch constraints;

[0263] To ensure that the reconstructed network forms a radial topology with the root node as the source, the number of closed branches should satisfy:

[0264] (15)

[0265] Constraint on the number of switch switching operations;

[0266] To avoid equipment damage and operational risks caused by frequent switching operations, a constraint is set on the number of switching operations between adjacent time periods:

[0267] (16)

[0268] In the formula, This represents the number of switching actions compared to the previous time period. This represents the maximum number of switching actions allowed in a single time segment.

[0269] In addition, the constraints of this optimization model should also include photovoltaic power output constraints, power flow constraints, and network security constraints.

[0270] Photovoltaic output constraints;

[0271] While outputting active power, in order to fully utilize the reactive power regulation capability of the photovoltaic inverter to support the voltage, its reactive power output needs to be limited by the inverter capacity and minimum power factor. Non-photovoltaic nodes do not need to consider active or reactive power injection.

[0272] (17)

[0273] In the formula, This is the minimum power factor allowed for a photovoltaic inverter. For its corresponding maximum power factor angle, this embodiment takes... .

[0274] Trend constraints;

[0275] For any closed branch Neglecting branch losses, the nonlinear term of network losses can be ignored. Assuming the per-unit voltage value is close to 1, the voltage squared difference approximately satisfies:

[0276] (18)

[0277] (19)

[0278] In the formula, , The points are nodes. and nodes The per-unit value of voltage amplitude, , These are the resistance and reactance of the branch circuit, respectively. This is a sufficiently large positive number to relax the constraint when the branch is disconnected. In this embodiment, it is determined based on the maximum voltage difference of the IEEE 33-bus system. The value of is chosen to ensure sufficient relaxation.

[0279] Cybersecurity constraints:

[0280] All node voltage magnitudes must meet the following requirements:

[0281] (20)

[0282] In the formula, the lower limit of the per-unit value of the voltage amplitude is taken. The per-unit upper limit of voltage amplitude (Per-unit value based on rated voltage). For the balancing node of a substation, the voltage amplitude is fixed at 1.0 pu.

[0283] In this embodiment, the network reconfiguration optimization model for a high-proportion distributed photovoltaic main-distribution coordinated system belongs to the mixed-integer nonlinear programming problem. For a given topology sample, the switch state variables in this model are fixed, thus degenerating into a conventional optimal power flow problem. This embodiment uses the commercial solver Gurobi on the MATLAB platform, calling the YALMIP toolbox for solution.

[0284] Solving this model yields the corresponding topological samples. Operating costs at different times .

[0285] 2-2) For each topology sample, extract the feature vector of the complex network corresponding to that topology.

[0286] In this embodiment, six complex network features are extracted from three aspects: network structure morphology, photovoltaic spatial distribution and electrical accessibility, and primary-distribution boundary support. Among them, the network structure morphology features include percolation threshold, fractal dimension, and network diameter; the photovoltaic spatial distribution and electrical accessibility features include photovoltaic weighted topology depth and photovoltaic weighted path resistance; and the primary-distribution boundary support features include primary-distribution boundary transmission capability.

[0287] Specifically, the percolation threshold is used to characterize the degradation sensitivity of the topology's connectivity scale under conditions of limited transmission in local branches. For a given topology sample ,set up Preserve the probability for the edge. In a given... Randomly retain topological samples Some branches in the middle are repeated. After the Monte Carlo test, the magnetic susceptibility of the network was calculated:

[0288] (twenty one)

[0289] In the formula, This represents the proportion of the largest connected component to the total number of nodes. This represents the Monte Carlo average. The number of Monte Carlo trials typically ranges from 50 to 500. In a specific embodiment of this invention, [the number of trials is...]. A specific embodiment of the present invention is as follows: The scan is performed within the interval with a step size of 0.01, and the value corresponding to the first peak value of the network magnetic susceptibility is taken. Value as seepage threshold .

[0290] Then topological sample The seepage threshold is defined as:

[0291] (twenty two)

[0292] The lower the value, the better the network can maintain large-scale connectivity under random branch disturbances, indicating lower topological vulnerability, which is beneficial for reducing the risk of local blockage in scenarios with extremely high photovoltaic penetration.

[0293] Fractal dimension is used to measure the degree of structural unfolding and self-similarity of a distribution network topology at different scales. For a given topology sample The box-covering algorithm is used to calculate the fractal dimension. Let the box diameter be... The minimum number of boxes required to cover all nodes is ,but and It approximately satisfies a power-law relationship:

[0294] (twenty three)

[0295] Therefore, the fractal dimension can be obtained from the slope of the linear fit in double logarithmic coordinates:

[0296] (twenty four)

[0297] This reflects the degree to which the network changes from a locally clustered structure to a globally uniformly expanded structure. A reasonable expanded structure helps improve the dispersion of power transmission paths, but excessive expansion may also cause voltage rise at the end point. Therefore, it needs to be analyzed in conjunction with the location-related characteristics of photovoltaics.

[0298] Network diameter is used to characterize the longest power supply scale of a network, reflecting the sensitivity of end nodes to voltage and power flow constraints. For a given topology sample... The network diameter can be expressed as:

[0299] (25)

[0300] In the formula, For a set of nodes, Representing topology Next node With nodes The number of branches traversed by the shortest path between them, i.e., the topological hop count. The larger the value, the longer the power supply link in the topology, and the more sensitive the voltage of the end node is to fluctuations in photovoltaic output.

[0301] To characterize the spatial distribution of photovoltaic resources in the topology at different times, for a given topology sample Time period The photovoltaic weighted topology depth can be expressed as:

[0302] (26)

[0303] In the formula, 0 represents the root node of the substation. For topology From the next root node to the photovoltaic node Topological hop count. Weight From time period The available photovoltaic power is normalized to obtain:

[0304] (27)

[0305] The larger the value, the longer the time period. More available photovoltaic power is located deep in the topology, which may cause more significant voltage rise and local reverse power flow constraints.

[0306] To further characterize the electrical distance of photovoltaic power transmission from the access node to the main grid, for a given topology sample Time period The photovoltaic weighted path resistance can be expressed as:

[0307] (28)

[0308] In the formula, For topology From the next root node to the photovoltaic node The only path branch road The resistance. The larger the indicator, the longer the time period. The higher the equivalent transmission resistance corresponding to the available photovoltaic power, the more significant the limitations of branch losses, voltage rise, and power flow constraints on photovoltaic absorption may be.

[0309] The primary-distribution boundary transmission capacity is used to characterize the power supply capacity and photovoltaic (PV) absorption support capacity of distribution network boundary nodes in a primary-distribution cooperative topology. For a given topology sample Define the primary and secondary boundary transmission capabilities:

[0310] (29)

[0311] In the formula, For topology The number of the main distribution boundary power supply node connected to the downstream distribution network. The larger the value, the stronger the power supply support capacity and photovoltaic absorption capacity of the distribution network under the current main distribution network structure.

[0312] In this embodiment, for each topological sample After performing complex calculations on the above six characteristics, the time period can be obtained. The feature vector of the complex network is represented as follows:

[0313] (30)

[0314] 2-3) Based on the results of steps 2-1) and 2-2), construct the XGBoost training dataset, and then use this dataset to fit a time-based runtime cost prediction model based on complex network feature vectors; the specific steps are as follows:

[0315] 2-3-1) Based on the results of steps 2-1) and 2-2), construct the XGBoost training dataset.

[0316] In this embodiment, for each topological sample Based on each time period Operating costs and feature vectors of complex networks Construct the XGBoost training dataset:

[0317] (31)

[0318] In the formula, For time period The following XGBoost training dataset; For topology During the period Feature vectors of complex networks; For topology During the period The following is the actual operating cost obtained through optimization calculation.

[0319] 2-3-2) Based on the XGBoost training dataset obtained in step 2-3-1), fit a time-period running cost prediction model based on complex network feature vectors.

[0320] In this embodiment, the feature vector of a complex network is used. As input, the time-period operating cost To produce the output, a time-period operating cost prediction model is established:

[0321] (32)

[0322] In the formula, For the predicted topology During the period The operating cost under this condition. In this embodiment, the model is for a time period. The corresponding XGBoost regression model.

[0323] Then topology The predicted total operating cost over the entire dynamic reconfiguration cycle is expressed as:

[0324] (33)

[0325] 3) Based on the time period operation cost prediction model obtained in step 2), the operation cost of each topology in the feasible topology set is predicted, and the topologies with low operation cost potential are selected to form a high-value candidate topology set after initial screening.

[0326] In this embodiment, based on the time-period operating cost prediction model, from the feasible topology set In this process, the operating costs are predicted according to the topology corresponding to each time period. Sort by size from smallest to largest and select the first few. A topology, The value range can be adjusted from 500 to 2000 based on the total number of candidate topologies and computing resources. In a specific embodiment of this invention, 1000 is used to form a time-segmented set of high-value topologies.

[0327] (34)

[0328] In the formula, This indicates sorting by the indicators in parentheses from smallest to largest and selecting the top ones. One element. Due to It is built according to time periods, thus preserving topologies with the potential for low operating costs in different time periods.

[0329] From feasible topology set In the middle, the predicted total operating cost over the entire reconstruction cycle corresponding to each topology is calculated. Sort by size from smallest to largest and select the first few. A topology, The range of values ​​is In a specific embodiment of this aspect, 2000 is used to form a multi-time period integrated high-value topology set:

[0330] (35)

[0331] Will and Taking the union of the sets yields the set of high-value candidate topologies after the initial screening by XGBoost:

[0332] (36)

[0333] In the formula, This is a set of high-value candidate topologies after initial screening by XGBoost. This set includes topologies that perform well across different time periods, as well as topologies with lower overall operating costs across multiple time periods, thus balancing time-specific differences and overall lifecycle economics.

[0334] 4) Cluster the high-value candidate topology set obtained in step 3) to obtain a high-value topology set that has both low operating cost potential and structural diversity.

[0335] In this embodiment, K-Medoids clustering is used to perform structural redundancy removal on the high-value candidate topology set obtained in step 4).

[0336] Specifically, Each topology Represented as length is Branch switch state vector:

[0337] (37)

[0338] in, Representing topology The Middle The switch status of each branch circuit. This represents the total number of operable switch branches in the distribution network. Each topology uses a symbol of length . The vector representation ensures the comparability of structural distances between different topologies.

[0339] Then, Any two topologies and The structural distance between them is represented by Hamming distance as follows:

[0340] (38)

[0341] In the formula, The smaller the value, the closer the switching states of the two topologies are, and the higher the structural similarity.

[0342] The Hamming distance is used here as a similarity metric for K-Medoids clustering, measuring the degree of structural difference in the switching states between two topologies. The smaller the Hamming distance, the closer the switching states of the two topologies are, and the higher their structural similarity.

[0343] Set the number of clusters ,right Perform K-Medoids clustering. Generally, the number of action space topologies is adjusted between 100 and 500 based on the desired size. In a specific embodiment of this invention, 300 is used, compressing the action space from 10,000 candidate topologies to 300 high-value topologies. After clustering, the topology with the lowest total XGBoost prediction cost within a cluster is selected as the representative topology of that cluster. The final set of high-value topologies is composed of the representative topologies of each cluster. This serves as the action space for subsequent dynamic reconstruction through reinforcement learning.

[0344] Thus far, this embodiment has compiled a set of feasible topologies. Compressed to a size of The high-value topology set reduces structural redundancy while retaining potential topologies with low operating costs, thereby reducing the size of the reinforcement learning action space and improving training efficiency.

[0345] 5) Based on the high-value topology set obtained in step 4), a Markov decision process model is performed, and the multi-period dynamic reconstruction problem is modeled as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network is obtained to generate multi-period dynamic reconstruction strategies.

[0346] In this embodiment, at each decision step, the agent perceives the current distribution network topology, node and branch operating attributes, complex network characteristics, XGBoost operating cost priors, and source load operating states, and selects a candidate topology from the high-value topology set as the reconfiguration action. To improve the agent's perception of topology structure and global operating cost potential, this embodiment uses a Graph Convolutional Network (GCN) to extract graph structure embeddings, and combines complex network characteristics and XGBoost prediction results to construct state representations. Finally, a Double Deep Q-Network (DDQN) is used to learn multi-period dynamic reconfiguration strategies. The specific steps are as follows:

[0347] 5-1) Construct a Markov decision process model.

[0348] In this embodiment, the dynamic reconstruction process can be viewed as a sequential interaction process between the agent and the environment. Each decision step observes the current state. Select Action The environment updates the topology and calculates operating costs based on actions, providing immediate rewards. and the next state This process can be represented as a Markov decision process:

[0349] (39)

[0350] In the formula, For state space; For action space; The state transition probability; For the reward function; This is the discount factor.

[0351] In this embodiment, the state space is defined as:

[0352] (40)

[0353] In the formula, For the current topology The corresponding graph structure includes the set of nodes, the set of closed branches, and their adjacency relationships; The node feature matrix; The characteristic matrix of the branch; For the current topology in the th The time period corresponding to each decision step The complex network feature vectors are the time-period running costs predicted by XGBoost; and These are the load level coefficient and photovoltaic output coefficient after considering disturbances, respectively.

[0354] The node feature matrix is ​​represented as follows:

[0355] (41)

[0356] In the formula, and They are nodes During the period Active and reactive loads; For nodes During the period Available photovoltaic power; For photovoltaic node marking; This serves as the root node identifier; For nodes In topology The degree of the nodes.

[0357] The branch feature matrix is ​​represented as follows:

[0358] (42)

[0359] In the formula, For the first Each decision-making branch The on / off state; and Branch roads Resistance and reactance; This is an indicator for an operable switch.

[0360] To characterize the impact of source load prediction errors on the reconfiguration strategy, this embodiment introduces source load perturbation variables into the state space. The perturbed load level coefficient and photovoltaic output coefficient are as follows:

[0361] (43)

[0362] (44)

[0363] In the formula, and The first The baseline load level and photovoltaic output factor for each decision step; and These are respectively the load and photovoltaic output disturbance terms; and These are the standard deviations of load and photovoltaic disturbance, respectively.

[0364] Actions are defined from high-value topology sets Select a candidate topology from the options, the first... The action of each decision step is defined as:

[0365] (45)

[0366] That is, action Indicates that the agent is in the first... Each decision step selects a high-value topology set. Middle Each candidate topology is used as a reconstruction action. After the action is executed, the topology is changed from [previous topology]. Updated to .

[0367] Since the distribution network operation topology needs to satisfy connectivity, radial constraints, and constraints on the number of switching operations in adjacent time periods, the set of possible actions after action masking is defined as follows:

[0368] (46)

[0369] In the formula, Indicates the state The set of possible actions after filtering by action mask; Number the candidate actions; To start from the current topology Switch to candidate topology The number of switching actions required.

[0370] During state transition, when the agent is in state Select action Afterwards, the environment will update the topology as follows:

[0371] (47)

[0372] Environment based on the next period The source load level is recalculated to determine node characteristics, branch characteristics, complex network characteristics, and XGBoost prediction running cost, generating the next state:

[0373] (48)

[0374] In the formula, This is the state transition function.

[0375] In this embodiment, the immediate reward uses a cost reduction relative to the baseline topology to guide the agent to search for a dynamic reconstruction strategy with lower operating costs and fewer switching actions. The immediate reward for each decision step is:

[0376] (49)

[0377] In the formula, For time period Operating costs of the baseline topology; For agent state Next action The reconstructed topology obtained later in the time period Operating costs below; This is the infeasibility penalty coefficient for Alternating Current Optimal Power Flow (ACOPF). The operating cost consists of photovoltaic reduction costs, grid loss costs, and switching operation costs.

[0378] (50)

[0379] In the formula, To reduce costs for photovoltaics; For network loss costs; For the first The number of switching actions per decision step relative to the reference topology. For the first decision period, Compared to the initial topology calculation; for subsequent time periods, Topology calculation has been selected compared to the previous period.

[0380] The cumulative reward for each training round is:

[0381] (51)

[0382] In the formula, The number of decision steps included in each training round.

[0383] 5-2) Fusion state coding of graph structure and complex network features.

[0384] Distribution networks naturally possess graph structure attributes. To improve state representation capabilities, this embodiment employs graph convolutional networks to extract the structural embedding of the current topology and inputs the complex network features and XGBoost-predicted operating costs as global prior information into the value network.

[0385] The output of the graph structure encoder is represented as follows:

[0386] (52)

[0387] In the formula, It is a GCN graph structure encoder. For the first A graph structure embedding for each decision step.

[0388] By concatenating the GCN graph structure embedding, complex network features, XGBoost prediction runtime cost, and source load perturbation variables, the fused state embedding of DDQN is obtained:

[0389] (53)

[0390] In the formula, This involves embedding the fused state of DDQN. This fusion approach enables the agent to learn both local electrical connectivity relationships and leverage the global runtime cost potential information provided by complex network features and XGBoost predictions.

[0391] 5-3) Training and decision-making of dual deep Q-networks.

[0392] In this embodiment, a dual-depth Q-network algorithm is used to solve the dynamic reconstruction strategy. The dual-depth Q-network selects the optimal action in the next state through an online network and evaluates the value of the action using a target network, thereby decoupling action selection from action value evaluation and reducing the risk of Q-value overestimation.

[0393] In the In a reinforcement learning decision-making step, the agent, based on... The -greedy strategy selects an action from the set of available actions:

[0394] (54)

[0395] In the formula, This indicates uniform random sampling from the set of possible actions; For online Q network; These are online network parameters.

[0396] The agent performs actions Afterwards, immediate rewards for environmental feedback. and the next state To form interactive samples:

[0397] (55)

[0398] In the formula, This is the termination marker. When hour, ;otherwise, Interaction samples are stored in the experience replay pool. During training, small batches of samples are randomly sampled from the experience replay pool to update the online Q-network.

[0399] The temporal difference objective value of a dual deep Q-network is defined as:

[0400] (56)

[0401] In the formula, For the target network parameters, Used to evaluate the value of the action. This is due to the introduction of a stop marker. The objective value of the last decision step is no longer added to the estimated value of the next state.

[0402] The loss function of a dual-depth Q-network is:

[0403] (57)

[0404] In the formula, This is the training loss for the action value network. The online Q-network approximates the action value function given by the DDQN objective value by minimizing the loss function.

[0405] In one specific embodiment of the present invention, the training parameters of the dual deep Q network are set with a learning rate of 0.0003, a graph structure encoder dimension of {173,128}, a GCN layer of 2, an embedding dimension of 64, a batch size of 128, a discount factor of 0.99, and a target network update cycle of every 250 rounds. The -greedy strategy has an initial exploration rate of 1.0, a final exploration rate of 0.01, and an exploration decay step count of 2000. Each training epoch contains 24 decision steps, corresponding to 24 time intervals. The training termination condition is set to reach the maximum number of training epochs of 8000. During training, the network model parameters are saved every 100 epochs, and the average cumulative reward for those 100 epochs is recorded. After training reaches 8000 epochs, the model with the highest average cumulative reward is selected as the final training result, and the parameters of the online Q-network at this point are saved. This is used for subsequent online decision-making applications.

[0406] This enables the construction of a solution framework of "complex network feature representation - high-value topology screening - graph reinforcement learning decision-making," which can effectively improve the solution efficiency and decision quality of dynamic reconfiguration of main and distribution coordinated systems under high-proportion photovoltaic access conditions.

[0407] 6) Use the online Q-network trained in step 5) to perform dynamic reconfiguration of the primary and secondary systems.

[0408] In one specific embodiment of the present invention, after reinforcement learning training is completed, the trained online Q-network parameters are... It can be used directly for online decision-making. The specific steps are as follows:

[0409] 6-1) Collect the current operating status information of the power distribution network.

[0410] In this embodiment, one decision step corresponds to one hour. At each decision step, the current operating status information of the distribution network is collected. The specific parameters collected include: the active power load of each node. reactive load Available photovoltaic power Photovoltaic node markers Root node marker In topology The degree of the nodes Resistance of each branch Reactance Current switch status Operable switch indicator The connectivity and adjacency relationships of the current topology; the baseline load level coefficient for the current time period. and photovoltaic power output coefficient .

[0411] 6-2) Based on the results of step 6-1), construct the current state.

[0412] In this embodiment, a graph structure of the current topology is constructed based on the data collected in 6-1). Node feature matrix and branch feature matrix Simultaneously, according to the definitions in step 3-2), the complex network feature vector composed of six characteristics of the current topology in the current time period is calculated: percolation threshold, fractal dimension, network diameter, photovoltaic weighted topology depth, photovoltaic weighted path resistance, and primary-distributor boundary transmission capacity. The trained XGBoost model is used to predict the operating cost of the current topology in the current time period. The above information is combined according to equation (34) to form the current state. .

[0413] 6-3) Perform graph structure encoding on the current state obtained in step 6-2), and then input it into the online Q network trained in step 5. Select the reconstruction action for the current decision step based on the output of the online Q network.

[0414] In this embodiment, the state The inputs are fed into the trained graph structure encoder and the online Q-network. The GCN encoder first extracts the graph structure embeddings of the current topology. Subsequently, the graph structure embedding, complex network features, XGBoost prediction runtime cost, and source-load perturbation variables are concatenated into a fused state embedding. The online Q-network estimates the action value of each candidate topology in the high-value topology set based on the fused state embedding. In this embodiment, the output dimension of the online Q network is consistent with the size of the high-value topology set, both being 300.

[0415] Since this is the online application phase, it will no longer be used. Instead of using a greedy exploration strategy, it directly selects the candidate topology with the highest action value as the reconstruction action for the current decision step:

[0416] (58)

[0417] Meanwhile, to ensure that the selected topology meets the distribution network operation constraints, the action masking mechanism defined in step 6-1) is also applied before action selection to eliminate candidate actions that do not meet the connectivity constraints, radial constraints, and constraints on the number of switching actions in adjacent time periods. Only actions that meet the safety constraints are allowed for the agent. Choose from the options provided.

[0418] 6-4) Execute the reconfiguration action obtained in step 6-3), which is to switch from the current topology to the corresponding candidate topology in the high-value topology set. Based on the result of this action, the distribution network dispatcher performs the corresponding opening and closing operations on each branch switch in the actual distribution network to complete the reconfiguration task for the current time period.

[0419] In this embodiment, after the reconfiguration action is performed, the distribution network operates into the next time period, and the agent enters the next decision step, repeating steps 6-1) to 6-4) until all 24 time periods of dynamic reconfiguration decisions are completed. Throughout the online decision-making process, no further optimization model solving or network parameter updates are required; only one GCN forward propagation and one Q-network forward inference calculation are needed. Therefore, the decision-making speed is extremely fast. In one specific embodiment of this invention, the online solution time is only 0.01 seconds, which can meet the real-time requirements of actual power grid operation.

[0420] To implement the above embodiments, a second aspect of the present invention proposes a dynamic reconfiguration device for a master-slave cooperative system guided by complex network features, comprising:

[0421] The topology sample generation module is used to use a graph search method to check the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; and to sample the set of feasible topologies to generate a set of labeled samples.

[0422] The time-period operation cost prediction model fitting module is used to calculate the operation cost of each reconstructed time period for each topology sample in the labeled sample set and extract the complex network feature vector corresponding to the topology sample, and fit the model to obtain the time-period operation cost prediction model based on the complex network feature vector.

[0423] The high-value candidate topology set initial screening module is used to select low-operating-cost potential topologies to form a high-value candidate topology set after initial screening by predicting the operating costs of each topology in the feasible topology set based on the time period operation cost prediction model.

[0424] The high-value topology set optimization module is used to cluster the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity.

[0425] The reinforcement learning module is used to model the Markov decision process based on the high-value topology set, and to model the multi-period dynamic reconstruction problem as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network is obtained to generate multi-period dynamic reconstruction strategies.

[0426] The online reconfiguration module is used to output reconfiguration actions using the online Q network when each reconfiguration period arrives, thereby realizing dynamic reconfiguration of the main and auxiliary systems.

[0427] It should be noted that the foregoing explanation of the embodiment of a dynamic reconfiguration method for a master-slave collaborative system guided by complex network features also applies to a dynamic reconfiguration device for a master-slave collaborative system guided by complex network features in this embodiment, and will not be repeated here. According to an embodiment of the present invention, a dynamic reconfiguration device for a main-distribution coordinated system guided by complex network characteristics is proposed. This device uses a graph search method to test the connectivity and radial constraints of candidate topologies in the distribution network to obtain a set of feasible topologies. The set of feasible topologies is sampled to generate a labeled sample set. For each topology sample in the labeled sample set, the operating cost for each reconfiguration period is calculated, and the complex network feature vector corresponding to the topology sample is extracted. A period-based operating cost prediction model based on the complex network feature vector is fitted. Based on the period-based operating cost prediction model, operating cost prediction is performed on each topology in the feasible topology set, and topologies with low operating cost potential are selected to form a high-value candidate topology set after initial screening. The high-value candidate topology set after initial screening is clustered to obtain a high-value topology set that combines low operating cost potential with structural diversity. Based on the high-value topology set, a Markov decision process model is performed, modeling the multi-period dynamic reconfiguration problem as a reinforcement learning sequential decision problem. After reinforcement learning training, an online Q-network is obtained to generate multi-period dynamic reconfiguration strategies. When each reconfiguration period arrives, the online Q-network outputs a reconfiguration action to realize the dynamic reconfiguration of the main-distribution system. This enables the rapid generation of dynamic reconfiguration strategies that balance economic efficiency and security in high-penetration scenarios with photovoltaic penetration rates of 200% to 250%, effectively reducing photovoltaic reduction rates and network losses, and enhancing the capacity for renewable energy consumption.

[0428] To implement the above embodiments, a third aspect of the present invention provides an electronic device, comprising:

[0429] At least one processor; and a memory communicatively connected to said at least one processor;

[0430] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to execute the above-described dynamic reconfiguration method for a master-slave cooperative system guided by complex network features.

[0431] To implement the above embodiments, a fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described dynamic reconfiguration method for a master-slave cooperative system guided by complex network features.

[0432] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0433] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a dynamic reconfiguration method for a master-slave cooperative system guided by complex network features according to the above embodiments.

[0434] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0435] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0436] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0437] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0438] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0439] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0440] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0441] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0442] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A dynamic reconfiguration method for a master-slave cooperative system guided by complex network characteristics, characterized in that, include: The connectivity and radial constraints of candidate topologies for the distribution network are checked using a graph search method to obtain a set of feasible topologies. Sample the feasible topology set to generate a labeled sample set; For each topology sample in the labeled sample set, calculate the operating cost for each reconstruction period and extract the complex network feature vector corresponding to the topology sample. Fit the model to obtain the period operating cost prediction model based on the complex network feature vector. Based on the time period operation cost prediction model, the operation cost of each topology in the feasible topology set is predicted, and a high-value candidate topology set after initial screening is selected from the topologies with low operation cost potential. Clustering is performed on the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity. Based on the high-value topology set, Markov decision process modeling is performed, and the multi-period dynamic reconstruction problem is modeled as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network for generating multi-period dynamic reconstruction strategies is obtained. When each reconfiguration period arrives, the online Q network is used to output a reconfiguration action to achieve dynamic reconfiguration of the main and auxiliary systems.

2. The method according to claim 1, characterized in that, The generated labeled sample set includes: 1) Calculate the photovoltaic center-oriented extension cost of the distribution network branch; Among them, branch roads The cost of the photovoltaic center-oriented expansion is: （1） In the formula, For distribution network branch edge weights, For distribution network Net photovoltaic injection center, Indicates distribution network node Distance to the net photovoltaic center As a weighting factor; The calculation expression for the net photovoltaic injection center is as follows: （2） In the formula, To reconstruct the time period set, For distribution network The set of nodes, For distribution network node During the period Available photovoltaic power output For distribution network node During the participation period The active load below, For distribution network node Topological location, Indicates taking the positive part; 2) Based on the results of step 1), generate candidate topologies; Among them, distribution network Using the boundary power supply node as the starting node, the Prim algorithm is used to progressively expand the spanning tree. Each time, the branch with the lowest photovoltaic center-guided expansion cost is selected from the candidate expansion branch set consisting of all connecting branches between the currently added node set and the node set not yet added to the spanning tree; when the resulting topology covers the distribution network... When all nodes are found, a candidate topology that satisfies connectivity and radial constraints is obtained; 3) By changing the combination of the main and distribution network boundary nodes, and by repeatedly executing steps 1)-2) based on different weight factors, multiple candidate topologies are generated; duplicate topologies are removed from all candidate topologies, and finally a set of feasible topologies oriented towards photovoltaic center is obtained; 4) Based on the results of step 3), sample topological samples without replacement uniformly from the feasible topology set to form a labeled sample set.

3. The method according to claim 1, characterized in that, The calculation of the operating cost for each reconstruction period for each topological sample in the labeled sample set includes: Annotated sample set Each topological sample in In each refactoring period Operating costs The network reconfiguration optimization model for a high-proportion distributed photovoltaic main-distribution coordinated system is obtained by solving the network reconfiguration optimization model under this topological sample; The objective function expression of the network reconfiguration optimization model for the high-proportion distributed photovoltaic main-distribution coordinated system is as follows: （3） in, （4） （5） （6） （7） （8） In the formula, To reconstruct the time period set; The length of a single time period; Total operating cost, , , , Time periods The costs of generating electricity from the units, the costs of photovoltaic reduction, the costs of grid losses, and the costs of switching penalties; For time period node The generator has active power output; , , They are nodes The quadratic coefficient, linear coefficient, and constant term of the generator's power generation cost; , and These are the photovoltaic reduction penalty coefficient, the grid loss cost coefficient, and the switching action penalty coefficient, respectively. , Separate time periods node Available photovoltaic power output and actual power consumption; For time period Reconstructed branch The switch status is 1, indicating that the branch is closed, and 0 indicates that the branch is open; A set of photovoltaic access nodes; Represents the set of branches in a distribution network; For time period branch road Active power loss, branch road electrical conductivity, For nodes During the period voltage amplitude, For nodes and nodes During the period The voltage phase angle difference; The constraints of the optimization model include: Parent node constraints; （9） （10） In the formula, For the set of distribution network nodes, Represents nodes The set of directly connected adjacent nodes. For nodes and nodes During the period The parent node is associated with the variable, if the node It is a node During the period The parent node, then Otherwise, it is 0; Constraints on the relationship between branch switch state variables and parent / child node variables; （11） （12） （13） Constraints on the main and distribution boundary power supply branches; （14） In the formula, For distribution network aggregation; For distribution network The set of candidate main and distribution boundary power supply branches; Indicates time period Branch closure, Indicates time period The branch circuit is disconnected; Closed branch constraints; （15） Constraint on the number of switch switching operations; （16） In the formula, This represents the number of switching actions compared to the previous time period. This represents the maximum number of switching actions allowed in a single time segment. Photovoltaic output constraints; （17） In the formula, This is the minimum power factor allowed for a photovoltaic inverter. Its corresponding maximum power factor angle; Trend constraints; （18） （19） In the formula, , The points are nodes. and nodes The per-unit value of voltage amplitude, , These are the resistance and reactance of the branch circuit, respectively. It is a positive number used to relax the constraint when the branch is disconnected; Cybersecurity constraints: （20） In the formula, This is the lower limit of the per-unit value of the voltage amplitude. This represents the per-unit upper limit of the voltage amplitude.

4. The method according to claim 3, characterized in that, The extraction of the complex network feature vector corresponding to the topology sample includes: 1) Calculate the seepage threshold; Among them, for topological samples ,set up Preserve the probability for the edge; given Randomly retain topological samples Some branches in the middle are repeated. After the Monte Carlo test, the magnetic susceptibility of the network was calculated: （21） In the formula, This represents the proportion of the largest connected component to the total number of nodes. Indicates Monte Carlo average; Then topology The seepage threshold is: （22） 2) Calculate the fractal dimension; For topological samples The fractal dimension is calculated using the box-covering algorithm; where the box diameter is . The minimum number of boxes required to cover all nodes is ,but and It approximately satisfies a power-law relationship: （23） The fractal dimension is obtained from the slope of the linear fit in double logarithmic coordinates: （24） 3) Calculate the network diameter; For topological samples The network diameter is represented as: （25） In the formula, For a set of nodes, Representing topology Next node With nodes The number of branches traversed by the shortest path between them, i.e., the topological hop count; 4) Calculate the photovoltaic weighted topology depth; For topological samples Time period The photovoltaic weighted topology depth is represented as follows: （26） In the formula, 0 represents the root node of the substation. For topology From the next root node to the photovoltaic node Topological hop count; weight From time period The available photovoltaic power is normalized to obtain: （27） 5) Calculate the photovoltaic weighted path resistance; For topological samples Time period The photovoltaic weighted path resistance is expressed as: （28） In the formula, For topology From the next root node to the photovoltaic node The only path branch road The resistance; 6) Calculate the transmission capacity at the primary / secondary boundary; For topological samples The primary-secondary boundary transmission capacity is expressed as: （29） In the formula, For topology The number of the main distribution boundary power supply node connected to the downstream distribution network; 7) Based on the results of steps 1)-6), the feature vectors of the complex network corresponding to the topology sample; Among them, time period Lower topological samples The feature vector of a complex network is represented as: （30）。 5. The method according to claim 4, characterized in that, The fitting process yields a time-period operating cost prediction model based on complex network feature vectors, including: 1) For each topological sample Based on each time period Operating costs and feature vectors of complex networks Construct the XGBoost training dataset: （31） In the formula, For time period The following XGBoost training dataset; 2) Based on the XGBoost training dataset obtained in step 1), fit a time-period running cost prediction model based on complex network feature vectors; Among them, the feature vector of complex network As input, the time-period operating cost To produce the output, a time-period operating cost prediction model is established: （32） In the formula, For the predicted topology During the period The operating costs under the specified conditions; this prediction model is for a specific time period. The corresponding XGBoost regression model; Then topology The predicted total operating cost over the entire dynamic reconfiguration cycle is expressed as: （33）。 6. The method according to claim 5, characterized in that, The process of constructing the set of high-value candidate topologies after initial screening is as follows: Based on the time-period operating cost prediction model, from the feasible topology set In this process, the operating costs are predicted according to the topology corresponding to each time period. Sort by size from smallest to largest and select the first few. These topologies constitute a time-segmented set of high-value topologies: （34） In the formula, This indicates sorting by the indicators in parentheses from smallest to largest and selecting the top ones. One element; From feasible topology set In the middle, the predicted total operating cost over the entire reconstruction cycle corresponding to each topology is calculated. Sort by size from smallest to largest and select the first few. These topologies constitute a multi-time-phase integrated high-value topology set: （35） Will and Taking the union of the sets yields the set of high-value candidate topologies after the initial screening by XGBoost: （36）。 7. The method according to claim 6, characterized in that, The clustering of the high-value candidate topology set after initial screening includes: K-Medoids clustering was used to remove redundancy from the high-value candidate topology set after initial screening; Among them, Each topology Represented as length is Branch switch state vector: （37） in, Representing topology Middle The switch status of each branch circuit; This indicates the total number of operable switch branches in the distribution network; Then, Any two topologies and The structural distance between them is represented by Hamming distance as follows: （38） Hamming distance is used as a similarity metric for K-Medoids clustering to measure the degree of structural difference in the switching states between two topologies; the number of clusters is set. ,right Perform K-Medoids clustering; after clustering, select the topology with the lowest total XGBoost prediction cost within each cluster as the representative topology of that cluster, and the final high-value topology set is composed of the representative topologies of each cluster. .

8. The method according to claim 7, characterized in that, The obtained online Q-network for generating multi-time dynamic reconstruction strategies includes: 1) Construct a Markov decision process model; Among them, the agent in the first Each decision step observes the current state. Select Action The environment updates the topology and calculates operating costs based on actions, providing immediate rewards. and the next state This process can be represented as a Markov decision process: （39） In the formula, For state space; For action space; The state transition probability; For the reward function; Discount factor; The state space is as follows: （40） In the formula, For the current topology The corresponding graph structure includes the set of nodes, the set of closed branches, and their adjacency relationships; The node feature matrix; The characteristic matrix of the branch; For the current topology in the th The time period corresponding to each decision step The complex network feature vectors are the time-period running costs predicted by XGBoost; and These are the load level coefficient and photovoltaic output coefficient after considering disturbances, respectively; The node feature matrix is ​​represented as follows: （41） In the formula, and They are nodes During the period Active and reactive loads; For nodes During the period Available photovoltaic power; For photovoltaic node marking; This serves as the root node identifier; For nodes In topology The degree of the nodes; The branch feature matrix is ​​represented as follows: （42） In the formula, For the first Each decision-making branch The on / off state; and Branch roads Resistance and reactance; Indicates an operable switch; Introducing source-load perturbation variables into the state space, the perturbed load level coefficient and photovoltaic output coefficient are respectively: （43） （44） In the formula, and The first The baseline load level and photovoltaic output factor for each decision step; and These are respectively the load and photovoltaic output disturbance terms; and These are the standard deviations of load and photovoltaic disturbance, respectively. Actions are taken from high-value topology sets Select a candidate topology from the options, the first... The action of each decision step is defined as: （45） action Indicates that the agent is in the first... Each decision step selects a high-value topology set. Middle Each candidate topology is used as a reconstruction action; after the action is executed, the topology is changed from... Updated to ; The set of possible actions after filtering by action mask is: （46） In the formula, Indicates the state The set of possible actions after filtering by action mask; Number the candidate actions; To start from the current topology Switch to candidate topology The required number of switching actions; During state transition, when the agent is in state Select action Afterwards, the environment will update the topology as follows: （47） Environment based on the next period The source load level is recalculated to determine node characteristics, branch characteristics, complex network characteristics, and XGBoost prediction running cost, generating the next state: （48） In the formula, This is the state transition function; No. The immediate reward for each decision step is: （49） In the formula, For time period Operating costs of the baseline topology; For agent state Next action The reconstructed topology obtained later in the time period Operating costs below; The penalty coefficient for the infeasibility of exchanging optimal trendlines; Operating costs consist of photovoltaic reduction costs, grid loss costs, and switching operation costs. （50） In the formula, To reduce costs for photovoltaics; For network loss costs; For the first The number of switching actions per decision step relative to the reference topology; The cumulative reward for each training round is: （51） In the formula, The number of decision steps included in each training round; 2) Fusion of graph structure and complex network features into state encoding; Specifically, a graph convolutional network is used to extract the structural embedding of the current topology, and the complex network features and the XGBoost prediction running cost are jointly input into the value network as global prior information. The output of the graph structure encoder is represented as follows: （52） In the formula, It is a GCN graph structure encoder. For the first Graph structure embedding for each decision step; By concatenating the GCN graph structure embedding, complex network features, XGBoost prediction runtime cost, and source load perturbation variables, the fused state embedding of DDQN is obtained: （53） In the formula, Embedding the fusion state of DDQN; 3) Dual-Depth Q-Network Training and Decision Making; Among them, in the first In a reinforcement learning decision-making step, the agent, based on... The -greedy strategy selects an action from the set of available actions: （54） In the formula, This indicates uniform random sampling from the set of possible actions; For online Q network; For online network parameters; The agent performs actions Afterwards, immediate rewards for environmental feedback. and the next state To form interactive samples: （55） In the formula, This is the end marker; when hour, ;otherwise, Interactive samples are stored in the experience replay pool. During training, small batches of samples are randomly sampled from the experience replay pool to update the online Q-network; The temporal difference objective value of a dual deep Q-network is: （56） In the formula, For the target network parameters, Used to assess the value of the action; This is a termination marker; The loss function of a dual-depth Q-network is: （57） In the formula, This represents the training loss of the action value network; Online Q-networks approximate the action value function given by the DDQN objective value by minimizing the loss function.

9. The method according to claim 8, characterized in that, The implementation of dynamic reconfiguration of the main and auxiliary systems includes: 1) Collect current operating status information of the power distribution network; At each decision step, the current operating status information of the distribution network is collected, including: the active power load of each node. reactive load Available photovoltaic power Photovoltaic node markers Root node marker In topology The degree of the nodes Resistance of each branch Reactance Current switch status Operable switch indicator The connectivity and adjacency relationships of the current topology; the baseline load level coefficient for the current time period. and photovoltaic power output coefficient ; 2) Based on the results of step 1), construct the current state; Specifically, based on the collected operational status information (1), a graph structure of the current topology is constructed. Node feature matrix and branch feature matrix Then construct the feature vector of the current complex network. Based on the time-period operating cost prediction model, the operating cost of the current topology in the current time period is obtained. Thus, the current state is generated. ; 3) Encode the current state obtained in step 2) into a graph structure, and then input it into the trained online Q network. Select the reconstruction action for the current decision step based on the output of the online Q network. The online Q-network output is an action value estimate for each candidate topology in the high-value topology set. And select the candidate topology with the highest action value as the reconstruction action for the current decision step: （58） 4) Based on the reconfiguration action obtained in step 3), perform the corresponding opening and closing operations on each branch switch in the distribution network to complete the reconfiguration task for the current time period.

10. A dynamic reconfiguration device for a master-slave cooperative system guided by complex network characteristics, characterized in that, include: The topology sample generation module is used to use a graph search method to check the connectivity and radial constraints of candidate topologies of the distribution network to obtain a set of feasible topologies; and to sample the set of feasible topologies to generate a set of labeled samples. The time-period operation cost prediction model fitting module is used to calculate the operation cost of each reconstructed time period for each topology sample in the labeled sample set and extract the complex network feature vector corresponding to the topology sample, and fit the model to obtain the time-period operation cost prediction model based on the complex network feature vector. The high-value candidate topology set initial screening module is used to select low-operating-cost potential topologies to form a high-value candidate topology set after initial screening by predicting the operating costs of each topology in the feasible topology set based on the time period operation cost prediction model. The high-value topology set optimization module is used to cluster the high-value candidate topology set after the initial screening to obtain a high-value topology set that has both low operating cost potential and structural diversity. The reinforcement learning module is used to perform Markov decision process modeling based on the high-value topology set, and to model the multi-period dynamic reconstruction problem as a reinforcement learning sequential decision problem. After the reinforcement learning training is completed, an online Q-network is obtained to generate multi-period dynamic reconstruction strategies. The online reconfiguration module is used to output reconfiguration actions using the online Q network when each reconfiguration period arrives, thereby realizing dynamic reconfiguration of the main and auxiliary systems.

Images