Human-machine fusion power distribution network load restoration system and method based on master-slave game mechanism

Abstract

The application relates to the technical field of power distribution network fault recovery, in particular to a man-machine fusion power distribution network load recovery system and method based on a master-slave game mechanism, which comprises a master-slave game modeling module, a resource dynamic identification module, an expert knowledge guiding module, a feature extraction and reinforcement learning module and a collaborative execution and closed-loop feedback module. After a fault occurs, the power distribution network operation side and the virtual power plant side are abstracted into master-slave decision-making subjects with a hierarchical relationship, the power distribution network operation side serves as the master party, the fault recovery target and the operation constraint condition are determined according to the real-time operation state information and the fault characteristics of the power distribution network, the virtual power plant side serves as the slave party, the available distributed resources in the fault area are dynamically identified and assembled, and the load transfer is coordinated and participated in, efficient load recovery after the fault of the power distribution network is realized, the multi-subject collaborative decision-making conflict problem is solved, the strategy executability and the learning efficiency are improved, and the state perception and decision-making capability in a complex scene are enhanced.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault recovery technology, specifically to a human-machine integrated power distribution network load recovery system and method based on a master-slave game mechanism. Background Technology

[0002] With the continuous expansion of distribution network scale and the large-scale integration of resources such as distributed power sources, energy storage systems, electric vehicles, and controllable loads, the operation mode and control objects of the distribution network under fault conditions have changed, making the operation and recovery process after a fault more complex. After a fault occurs, the original power supply path faces the risk of interruption, and some areas may be disconnected from the main grid, requiring load restoration through network reconfiguration and local resource support. In this process, not only is it necessary to adjust the distribution network topology, but also to quickly identify and integrate available distributed resources around the fault area, and assess their output capacity and power supply continuity to support load restoration in the fault area. Due to the limited capacity of distributed resources, their time-varying operating states, and their discrete spatial distribution, coupled with the interdependent processes of network reconfiguration and resource support, the fault recovery problem is gradually exhibiting complex characteristics of multi-constraint coupling and dynamic evolution.

[0003] Currently, distributed resources are gradually moving away from individual entity-based scheduling and towards aggregated management through virtual power plants. While this operational model improves resource utilization efficiency, it also introduces more complex coordination issues in fault recovery scenarios. On one hand, the distribution network needs to reconfigure its network topology through switching operations to improve overall load recovery rates; on the other hand, the virtual power plant needs to formulate output strategies under constraints such as energy storage capacity, charging and discharging power limitations, and resource availability. Network reconfiguration schemes affect available power paths and power allocation space, while resource output capacity, in turn, constrains the feasibility of topology schemes, creating a mutually influential and hierarchically coupled decision-making relationship. Without a systematic characterization of this hierarchical structure, relying solely on centralized unified optimization or simple coordination assumptions makes it difficult to achieve stable and efficient coordinated recovery under complex topology conditions.

[0004] To improve decision-making efficiency, intelligent methods such as reinforcement learning have been introduced into the field of distribution network fault recovery in recent years. These methods automatically generate network reconfiguration strategies by learning the mapping relationship between operating states and control actions. While these methods have certain advantages in strategy generation capabilities, they still face many challenges in complex electrical constraints and multi-agent coupling environments. On the one hand, in the absence of physical constraints and embedded safety mechanisms, reinforcement learning methods may generate control operations that do not conform to power grid operation procedures during the exploration process, such as forming unacceptable loop network structures or causing voltage overruns, line overloads, and other problems. On the other hand, under high-dimensional action spaces and multiple constraints, the algorithm's exploration efficiency is limited and its convergence speed is slow, making it difficult to simultaneously consider recovery efficiency and operational safety. Furthermore, existing research mostly focuses on single decision-making layers or static optimization frameworks, lacking systematic consideration of dynamic resource building mechanisms, hierarchical interaction relationships, and the embedding methods of safety prior knowledge.

[0005] In summary, in distribution network fault recovery scenarios with extensive participation from virtual power plants, existing methods still have the following shortcomings:

[0006] 1. The lack of a dynamic identification and reorganization mechanism for distributed resources after a failure makes it difficult to quickly construct temporary support units with black start capability and continuous power supply capability;

[0007] 2. The hierarchical coupling relationship between distribution network reconfiguration and virtual power plant resource scheduling is not fully characterized, and a systematic master-slave coordination modeling method is lacking;

[0008] 3. Reinforcement learning methods lack the embedding mechanism of physical laws, operating procedures and prior knowledge, which easily generates infeasible or high-risk operations, and the exploration efficiency is low in the early stage of training;

[0009] 4. In a multi-agent interactive environment, the lack of an effective dynamic feedback and iterative optimization mechanism makes it difficult to guarantee the convergence and stability of the strategy;

[0010] 5. The lack of a unified decision-making framework that can simultaneously consider dynamic resource allocation, hierarchical decision-making relationships, and operational safety constraints makes it difficult to maximize the load recovery rate under fault scenarios. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, this invention proposes a human-machine integrated distribution network load restoration system and method based on a master-slave game mechanism. After a fault occurs, this method abstracts the distribution network operation side and the virtual power plant side into hierarchical master-slave decision-making entities. The distribution network operation side acts as the master, determining the fault restoration target and operational constraints based on the real-time operating status information and fault characteristics of the distribution network. The virtual power plant side acts as the slave, dynamically identifying and assembling available distributed resources in the fault area while meeting its own operational constraints, and coordinating their participation in load transfer. On this basis, by introducing an expert knowledge guidance mechanism based on historical scheduling experience and operating rules into the master-slave decision-making process, the game strategy is constrained and guided, thereby achieving orderly coordination between distribution network operation decisions and virtual power plant resource responses. This solves the problems of multi-entity collaborative decision-making, high-risk operation of intelligent algorithms, and low strategy learning efficiency in distribution network fault restoration with the participation of virtual power plants, and improves the distribution network fault restoration rate, system operation stability, and strategy implementation reliability.

[0012] The present invention discloses a human-machine integrated distribution network load restoration method based on a master-slave game mechanism, the specific steps of which include:

[0013] S1. Modeling the master-slave game environment: Constructing a hierarchical master-slave game decision-making model; S2. Dynamic identification and assembly of available resources in virtual power plant: Obtain the real-time operating status of each distributed resource in the virtual power plant, evaluate the adjustability of various resources, determine and screen the participation of resources according to the power supply restoration needs of the fault area, assemble a set of supporting resources that can participate in fault restoration, and construct resource capability vectors and virtual power plant aggregation capability state variables. S3. Construct an expert knowledge constraint and guidance mechanism for master-slave game decision-making: embed power system experience-based knowledge and rule-based knowledge into the decision-making process, construct a unified decision state vector under the fault recovery scenario, construct an expert demonstration sample set and a joint initialization objective function, establish a hierarchical sample reinforcement mechanism, and construct a safety action constraint mapping function to reconstruct the set of executable actions. S4. Master-Slave Game Reinforcement Learning: Graph attention network is used to extract the topological features of the distribution network, and Transformer structure is used to extract the timing features of the power grid operation. The system-level state representation is obtained by fusion, and master-slave strategy models are constructed respectively. The game is iterated through alternating master-slave updates.

[0014] Furthermore, this invention also provides a human-machine integrated power distribution network load restoration system based on a master-slave game mechanism, comprising:

[0015] Master-Slave Game Modeling Module: Used to construct a hierarchical game decision-making model between the distribution network master and the virtual power plant slave, defining the state space, action space and reward / payout function of both parties;

[0016] Resource dynamic identification module: used to obtain the real-time status of distributed resources in the virtual power plant, assess resource adjustability, screen and build a set of fault recovery support resources, and construct resource capability vectors and aggregated capability status variables;

[0017] Expert knowledge guidance module: This module is used to embed experiential and rule-based knowledge of the power system into the decision-making process, construct a joint initialization objective function, a hierarchical sample reinforcement mechanism, and a safety action constraint mapping function, thereby achieving the constraint and guidance of policy learning.

[0018] Feature Extraction and Reinforcement Learning Module: Used to extract power grid topology and time-series features through graph attention networks and Transformers and fuse them into system-level state representations, construct master-slave strategy models and optimize collaborative decision-making strategies through game theory iteration;

[0019] Collaborative execution and closed-loop feedback module: used to control the topology reconfiguration operation on the distribution network side and the resource coordination and allocation on the virtual power plant side, collect the grid operation status and resource output status after execution and feed them back to the preceding module to realize the continuous correction and optimization of the strategy.

[0020] The beneficial effects of this invention are as follows:

[0021] 1. This invention constructs a master-slave game collaborative decision-making architecture between the distribution network and the virtual power plant, clarifies the dominant decision-making position of the distribution network, characterizes the optimal response mechanism of the virtual power plant, and considers the power grid operation safety constraints and the internal resource regulation capabilities of the virtual power plant. It achieves unified optimization of system-level safety objectives and multi-entity collaborative scheduling objectives, effectively solving the contradiction between the traditional centralized scheduling and the difficulty in coordinating the autonomous operation of distributed resources and the overall system safety objectives.

[0022] 2. This invention introduces an expert knowledge guidance mechanism, embedding power system operation experience and scheduling rules into the reinforcement learning process. Through joint initialization of the objective function, hierarchical sample reinforcement mechanism, and safety action constraints, the policy search always meets the power grid safety requirements, reduces ineffective exploration behavior, improves the stability and convergence efficiency of policy learning, and enables the system to obtain a fault recovery strategy with engineering feasibility within a short training period.

[0023] 3. This invention combines graph attention networks and Transformers to construct a joint feature representation model of power grid topology and time-series information, and comprehensively models the structural changes, source-load fluctuations and dynamic changes of the power grid after a fault. This enables the decision model to maintain stable decision-making ability when the power grid topology is reconstructed or when there are uncertain disturbances in the operating data, thereby improving the adaptability and operational reliability of the distribution network under complex operating conditions.

[0024] 4. This invention uses a closed-loop control mechanism of "strategy generation - collaborative execution - status feedback" to feed back the execution status of distribution network topology reconstruction and virtual power plant resource scheduling to the decision system in real time, continuously correct the master-slave game strategy, and enable the distribution network and virtual power plant to form a stable collaborative relationship, which significantly improves fault recovery efficiency and power supply reliability. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 Flowchart of the human-machine integrated power distribution network load restoration method based on master-slave game mechanism provided by the present invention;

[0027] Figure 2 A diagram of an expert knowledge-guided reinforcement learning framework for master-slave game decision-making provided by this invention;

[0028] Figure 3 The diagram shows the master-slave game collaborative decision-making structure of the power distribution network and virtual power plant provided by this invention. Detailed Implementation

[0029] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0030] Example 1:

[0031] Please see Figures 1-3 As shown, the present invention provides a human-machine integrated distribution network load restoration method based on a master-slave game mechanism, applicable to load transfer and power restoration scenarios after a distribution network fault. Specific steps include:

[0032] S1. Modeling of a master-slave game environment;

[0033] After a fault occurs in the distribution network, some power supply paths may be interrupted, and the load in the faulty area needs to be restored through network reconfiguration and distributed resource support. To describe the collaborative decision-making relationship between the distribution network operator and the virtual power plant resource aggregator, this invention constructs a hierarchical master-slave game decision-making model. In this model, the distribution network operator acts as the master, primarily responsible for formulating network reconfiguration strategies and load restoration requirements based on the grid's operating status; the virtual power plant resource aggregator acts as the slave, responding to the master's demands based on its own adjustable resource capabilities after receiving the master's decision information, thus forming a hierarchical collaborative decision-making process with a sequential decision-making order.

[0034] In this master-slave game structure, the master first makes a decision based on the current system operating state and sends a resource support demand signal to the slave. After observing the master's decision, the slave responds by scheduling resources while satisfying its own resource operating constraints. When making decisions, the master takes into account the slave's response behavior, thus forming a master-slave (Stackelberg) hierarchical game structure where the master makes the decision first and the slave responds later. This hierarchical master-slave game decision-making model is constructed by taking the distribution network operator as the master and the virtual power plant resource aggregation entity as the slave, thus forming a master-slave hierarchical game structure where the master makes the decision first and the slave responds later.

[0035] The aforementioned principal decision-making model is expressed as follows:

[0036] ;

[0037] in, This represents the decision-making model of the principal. Represents the master's state space. Indicates the action space of the main party. Indicates the state Next action Then transition to state State transition function, Actions taken by the host From the actions taken by Fang, This represents the principal reward function, used to measure the overall impact of the current decision on system recovery and safe operation. This is the discount factor, and its value range is... This is used to balance immediate recovery effects with long-term operating performance;

[0038] The principal state space is defined as follows:

[0039] ;

[0040] Among them, state variables This indicates the operating status of each section switch and tie switch in the distribution network. This indicates the power supply status of each node's load and its recovery rate. This indicates the state of charge and adjustable capacity information of the energy storage device. This indicates the current aggregated resource capacity and adjustable range that the virtual power plant can provide. This status comprehensively reflects the system's recoverability and operational boundaries after a failure.

[0041] The aforementioned principal action space is defined as:

[0042] ;

[0043] Among them, action variables This indicates a topology switching operation for sectionalizing switches and tie switches. This indicates a power support demand instruction issued to the virtual power plant, specifying the load restoration power level that needs to be supported by the virtual power plant.

[0044] The state transition function Determined by power flow calculations and power grid physical constraints;

[0045] The principal reward function is as follows:

[0046] ;

[0047] in, The main action, This refers to the resource response actions taken by the subordinate party in response to the demands of the primary party. This indicates the percentage of system load recovery under the current decision. Indicates the system's operating power loss. This indicates the degree of violation of constraints such as voltage exceeding limits and line overload. These are non-negative weighting coefficients used to adjust the relative importance of each objective in the optimization process.

[0048] The aforementioned decision-making model is expressed as follows:

[0049] ;

[0050] in This represents the decision-making model from the perspective of the party. Represents the state space from the side. Indicates the action space from the side. This represents the transition relationship of resource status under the combined influence of the main party's demand and its own adjustment behavior. This transition function is affected by the upper limit of distributed power generation output, energy storage capacity constraints, and load regulation capabilities. Represents the profit function from the other side. Discount factor;

[0051] The slave state space is defined as follows:

[0052] ;

[0053] in, This indicates the currently available output capacity of the distributed power source. Indicates the energy storage's state of charge and remaining regulation capacity. This indicates the power range that the adjustable load can participate in regulation. This indicates the power support requirement issued by the master, and this state describes the feasible adjustment space of the slave after receiving the master's scheduling instructions;

[0054] The slave action space is defined as follows:

[0055] ;

[0056] in This represents the output adjustment decision of distributed power sources. This indicates the energy storage charging and discharging control decision. This indicates a controllable load adjustment strategy.

[0057] The revenue function is mainly used to characterize the response of a virtual power plant to the demand for distribution network load transfer and the degree to which operational constraints are met. Its formal definition is:

[0058] ;

[0059] in This indicates the degree of matching between the actual supporting power provided by the supplier and the power required by the supplier. This value can be constructed from the absolute value of the power deviation. This indicates the degree to which resource boundaries have been exceeded or operational restrictions have been violated. These are the weighting coefficients.

[0060] S2. Dynamic identification and assembly of available resources in virtual power plants;

[0061] After a distribution network fault occurs and the master-slave game-theoretic collaborative decision-making model is constructed, it is necessary to identify and dynamically assemble the internal resources of the virtual power plant that can participate in load transfer to the faulted area. Since virtual power plants are typically composed of multiple types of distributed resources, the availability, output capacity, and operational constraints of different resources vary at the time of the fault. Therefore, it is necessary to uniformly identify and assess the capabilities of various resources within the virtual power plant, and dynamically screen and combine resources according to the power restoration needs of the faulted area. This forms a set of virtual power plant support resources that can be used for load restoration, providing a foundation for subsequent collaborative decision-making.

[0062] Acquire the real-time operating status of resources such as distributed power sources, energy storage devices, and controllable loads within the virtual power plant, and evaluate the adjustability of each type of resource.

[0063] The various resources mentioned include:

[0064] Distributed power generation: Adjustable output range ;in, Representing resources Actual output power This indicates the minimum output power of the resource in its current operating state. This indicates its maximum available output power, which is determined by the equipment's rated capacity, current operating status, and grid connection constraints.

[0065] Energy storage devices: Both their power regulation capability and state of charge need to be considered. Assume energy storage resources... At any moment The state of charge is represented as Its range of values ​​satisfies:

[0066] ;

[0067] in, This indicates the state of charge of the energy storage device at time t. and These represent the minimum and maximum permissible states of charge (SOC) of the energy storage device, respectively. Simultaneously, the charging and discharging power of the energy storage device satisfies:

[0068] ;

[0069] in, This indicates the charging power of the energy storage device. This indicates the discharge power of the energy storage device. , , and These represent the corresponding upper and lower power limits, respectively.

[0070] Controllable load: Its regulating capability is usually expressed by the power reduction capability. Let the first... The original load demand of each controllable load node is Therefore, its adjustable power range is:

[0071] ;

[0072] in, Indicates load node Adjustable power refers to the power capacity that can be released through load reduction or peak shifting.

[0073] After completing the assessment of the adjustability of various resources, it is necessary to screen and dynamically assemble resources based on the power restoration needs of the fault area. Let the set of nodes in the fault area be denoted as . The set of resource nodes whose electrical distance is within a certain range is represented as: To ensure that resources can effectively participate in load restoration in fault areas, participation ability assessment of resource units is necessary. Specifically, a resource unit should satisfy the requirement that there is a feasible power supply path between its node and the nodes in the fault area, thus ensuring electrical reachability; simultaneously, the resource should have a certain adjustable power capacity under its current operating state to provide power support to the fault area; furthermore, it must meet operational constraints such as equipment capacity, energy storage state of charge, and load regulation range. Only resource units that meet the above conditions can be included in the candidate resource range of the virtual power plant for fault restoration support.

[0074] Resource units that meet the above conditions are included in the virtual power plant support resource set:

[0075] ;

[0076] in, This represents the set of resources that can participate in fault recovery. This set consists of resource units that meet the requirements of operational availability, sufficient power regulation capability, and network operation constraints, and is dynamically updated as the system operating status and fault location change, thereby realizing the dynamic assembly of virtual power plant resources.

[0077] Based on the power restoration needs of the fault area, the availability of resources is assessed. The assessment criteria are: the resource is electrically accessible to the fault area, has adjustable power capacity, and meets equipment operating constraints. Resources that meet these criteria are included in the support resource set. Construct resource capability vectors And aggregate them to obtain the overall adjustable support capacity of the virtual power plant, forming the virtual power plant state variables in the master-slave game model;

[0078] Where C represents the resource capability vector, and m represents the number of resources participating in fault recovery. This represents the adjustable power capacity of the k-th resource unit, which is determined by the upper limit of the output or the adjustable power range of the corresponding resource.

[0079] S3. Construct an expert knowledge constraint and guidance mechanism for master-slave game decision-making; After completing the construction of the master-slave game decision-making model and the dynamic identification and assembly of available resources in the virtual power plant, to avoid problems such as large-scale ineffective exploration, illegal operations, and response mismatch in the early stages of strategy learning for both the master and slave, this invention further constructs an expert knowledge constraint and guidance mechanism for master-slave game decision-making. This mechanism does not directly rely on random trial and error to generate recovery strategies, but rather transforms the long-term accumulated operational experience and safety rules in power system dispatching into prior information that can be embedded in the decision-making process. This allows both the master, when making network reconfiguration decisions, and the slave, when responding to resource dispatching, to form a feasible initial strategy foundation while meeting operational procedures. Its framework is as follows: Figure 2 As shown.

[0080] The decision-making process incorporates empirical and rule-based knowledge of the power system. The empirical knowledge includes historical fault recovery and dispatch records, offline optimization results, and recovery trajectory samples for typical scenarios. The rule-based knowledge includes distribution system control procedures, equipment operation criteria, and power grid physical mechanisms. Specifically, it prohibits the formation of closed-loop power supply paths, prohibits the formation of multiple power supply paths that are not allowed, avoids load islanding, and restricts the islanding operation of distributed power sources that do not meet the conditions.

[0081] The expert knowledge introduced in this invention comprises two parts: empirical knowledge and rule-based knowledge. Empirical knowledge primarily originates from scheduling records during historical fault recovery processes, offline optimization results, and recovery trajectory samples formed under typical scenarios. It is used to characterize optimal operating modes under different fault locations, load levels, and resource availability conditions. Rule-based knowledge mainly originates from distribution system control procedures, equipment operating criteria, and the physical mechanisms of the power grid. It is used to define inviolable operational boundaries during the recovery process, including constraints such as prohibiting the formation of closed-loop power supply paths, prohibiting the formation of disallowed multi-source power paths, avoiding load islanding, and restricting the islanded operation of distributed power sources that do not meet the conditions. By applying these two types of knowledge to the strategy preference formation process and the security boundary constraint process, respectively, the decision generation in the master-slave game can be both experience-oriented and meets engineering operation constraints.

[0082] In this mechanism, to adapt to reinforcement learning solutions, the system states in the master-slave game model constructed in step one are further refined and expanded to construct decision state vectors for the fault recovery scenario. Used to uniformly represent time The decision-making basis for the principal and subordinate parties. Let's assume the time... The state vector is represented as:

[0083] ;

[0084] in, This represents the state sub-vector of the distribution network structure, used to describe the on / off state of each controllable branch and tie branch after a fault. This represents a load status subvector used to describe the load level of each node, the proportion of restored load, and the power supply status of critical loads; This represents a sub-vector representing the node's operating status, used to describe the voltage amplitude and voltage over-limit information of each node; This represents a sub-vector representing the line's operating status, used to describe the line's power flow and capacity occupancy. This represents a distributed resource state sub-vector, used to describe the adjustable output of distributed power sources, the state of charge of energy storage, and the controllable load regulation capability within a virtual power plant. The above state vector... This is used to uniformly characterize the decision-making basis of the master and slave at the current moment. The master's decision primarily relies on the network structure's state sub-vectors. Load state sub-vector and node and line operating state sub-vectors , The decision-making process mainly relies on the distributed resource state sub-vectors. And the recovery request information issued by the main party, thereby forming a resource scheduling response strategy.

[0085] Construct an expert demonstration sample set based on experiential expert knowledge. Its expression is:

[0086] ;

[0087] in, Indicates the first The system state corresponding to each demonstration sample This indicates that the demonstration actions given by the expert in this state will be demonstrated by the host. Demonstrating the actions from the other side Together they constitute. This indicates the evaluation of the recovery effect obtained after performing this action. This represents the total number of expert samples. The demonstration action... For the master side, this manifests as switching operations and resource demand assignment schemes; for the slave side, it manifests as distributed power generation output allocation, energy storage charging and discharging decisions, and controllable load adjustment decisions. Using the aforementioned sample set, empirically based action preferences can be provided for both the master side's network reconfiguration decision-making strategy and the slave side's resource scheduling response strategy during the initial stages of strategy learning.

[0088] To enable the primary decision-making evaluation model and the secondary response evaluation model to identify the preferred action in the initial stage, this invention no longer directly uses a single value pre-training expression, but instead constructs a joint initialization objective function that includes a recovery benefit fitting term, an expert preference correction term, and parameter constraint terms. Let the parameters of the primary-secondary joint action evaluation model be... This is used to uniformly characterize the joint action value of the master network reconstruction decision and the slave resource scheduling response under a given state. The initial objective function is then denoted as:

[0089] ;

[0090] in, This represents the overall optimization objective during the initialization phase. This represents the recovery payoff fitting term, used to ensure that the strategy evaluation result is consistent with the recovery payoff after the state transition; This indicates an expert preference correction item, used to increase the priority of the expert demonstration action in the corresponding state; This represents parameter constraints used to limit model complexity and improve generalization ability; and , where represents the weighting coefficients. Unlike the original expression that directly focuses on the single Q-value error, the above objective function emphasizes the joint constraints of "revenue consistency + expert preference consistency + model smoothness".

[0091] The recovery return fitting term can be expressed as:

[0092] ;

[0093] in, Represents the training batch sample set, actions This represents a master-slave joint action variable, where the action is performed by the master. With the action of the other party Together constitute The parameter is The policy evaluation function in the state Next action The evaluation value, Indicates the execution of an action The immediate recovery benefits obtained afterward This indicates the next state after the action is performed. Indicates the discount factor. Representing state The following is a set of candidate actions. This represents the reference parameters used for stable training.

[0094] The expert preference correction term can be expressed as:

[0095] ;

[0096] in, This indicates a batch of samples drawn from the expert demonstration sample set. This indicates that the expert demonstrated the action. This represents the minimum evaluation interval constant between expert actions and non-expert actions. Its purpose is to impose an additional penalty on the model when the evaluation result of a non-expert action is too high, thereby encouraging expert demonstration actions to achieve a higher ranking in the corresponding state.

[0097] The parameter constraint terms can be written as:

[0098] ;

[0099] in, Indicates parameters The squared L2 norm is used to suppress model overfitting.

[0100] After establishing initial preferences, to enhance the policy learning process's ability to utilize key samples, this invention further constructs a hierarchical sample reinforcement mechanism. The sample storage unit consists of three subsets, denoted as follows:

[0101] ;

[0102] in, This represents a subset of expert experience used to store historical demonstration samples and offline recovery trajectories; This represents a subset of interactive experiences, used to store normal samples formed by the interaction between the master and slave parties and the environment during the training process; This represents a subset of risk experiences used to store samples that cause constraint out-of-bounds errors, power mismatches, or recovery failures.

[0103] Since the primary and secondary parties form a joint action based on a unified state during the decision-making process, the samples in the sample storage unit are recorded in a joint form. During the policy learning phase, the joint samples are mapped according to the different decision-making entities, with the principal entity based on... Update network reconstruction decision-making strategies from the perspective of [the relevant authority / entity]. Resource scheduling response strategy updates are performed, thereby decoupling the strategy learning process while maintaining the master-slave collaborative relationship.

[0104] Suppose that in one training iteration, the proportions of samples drawn from the three subsets are respectively , and Then we have:

[0105] ;

[0106] in, , , All are intervals The sampling weights within the range. Increase them during the initial training phase. The value of is chosen to strengthen the guiding role of expert samples in updating the parameters of the master-follower decision-making model and forming action preferences; the value is gradually increased in the later stages of training. The proportion of [something] is adjusted to enhance the autonomous adaptability of the master-slave strategy; for The corresponding risk samples retain a certain sampling weight to remind the model to avoid repeating high-risk operations.

[0107] To further highlight the role of high-value samples in training, the priority index for sample z is defined as follows:

[0108] ;

[0109] in, To prevent tiny positive numbers with priority of zero, This represents the evaluation bias of sample z. This is a flag function indicating whether a sample belongs to the high-risk or critical recovery sample category. It takes a value of 1 when the sample belongs to the high-risk or critical load recovery sample category, and a value of 0 otherwise. To add weights, the sample selection probability can be further written as:

[0110] ;

[0111] in, This represents the priority sensitivity coefficient. This mechanism allows samples more closely related to the success or failure of recovery to be used more frequently to update the master-slave game decision model, thereby improving the collaborative learning efficiency of the master's network reconstruction decision and the slave's resource scheduling response.

[0112] Regarding the role of rule-based expert knowledge, this invention no longer simply describes it as the shielding of invalid actions, but rather constructs a safe action constraint mapping function. Let the state be... The following is the candidate action set The security discrimination mapping is defined as follows:

[0113] ;

[0114] in, Used to determine actions In state Is the following a permitted safety action? When an action results in an unacceptable loop in the network, multiple paths between substations, a power outage load island, or a distributed power source island that does not meet the conditions, The value is 0. Based on this security discrimination mapping, a set of constrained executable actions is further constructed:

[0115] ;

[0116] This means that decisions are only allowed from a subset of actions that satisfy the rule constraints. For the case where an evaluation function is used for action selection, the final action can be written as:

[0117] ;

[0118] in, Indicates the state The final selected action is then described. Compared to the traditional method of directly introducing a Boolean mask and subtracting a large number, this expression emphasizes that the invention achieves rule embedding through "reconstruction of a set of safe and executable actions," and its form also differs from the original text.

[0119] Through the aforementioned expert knowledge constraint and guidance mechanism, experiential knowledge is primarily used to establish initial strategy preferences for both the master and slave sides and to enhance the learning effect of high-quality recovery samples. Rule-based knowledge is mainly used to limit the safety boundaries of master-slave decisions and eliminate potentially high-risk actions. The combined effect of these two types of knowledge enables the master-slave game strategy to possess strong feasibility and security in the early stages of training, reduces ineffective exploration behavior, improves the convergence speed of the strategy under complex failure scenarios, and provides a reliable initial strategy foundation for subsequent reinforcement learning-based collaborative optimization of master-slave games.

[0120] S4, Master-Slave Game Reinforcement Learning;

[0121] After initializing the strategy guided by expert knowledge, it is necessary to further optimize the master-slave game strategy between the distribution network operator and the virtual power plant resource aggregator to obtain a collaborative decision-making strategy with high load recovery capability and operational safety under different fault scenarios. To this end, this invention introduces reinforcement learning methods into the master-slave game decision-making framework, gradually optimizing the collaborative strategy through an interactive learning process between the master and slave parties. Figure 3 As shown:

[0122] To accurately describe the operating state of the distribution network during the fault recovery process, the state vector defined in step three... Based on this, graph neural networks are further used to extract the topological features of the distribution network, and Transformer is combined to extract the temporal features during operation, thereby forming a system-level state representation for master-slave game strategy learning.

[0123] Suppose the state of the distribution network at time t is represented by a graphical structure:

[0124] ;

[0125] in, This represents a set of nodes, corresponding to bus nodes and distributed energy nodes in a power distribution network. The set of edges at time t is used to describe the connections between nodes; This represents the node characteristic matrix, used to describe the electrical operating status of each node; This represents the edge attribute matrix, used to describe the electrical parameter characteristics of the line. The node feature vector contains operational information such as node voltage amplitude, node load power, energy storage state of charge, and distributed generation output.

[0126] Based on the graph structure representation described above, a Graph Attention Network (GAT) is used to encode node states in order to extract power grid topology features. In the The layer's features are updated as follows:

[0127] ;

[0128] in, Represents a node In the Layer feature representation, Represents a node The set of neighboring nodes, Represents the trainable weight matrix and attention coefficients. Used to represent nodes For nodes The relative importance of each node is calculated by concatenating the features of its neighboring nodes:

[0129] ;

[0130] in, This represents the activation function. For the first The trainable weight matrix of the layer, Represents a node The neighborhood group, This represents a vector concatenation operation. The trainable weight vector in the attention mechanism represents the initial features of the node. It includes real-time operating data such as the node's voltage amplitude, load power, energy storage status, and switch status.

[0131] Simultaneously, to reflect the dynamic characteristics of system operating status changes over time, time-series modeling is performed on the historical operating data of the nodes. Let the nodes... The sequence of running states within a continuous time window is represented as follows:

[0132] ;

[0133] Each of them This includes information on the node's voltage, current, load, and distributed energy output at time t, where... This represents the length of the time window, used to describe the historical sequence range of the system's operating states. The self-attention mechanism in the Transformer structure is used to encode the above sequence to extract key temporal features, expressed as follows:

[0134] ;

[0135] in , , These are the query matrix, key matrix, and value matrix obtained by linearly mapping the input sequence, respectively:

[0136] ;

[0137] in , , For trainable weight matrix, This represents the dimension of the key vector. A self-attention mechanism is used to calculate the correlation weights between states at different time steps, thereby highlighting historical operational characteristics that have a significant impact on the current decision.

[0138] Finally, topological features With time series characteristics The nodes are fused to form a node state feature vector. The fusion method employs vector concatenation followed by mapping through a fully connected layer to ensure that topological and temporal information are fully represented in the same feature space.

[0139] ;

[0140] in This represents a fully connected mapping function. This represents the feature vector concatenation operation. For nodes The system-level state vector is obtained by fusing state features, which simultaneously contains topological information about the nodes and key dynamic temporal information. Furthermore, the features of all nodes are aggregated using a readout function to form a system-level state vector.

[0141] ;

[0142] in, This represents the set of nodes in the distribution network diagram structure. This represents the total number of nodes. The system-level state vector... It serves as a unified input to the strategy networks of both the master and slave sides, supporting subsequent master-slave game decision-making.

[0143] In this state Based on this representation, strategy functions for the master and slave sides are constructed. For the master distribution network operator, a strategy model is constructed using Dueling Double Deep Q Network (D3QN), and its strategy function is expressed as follows:

[0144] ;

[0145] Where parameters The initialization mechanism of the expert knowledge-guided strategy constructed in step three is then initialized. The principal's optimization objective is to maximize the overall system recovery benefit, and its reward function... Taking into account factors such as load recovery ratio, system power loss, and grid operation safety constraints, the strategy generation process involves the master party determining the current system state. Select topology adjustment and resource requirement actions from the set of executable actions that satisfy the safety action constraint mechanism described in step three. And through the policy function The resource response behavior of the slave is predicted to assess the impact of different master actions on the system's operating state and recovery benefits, and to guide the fault recovery process. The master strategy optimization objective function is defined as:

[0146] ;

[0147] For the slave resource aggregator, this invention employs multiple Deep Q Networks (DQNs) to learn strategies for different types of adjustable resources, including decision variables such as distributed power generation output regulation, energy storage charging and discharging control, and controllable load regulation. From a decision-making perspective, the slave still participates in the master-slave game as a unified virtual power plant resource aggregator; in the specific solution process, corresponding sub-policy networks are set up for different types of resources for decomposition learning to improve the flexibility and solution efficiency of slave resource response modeling.

[0148] Its policy network is denoted as:

[0149] ;

[0150] Where parameters This represents the network parameters of the slave's policy. The slave observes the master's actions. Subsequently, resource response actions are generated from the set of feasible actions that meet the equipment operation constraints and the safety action constraint mechanism established in step three. This is to achieve collaborative support for the needs of the primary party. (Support function for the secondary party) It primarily characterizes the resource response effect and the degree to which operational constraints are met; its optimization objective is defined as:

[0151] ;

[0152] During the strategy-solving process, the master and slave players engage in iterative game play through an alternating update mechanism. Specifically, within each decision cycle, the master player first updates the strategy based on the current system state. Generate Actions Subsequently, the server generates a resource response action after observing the client's action. The environment updates the distribution network operating status based on the actions of both the master and slave parties and calculates the corresponding reward signals. Subsequently, the master and slave parties each draw from the hierarchical experience sample pool constructed in step three. According to sampling weight Training samples are extracted for experience replay, and the policy network parameters are iteratively optimized using the gradient descent algorithm, thereby enhancing the policy's learning ability in key operating scenarios and risky states.

[0153] Through the aforementioned master-slave reinforcement learning joint optimization process, the main body of distribution network operation and the main body of virtual power plant resource aggregation form a stable collaborative decision-making relationship under the premise of meeting the constraints of power grid operation safety and resource regulation capacity, thereby significantly improving the efficiency of distribution network fault recovery, power supply reliability and the robustness of dispatching strategies.

[0154] S5, Collaborative Response Execution and Closed-Loop Feedback;

[0155] On the distribution network side, a topology reconfiguration operation is performed. On the virtual power plant side, internal resource coordination and power support are achieved through a constraint optimization model. The grid operation status and resource output status after execution are fed back to the decision-making system, forming a closed-loop optimization control of "strategy generation - collaborative execution - status feedback", which continuously corrects the master-slave game decision-making strategy.

[0156] Coordinated execution: The distribution network side executes actions based on the master's actions. Performing the opening and closing operations of sectionalizing switches and tie switches enables topology reconfiguration. The virtual power plant side responds according to the actions of the slave side. The model achieves coordinated allocation of internal resources through a constrained optimization model. The model aims to minimize operating costs and unmet power of loads. The constraints include power balance, upper and lower limits of resource output, and energy storage state of charge restrictions, so as to achieve priority recovery of critical loads and coordinated support of multiple resources.

[0157] The constrained optimization model is as follows:

[0158] ;

[0159] ;

[0160] ;

[0161] in, This represents the set of adjustable resources within a virtual power plant, including resource units such as distributed power sources, energy storage devices, and controllable loads. This represents the set of load nodes participating in the recovery. This indicates the total power support demand of the fault region generated by the master-slave game decision. This represents the actual output power of resource i. and Let represent the minimum and maximum adjustable output boundaries of resource i, respectively. This represents the operating cost function or output deviation penalty function for resource i. Indicates load node The original load demand power, Indicates load node Unmet power, This is a load priority weighting coefficient used to reflect the requirement for priority restoration of critical loads. Indicates energy storage resources At any moment The state of charge, and These represent the permissible ranges for the state of charge of energy storage;

[0162] Closed-loop feedback: Data such as the operation results of distribution network switches, the output status of virtual power plant resources, and the charging status of energy storage are fed back to the master-slave game decision-making system in real time. The system updates the grid operation status based on the feedback information, which serves as the input for the next round of decision-making, forming a closed-loop optimization control mechanism to continuously correct decision-making strategies and improve the dynamic adaptability of fault recovery.

[0163] Example 2:

[0164] In addition, corresponding to the above method, the present invention also provides a human-machine integrated distribution network load restoration system based on a master-slave game mechanism, comprising:

[0165] Master-Slave Game Modeling Module: Used to construct a hierarchical game decision-making model between the distribution network master and the virtual power plant slave, defining the state space, action space and reward / payout function of both parties;

[0166] Resource dynamic identification module: used to obtain the real-time status of distributed resources in the virtual power plant, assess resource adjustability, screen and build a set of fault recovery support resources, and construct resource capability vectors and aggregated capability status variables;

[0167] Expert knowledge guidance module: This module is used to embed experiential and rule-based knowledge of the power system into the decision-making process, construct a joint initialization objective function, a hierarchical sample reinforcement mechanism, and a safety action constraint mapping function, thereby achieving the constraint and guidance of policy learning.

[0168] The feature extraction and reinforcement learning module is used to extract power grid topology and time-series features through graph attention network and Transformer and fuse them into a system-level state representation, construct a master-slave strategy model and optimize the collaborative decision-making strategy through game iteration.

[0169] Collaborative execution and closed-loop feedback module: used to control the topology reconfiguration operation on the distribution network side and the resource coordination and allocation on the virtual power plant side, collect the grid operation status and resource output status after execution and feed them back to the preceding module to realize the continuous correction and optimization of the strategy;

[0170] The above modules work together to achieve human-machine integrated load restoration after a power distribution network fault.

[0171] The feature extraction and reinforcement learning module includes a topological feature extraction unit, a temporal feature extraction unit, a feature fusion unit, and a master-slave game optimization unit. The topological feature extraction unit deploys a graph attention network model, the temporal feature extraction unit deploys a Transformer self-attention model, and the master-slave game optimization unit deploys a master-side Dueling Double Deep Q Network policy model and a slave-side multi-deep Q network sub-policy model, respectively.

[0172] The collaborative execution and closed-loop feedback module includes a power grid execution unit, a virtual power plant scheduling unit, and a status feedback unit. The virtual power plant scheduling unit has a built-in resource coordination and allocation constraint optimization model to realize the collaborative scheduling of distributed power sources, energy storage devices, and controllable loads. The status feedback unit is used to collect data such as the status of distribution network switches, node voltages, line power flow, virtual power plant resource output, and energy storage charge status in real time and update the system status.

[0173] This invention constructs a master-slave game-theoretic collaborative decision-making mechanism between the distribution network operator and the virtual power plant resource aggregator, enabling stronger multi-entity collaboration and higher load recovery efficiency after a distribution network fault. The invention models the distribution network operator as the master and the virtual power plant resource aggregator as the slave, achieving coordinated optimization between grid dispatching objectives and distributed resource operation objectives. During fault recovery, the master formulates switching operations and load recovery strategies based on the grid operating status and security constraints, while simultaneously predicting the slave's possible response behavior; the slave, after observing the master's instructions, responds with resource dispatch based on its own resource constraints. Through this master-slave game-theoretic collaborative decision-making mechanism, load transfer schemes that meet grid security constraints can be generated under complex fault scenarios, multiple types of distributed resource participation, and dynamic changes in network topology or load, thereby improving the stability of fault recovery decisions and multi-entity collaborative dispatching capabilities. Furthermore, this invention introduces an expert knowledge guidance mechanism into the master-slave game-theoretic decision-making process, integrating power system operating experience and dispatching rules into the strategy generation process. This allows reinforcement learning decision strategies to be optimized under the conditions of meeting grid physical constraints and operating procedures, further improving the engineering feasibility of the decision strategies and the system's operational security.

[0174] Based on this, the present invention achieves the above-mentioned technical advantages through four key technical modules: First, a power grid topology and timing feature extraction module, which extracts the topology and timing features of the power grid through joint modeling of graph neural networks and Transformers, realizing a comprehensive perception of the power grid's operating state after a fault, thereby providing an accurate state expression for subsequent decision-making strategy generation; Second, an expert knowledge-guided strategy initialization module, which embeds power system operating experience and scheduling rules into the reinforcement learning strategy learning process, constraining and guiding the strategy generation process, thereby improving strategy learning efficiency and avoiding unsafe operations that violate power grid operating procedures; Third, a master-slave game reinforcement learning decision-making module, which realizes the collaborative optimization between distribution network scheduling decisions and virtual power plant resource responses through a reinforcement learning game mechanism between the master and slave sides, enabling the system to gradually form a stable collaborative decision-making strategy in a dynamic operating environment; Fourth, a collaborative response and resource coordination execution module, which implements topology reconstruction operations on the distribution network side after strategy generation, and performs resource coordination scheduling on the virtual power plant side for distributed power sources, energy storage, and controllable loads, forming a closed-loop control process of "strategy generation - collaborative execution - state feedback" through state feedback. Through the synergistic effect of the above modules, the present invention can achieve efficient multi-entity collaboration and rapid load recovery under different fault scenarios and complex operating conditions, thereby significantly improving the fault recovery efficiency and system operation reliability of the distribution network.

[0175] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A human-machine integrated load restoration method for power distribution networks based on a master-slave game mechanism, characterized in that, The specific steps include: S1. Modeling the master-slave game environment: Constructing a hierarchical master-slave game decision-making model; S2. Dynamic identification and assembly of available resources in virtual power plant: Obtain the real-time operating status of each distributed resource in the virtual power plant, evaluate the adjustability of various resources, determine and screen the participation of resources according to the power supply restoration needs of the fault area, assemble a set of supporting resources that can participate in fault restoration, and construct resource capability vectors and virtual power plant aggregation capability state variables. S3. Construct an expert knowledge constraint and guidance mechanism for master-slave game decision-making: embed power system experience-based knowledge and rule-based knowledge into the decision-making process, construct a unified decision state vector under the fault recovery scenario, construct an expert demonstration sample set and a joint initialization objective function, establish a hierarchical sample reinforcement mechanism, and construct a safety action constraint mapping function to reconstruct the set of executable actions. S4. Master-Slave Game Reinforcement Learning: Graph attention network is used to extract the topological features of the distribution network, and Transformer structure is used to extract the operation sequence features of the power grid. The system-level state representation is obtained by fusion. Master-slave strategy models are constructed respectively, and the collaborative decision-making strategy is optimized through the game iteration mechanism of alternating master and slave updates. S5. Coordinated Response Execution and Closed-Loop Feedback: Based on the master-slave game decision results generated in S4, the distribution network operation side and the virtual power plant side execute the corresponding control commands respectively, thereby forming a coordinated response execution mechanism in the fault recovery process.

2. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that: In S1, the main entity operating the distribution network is considered the master, and the entity aggregating virtual power plant resources is considered the slave. The decision model of the master in the master-slave game decision-making model is expressed as follows: ； in, This represents the decision-making model of the principal. Represents the master's state space. Indicates the action space of the main party. Indicates the state Next action Then transferred to state State transition function, The main action, This refers to the resource response actions taken by the subordinate party in response to the demands of the primary party. This represents the principal reward function, used to measure the overall impact of the current decision on system recovery and safe operation. This is a discount factor used to balance immediate recovery effects with long-term performance.

3. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 2, characterized in that: In S1, the principal reward function is specifically as follows: ； in, The main action, This refers to the resource response actions taken by the subordinate party in response to the demands of the primary party. This indicates the percentage of system load recovery under the current decision. Indicates the system's operating power loss. This indicates the degree of violation of constraints such as voltage exceeding limits and line overload. These are non-negative weighting coefficients used to adjust the relative importance of each objective in the optimization process.

4. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that: In S2, the conditions for determining the availability of resources include: there is a feasible power supply path between the node where the resource is located and the node in the fault area, and the resource has electrical reachability; The resource has adjustable power capacity in its current operating state; Resources meet the operational constraints of equipment capacity, energy storage state of charge, and load regulation range.

5. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that: In S3, the joint initialization objective function includes a recovery benefit fitting term, an expert preference correction term, and a parameter constraint term. Specifically, the joint initialization objective function is as follows: ； in, This represents the overall optimization objective during the initialization phase. This represents the recovery payoff fitting term, used to ensure that the strategy evaluation result is consistent with the recovery payoff after the state transition. This indicates an expert preference correction item, used to increase the priority of the expert's demonstrated action in the corresponding state. This represents parameter constraint terms, used to limit model complexity and improve generalization ability. and These are the weighting coefficients.

6. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that: In S3, the safety action constraint mapping function is: ； Reconstruct the set of executable actions based on this mapping function. ; in, Used to determine actions In state The following actions are considered permissible safety actions, particularly when they result in the formation of prohibited loops, multiple paths between substations, isolated power outage loads, or unqualified distributed power source islands. The value is 0.

7. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that: In S4, when integrating the distribution network topology and temporal characteristics to obtain the system-level state representation, a graph attention network is used to analyze the distribution network graph structure. The node topology features are obtained by encoding. The Transformer self-attention mechanism is used to encode the running state sequence of the node in a continuous time window to obtain the temporal features. The topology features and temporal feature vectors are concatenated and mapped through a fully connected layer to obtain the node fusion state features. Finally, the readout function aggregates all node features to form a system-level state representation.

8. The method for load restoration of a distribution network based on a master-slave game mechanism according to claim 1, characterized in that, In S4, the game-theoretic iterative mechanism of alternating master-slave updates is specifically as follows: Within each decision cycle, the master party generates master party actions based on the current system-level state representation, and the slave party generates resource response actions after observing the master party actions; The distribution network environment updates its operating status and calculates reward signals based on the actions of both the master and slave parties; The master and slave sides extract training samples from the hierarchical experience sample pool according to the sampling weight and perform experience playback. They then iteratively optimize the network parameters of their respective strategies using the gradient descent algorithm.

9. A system applicable to the human-machine integrated distribution network load restoration method based on master-slave game mechanism as described in any one of claims 1-8, characterized in that, include: Master-Slave Game Modeling Module: Used to construct a hierarchical game decision-making model between the master of the distribution network and the slave of the virtual power plant, defining the state space, action space and reward / payout function of both parties; Resource dynamic identification module: used to obtain the real-time status of distributed resources in the virtual power plant, assess resource adjustability, screen and build a set of fault recovery support resources, and construct resource capability vectors and aggregated capability status variables; Expert knowledge guidance module: This module is used to embed experiential and rule-based knowledge of the power system into the decision-making process, construct a joint initialization objective function, a hierarchical sample reinforcement mechanism, and a safety action constraint mapping function, thereby achieving the constraint and guidance of policy learning. Feature Extraction and Reinforcement Learning Module: Used to extract power grid topology and time-series features through graph attention networks and Transformers and fuse them into system-level state representations, construct master-slave strategy models and optimize collaborative decision-making strategies through game theory iteration; Collaborative execution and closed-loop feedback module: used to control the topology reconfiguration operation on the distribution network side and the resource coordination and allocation on the virtual power plant side, collect the grid operation status and resource output status after execution and feed them back to the preceding module to realize the continuous correction and optimization of the strategy.

10. A human-machine integrated power distribution network load restoration system based on a master-slave game mechanism according to claim 9, characterized in that, The feature extraction and reinforcement learning module includes a topological feature extraction unit, a temporal feature extraction unit, a feature fusion unit, and a master-slave game optimization unit; The topology feature extraction unit deploys a graph attention network model, the temporal feature extraction unit deploys a Transformer self-attention model, and the master-slave game optimization unit deploys a master-side Dueling Double DeepQ Network strategy model and a slave-side multi-deep Q network sub-strategy model, respectively. The collaborative execution and closed-loop feedback module includes a power grid execution unit, a virtual power plant scheduling unit, and a status feedback unit. The virtual power plant scheduling unit has a built-in resource coordination and allocation constraint optimization model, which is used to realize the collaborative scheduling of distributed power sources, energy storage devices, and controllable loads. The aforementioned status feedback unit is used to collect data such as the status of distribution network switches, node voltage, line power flow, virtual power plant resource output, and energy storage charge status in real time and update the system status.

Images