Distribution network architecture self-adaptive reconstruction method, system and equipment based on distributed flexible control, and medium

Abstract

The invention relates to the technical field of power distribution network reconstruction, and discloses a power distribution network architecture adaptive reconstruction method, system and device based on distributed flexible control, and a medium, and the method comprises the steps: obtaining the operation data and state data of a power distribution network in real time; dividing the power distribution network into a plurality of control areas, and deploying a control agent in each control area; on the basis of the system state information, if it is detected that operation constraint violation is conducted, adjustment is conducted preferentially through the flexible control device; if the operation constraint is still violated after the flexible adjustment, each control agent generates a candidate net rack reconstruction scheme based on a reinforcement learning model; through information interaction among multiple agents, distributed coordination is carried out on candidate network frame reconstruction schemes, feedback information is generated based on an execution result to reinforce a learning model, centralized optimization is avoided, a reconstruction decision can be completed and executed within second-level time, dynamic adaptive adjustment of a power distribution network architecture is realized, and the network frame reconstruction efficiency is improved. The system is ensured to operate in an optimal or suboptimal state, and has good expandability.

Description

Technical Field

[0001] This invention relates to the field of distribution network reconfiguration technology, and in particular to an adaptive reconfiguration method, system, device and medium for distribution network architecture based on distributed flexible control. Background Technology

[0002] With the large-scale integration of distributed generation into the distribution network, the operating characteristics of the distribution network have changed significantly. The randomness and volatility of power output and the time-varying characteristics of loads have placed higher demands on system optimization. Traditional distribution network reconfiguration techniques mainly change the network topology by adjusting switch states to achieve goals such as reducing network losses, improving voltage distribution, and increasing reliability. In deterministic operating environments with low distributed generation penetration, existing reconfiguration methods typically employ offline optimization, making scheduling decisions on a day-ahead or week-ahead basis, which can basically meet the system's operational needs.

[0003] However, existing distribution network reconfiguration technologies still have shortcomings. These are mainly manifested in their static optimization characteristics, making them unsuitable for the dynamic operating environment under high distributed generation penetration. When actual operating conditions deviate from predicted scenarios, pre-defined reconfiguration schemes may lose their optimality, even leading to violations of constraints such as voltage exceeding limits or branch overload. Furthermore, existing methods often employ a centralized optimization architecture, requiring the collection of network-wide data and complex calculations at a central location. This results in significant communication and computational delays, slow response times, and difficulty in meeting the real-time control requirements of rapidly changing distribution networks, especially in fault conditions where rapid load transfer or islanding operation is impossible. In addition, traditional reconfiguration primarily focuses on network topology adjustments, neglecting the coordinated utilization of flexible control resources such as energy storage systems, flexible interconnection devices, and distributed generation inverters. Typically, network reconfiguration and flexible control are optimized separately, failing to fully leverage their complementary advantages and limiting the overall system control efficiency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an adaptive reconfiguration method and system for distribution network architecture based on distributed flexible control to solve the technical problems of current static optimization, slow response, and insufficient coordination in planning and operation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an adaptive reconfiguration method for distribution network architecture based on distributed flexible control, comprising: Real-time acquisition of operation and status data of the power distribution network to form system status information; The power distribution network is divided into multiple control zones, and control agents are deployed in each control zone; Based on the system state information, it is detected whether the operating constraints are violated; if the operating constraints are violated, the flexible control device is used to make adjustments first; if the operating constraints are still violated after the flexible control adjustment, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model. The candidate network reconfiguration scheme is distributed and coordinated through information interaction among multiple agents to obtain the coordinated reconfiguration scheme. The reconstruction scheme is executed, and feedback information is generated based on the execution results. The reinforcement learning model is then optimized using the feedback information.

[0007] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, the distribution network is divided into multiple control areas, and a control agent is deployed in each control area, including: The region is divided by clustering based on electrical distance, and the adjacency relationship between regions is established by using electrical distance as the clustering metric. A control agent responsible for local data processing and decision-making is deployed in each region. After the agents are deployed, the system establishes the communication topology between regions and defines the adjacency matrix. Indicates the connection relationship between regions: In the formula, For elements of the adjacency matrix, and This is the area code.

[0008] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, wherein: the preferential adjustment via flexible control equipment includes: Calculate the adjustment requirements for constraint violations and search for available flexible control resources; Select a flexible device and issue an adjustment command based on the priority of adjustment speed and cost. If the constraint is eliminated after adjustment, the current process ends.

[0009] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, if the operational constraints are still violated after flexible control adjustment, each control agent generates candidate network reconfiguration schemes based on a reinforcement learning model, including: State space information is generated based on system state information; Using the state space as input, actions including switching operations are selected using a pre-trained deep reinforcement learning strategy. A reward function that comprehensively considers network loss, voltage deviation, and switching operation costs; The agent calculates the Q-value of each action based on the state space information through a neural network and selects the action with the largest Q-value as the reconstruction scheme.

[0010] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, the candidate network reconfiguration schemes are distributedly coordinated through information interaction among multiple agents to obtain a coordinated reconfiguration scheme, including: The coordination variables are updated iteratively using a consensus protocol; The updated coordination variables are used as Lagrange multipliers to introduce a local objective function adjustment scheme until the coordination variables of each agent tend to be consistent and satisfy the global constraints.

[0011] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, the step of executing the reconfiguration scheme and generating feedback information based on the execution result includes: Perform switching operations sequentially to ensure the radial topology is maintained; After execution, operational and status data are re-collected, the network loss reduction rate after reconstruction is calculated, the reconstruction effect is evaluated, and feedback information is generated.

[0012] As a preferred embodiment of the adaptive reconfiguration method for distribution network architecture based on distributed flexible control described in this invention, wherein: optimizing the reinforcement learning model using the feedback information includes: The current state, action, reward, and new state are combined into an experience tuple for storage. This tuple is used for experience replay training to update the reinforcement learning model parameters, thereby achieving continuous adaptive optimization of the model.

[0013] Secondly, the present invention provides an adaptive reconfiguration system for a distribution network architecture based on distributed flexible control, comprising: The acquisition module is used to acquire real-time operating data and status data of the power distribution network to form system status information; The agent allocation module is used to divide the power distribution network into multiple control areas and deploy control agents in each control area; The detection module is used to detect whether the operating constraints are violated based on the system state information; if the operating constraints are violated, the flexible control device is used to make adjustments first; if the operating constraints are still violated after the flexible control adjustment, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model. The coordination module is used to perform distributed coordination of the candidate network reconstruction scheme through information interaction among multiple agents to obtain the coordinated reconstruction scheme. The execution feedback module is used to execute the reconstruction scheme, generate feedback information based on the execution results, and optimize the reinforcement learning model using the feedback information.

[0014] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of an adaptive reconfiguration method for a power distribution network architecture based on distributed flexible control.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the aforementioned adaptive reconfiguration method for a power distribution network architecture based on distributed flexible control.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By establishing an online decision-making model based on reinforcement learning, this invention can autonomously learn the optimal reconfiguration strategy based on real-time operational data, without the need for manual pre-setting of rules or offline optimization calculations. When distributed power output fluctuates, load changes, or faults occur, reconfiguration decisions can be completed and executed within seconds, realizing dynamic adaptive adjustment of the distribution network architecture and ensuring that the system always operates in an optimal or suboptimal state. A joint optimization model of the grid structure and flexible control equipment is established, attempting to respond quickly to operational changes through flexible control. When the flexible adjustment capability is insufficient, grid reconfiguration is initiated, achieving hierarchical coordination between the two. The collaborative mechanism can leverage the rapid response advantage of flexible control and minimize the number of switching actions, thereby improving system operating efficiency. A multi-agent distributed control architecture is adopted, where each agent achieves global coordination through limited information interaction, avoiding centralized optimization and exhibiting good scalability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of an adaptive reconfiguration method for distribution network architecture based on distributed flexible control, according to an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Example 1, referring to Figure 1 As an embodiment of the present invention, an adaptive reconfiguration method for distribution network architecture based on distributed flexible control is provided, comprising: S100: Real-time acquisition of operation and status data of the distribution network to form system status information; S200: Divide the power distribution network into multiple control areas and deploy control agents in each control area; S300: Based on the system state information, detect whether the operating constraints are violated; if the operating constraints are violated, adjust them first through the flexible control device; if the operating constraints are still violated after the flexible control adjustment, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model. S400: Distributed coordination of the candidate network reconfiguration scheme is performed through information interaction among multiple agents to obtain a coordinated reconfiguration scheme; S500: Execute the reconstruction scheme, generate feedback information based on the execution result, and optimize the reinforcement learning model using the feedback information.

[0021] Specifically, this invention constructs an adaptive reconfiguration method based on real-time operation data acquisition of the distribution network, a mixed-integer programming model, reinforcement learning algorithms, and distributed control theory using S100-S500. First, operational data such as voltage, current, and power at each node, as well as status information of flexible control equipment, are collected through distribution network monitoring terminals. Then, the distribution network is divided into multiple control zones, with one control agent deployed in each zone. Each agent trains for reconfiguration decisions using a reinforcement learning model based on its own zone's operational data and interaction information with neighboring zones. In actual operation, the agent selects the optimal action based on the current state, including switching operations and flexible equipment adjustment commands. The system adopts a hierarchical control strategy, first attempting to eliminate constraint violations through flexible control, and then performing network reconfiguration when the flexible adjustment capability is insufficient. All agents maintain global coordination through a consensus protocol to ensure that local optimization decisions conform to the overall system objectives.

[0022] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, an adaptive reconfiguration method for distribution network architecture based on distributed flexible control is provided. The method includes: S100: Real-time acquisition of operation and status data of the distribution network to form system status information; Specifically, the collected data may include: the three-phase voltage amplitude of each node. Phase angle Three-phase current Active power reactive power and the output status of flexible control equipment. State of charge of energy storage system Switch status wait.

[0023] In one optional implementation, S100 can collect real-time operating data from each node of the distribution network via distribution automation terminals (DTUs), smart meters, flexible interconnection devices, and distributed power monitoring systems. The data acquisition frequency is set to once per second to ensure that the system can promptly detect changes in operating status.

[0024] In another alternative implementation, based on the above implementation, to ensure data reliability, the collected data can also be quality-verified. First, the consistency of the data timestamps is checked, requiring that the timestamp deviation of each measuring point at the same time does not exceed 100 milliseconds. Data exceeding this range is marked as abnormal and re-acquired. Then, the data rationality is verified; the voltage data should meet the following requirements. The current data should meet the requirements. ;in, Rated voltage, This represents the line's rated current. Data that does not meet the rationality criteria is discarded and replaced with linear interpolation of data from adjacent time points.

[0025] Specifically, in S100, system state information is generated based on the collected operational data, and a state vector is constructed. , indicating time System operating status: In the formula, For nodes At any moment voltage, branch road The current, and They are nodes Active power and reactive power, This represents the state of switch k (value 0 or 1, where 0 indicates open and 1 indicates closed). For flexible equipment Apparent power For energy storage systems The state of charge.

[0026] It should be noted that the state vector The information is passed to the S200 for processing as input for subsequent decisions. Through real-time status awareness, problems such as voltage exceeding limits, branch overload, and power imbalance during operation can be detected in a timely manner, thus providing a data foundation for adaptive reconfiguration.

[0027] S200: Divide the power distribution network into multiple control areas and deploy control agents in each control area; It should be noted that, in order to achieve distributed control, the distribution network is divided into multiple relatively independent control areas. The division of areas follows these principles: each area contains several nodes and branches, the electrical connections within the area are close, and the areas are connected by tie switches; the load capacity of each area is relatively balanced to avoid overloading of any single area; the area boundary is set at the branch point of the main line to facilitate power exchange and fault isolation.

[0028] In this embodiment of the application, step S200 divides the power distribution network into multiple control areas and deploys a control agent in each control area, including the following steps A1-A2: A1: Based on electrical distance, a clustering method is used to divide the region, with electrical distance as the clustering metric, to establish adjacency relationships between regions, and to deploy control agents responsible for local data processing and decision-making in each region; Specifically, a clustering method based on electrical distance is adopted; nodes are defined. and nodes The electrical distance between them is expressed as: In the formula, For nodes and nodes electrical distance, The equivalent impedance between the two nodes is calculated using the distribution network impedance matrix. Electrical distance reflects the electrical coupling strength between nodes; a smaller distance indicates stronger coupling.

[0029] In one alternative implementation, A1 can employ the K-means clustering algorithm, using electrical distance as the clustering metric, to divide the distribution network nodes into... Number of regions. The principle is that each region should contain between 15 and 30 nodes, ensuring that the optimization problem within the region can be solved quickly while avoiding excessive coordination complexity due to too many regions. For a typical 110-node distribution network, it is usually divided into 4 to 6 regions.

[0030] In another alternative implementation, A1 can also consider using a hierarchical clustering algorithm. Specifically, using electrical distance as the clustering metric, nodes or subclusters with the tightest electrical coupling are gradually merged from the bottom up until a preset number of regions or inter-cluster distance threshold is reached. This method does not require pre-specifying the precise number of regions; instead, it flexibly determines the partitioning hierarchy by constructing a tree-like clustering structure, making it suitable for scenarios with complex distribution network topologies and uneven electrical coupling strength.

[0031] Furthermore, A1 deploys control agents responsible for local data processing and decision-making in each region. Specifically, this could be done in each region... Deploy a control agent Intelligent agents are carried by edge computing devices and deployed in regional power distribution rooms or switching stations. Responsible for collecting operational data in this area (From the state vector of S100) Extracting regions (part of the process) performs local decision-making calculations and exchanges boundary information with neighboring agents. Boundary information includes key parameters such as the voltage of the regional boundary nodes and the power flowing through the tie lines, and the information exchange frequency is once per second.

[0032] A2: After the agents are deployed, the system establishes the communication topology between regions and defines the adjacency matrix. Indicates the connection relationship between regions: In the formula, For elements of the adjacency matrix, and This is the area code.

[0033] It should be noted that the adjacency matrix Used for subsequent distributed coordination and optimization by the S400, it determines the information interaction objects for each agent. The region division results and agent deployment information are transmitted to the S300 as the execution unit for flexible control response.

[0034] S300: Based on the system state information, detect whether the operating constraints are violated; if the operating constraints are violated, adjust them first through the flexible control device; if the operating constraints are still violated after the flexible control adjustment, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model. Specifically, operating constraints include voltage constraints and branch current constraints. First, determine if any constraints are violated. The voltage constraint check condition is: In the formula, The minimum permissible voltage, The maximum permissible voltage, This is the rated voltage (10kV or 0.4kV). When the voltage at any node... satisfy or When this occurs, it is determined that the voltage exceeds the limit.

[0035] The branch current constraint check conditions are: In the formula, branch road At any moment The current, Let k be the rated current carrying capacity of branch k. If any branch does not meet this condition, it is determined that the branch is overloaded.

[0036] In this embodiment of the application, the adjustment in step S300, which prioritizes the use of a flexible control device, includes the following steps A3-A4: A3: Calculate the adjustment requirements for constraint violations and search for available flexible control resources; Specifically, when a constraint violation is detected, the system calculates the required adjustment. For voltage exceeding limits, the voltage deviation is calculated: In the formula, For nodes Voltage deviation, The target voltage is (usually the rated voltage). Based on the voltage regulation sensitivity, the required reactive power to be provided by the flexible equipment is determined. ,in This is the reactive power sensitivity coefficient to voltage, calculated based on distribution network parameters, with a typical value of 0.02 to 0.05 (unit: kV / Mvar).

[0037] Specifically, based on the above implementation methods, the system is located in the region. Internally search for available flexible control resources and establish a resource inventory. For energy storage systems, check their state of charge. Is it within the permitted range? ,generally , ), and remaining adjustable capacity ,in This is the rated power of the energy storage. Provide power for the current situation. For flexible interconnect devices, check whether their apparent power has reached the capacity limit.

[0038] A4: Select flexible equipment and issue adjustment instructions based on adjustment speed and cost priority. If the constraint is eliminated after adjustment, the current process ends.

[0039] For example, energy storage systems and flexible interconnect devices, with their fast response times (milliseconds), are prioritized; distributed power inverters are next; and reactive power compensation devices are last. The system issues adjustment commands to the selected flexible devices. After the flexible control is executed, the system re-collects operating data and checks whether the constraints have been eliminated. If the constraints have been eliminated, the adjustment ends, grid reconfiguration is not performed, and the system state remains stable. Furthermore, if the constraints are still not eliminated, it indicates insufficient flexibility, requiring the initiation of network reconfiguration to transfer control to the subsequent reinforcement learning model. The execution results of flexible control (whether the constraints were successfully eliminated, the type and capacity of the flexible resources used) serve as feedback information for training the reinforcement learning model in subsequent steps.

[0040] In this embodiment of the application, if the operational constraints are still violated after the flexible control adjustment in step S300, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model, including steps B1-B4: B1: Generate state space information based on system state information; B2: Using the state space as input, select actions including switching operations using a pre-trained deep reinforcement learning strategy; B3: A reward function that comprehensively considers network loss, voltage deviation, and switching operation costs; B4: The agent calculates the Q-value of each action based on the state space information through the forward calculation of the neural network, and selects the action with the largest Q-value as the reconstruction scheme.

[0041] Specifically, in steps B1-B4, the Deep Q-Network (DQN) algorithm is used for decision training. The agent's state space S, action space A, and reward function R are defined. The state space S consists of the state vectors acquired in the first step. The structure includes information such as node voltages, branch currents, and flexible device status. Action space A represents executable combinations of switching operations, including opening a closed switch, closing a disconnected switch, and maintaining the current topology. For operations containing... In a distribution network with a single switch, the theoretical operating space is... However, in practice, considering radial constraints and operational feasibility, the number of effective actions is far less than this value.

[0042] Define a single-step reward function Taking into account network losses, voltage deviations, and switching operation costs: In the formula, For a moment Instant rewards System network loss power (unit: kW). For nodes Voltage, Rated voltage, This represents the number of switching actions performed at this moment. These are the weighting coefficients.

[0043] The method for determining the weighting coefficients is as follows: First, normalize the network loss, voltage deviation, and number of switching operations to ensure that the three values ​​are on the same order of magnitude; then, set the weights according to the priority of the operational objectives. If the primary objective is to reduce losses, then... , , If the primary goal is to improve voltage quality, then , , To reduce the number of switching operations, increase the speed appropriately. value.

[0044] The agent approximates the Q-value function using a neural network. This function represents the state. Next action The long-term cumulative reward expectation is determined by the neural network, which employs a three-layer fully connected structure. The number of nodes in the input layer equals the dimension of the state vector, the number of nodes in the hidden layer is 128, and the number of nodes in the output layer equals the size of the action space. The training process uses an experience replay mechanism, storing the experience tuples (state, action, reward, next state) generated by the agent's interaction with the environment in an experience pool. Each time, a batch of samples is randomly selected from the experience pool for training, avoiding training instability caused by sample correlation.

[0045] During training, the agent adjusts its behavior based on the current state. And the Q-value of the neural network output to select actions .use - Greedy strategy balance exploration and utilization: using probability Randomly select an action (exploration) with probability. Choose the action with the highest Q value (exploit). Initially =1.0, gradually decreasing to 0.1 as training progresses, allowing the agent to transition from full exploration to full utilization of learning outcomes. Execute actions. Afterwards, the environment shifted to a new state. The intelligent agent receives a reward. Update the Q value according to the Bellman equation: In the formula, The Q-value of the state-action pair. The learning rate (value 0.001). This is the discount factor (value 0.95). The maximum Q-value among all possible actions in the next state is obtained; the neural network parameters are updated through backpropagation to make the predicted Q-value approximate the target Q-value.

[0046] Training is conducted offline, using historical runtime data and diverse scenarios generated through simulation. After training converges, the agent learns a mapping strategy from state to action. It can quickly make reconstruction decisions based on the current operating status. During online operation, the agent uses the real-time status data collected in the first step... The Q-value of each action is obtained through forward computation using a neural network, and the action with the largest Q-value is selected as the reconstruction scheme. The decision result (the selected sequence of switching operations) is then passed to the fifth step for distributed coordination and verification.

[0047] It should be noted that in steps B1-B4, when the flexible control fails to eliminate the constraint violation, the system initiates the reconstruction decision module based on reinforcement learning. Reinforcement learning learns the optimal strategy through interaction with the environment, without the need to establish an accurate system model in advance, and is particularly suitable for handling the complex dynamic characteristics of the distribution network.

[0048] S400: Distributed coordination of the candidate network reconfiguration scheme is performed through information interaction among multiple agents to obtain a coordinated reconfiguration scheme; It should be noted that, since the distribution network is divided into multiple regions, independent reconfiguration decisions by agents in each region may lead to locally optimal but globally suboptimal results. Therefore, a distributed coordination mechanism needs to be established in S400 to ensure that the decisions of each agent satisfy local optimization while guaranteeing the achievement of the global objective.

[0049] In this embodiment of the application, step S400 involves distributed coordination of the candidate network reconfiguration scheme through information interaction among multiple agents to obtain the coordinated reconfiguration scheme, including the following steps C1-C2: C1: Iteratively update the coordination variables using a consensus protocol; Specifically, in C1, each agent maintains a coordination variable. , indicating region Marginal cost estimation for the global objective. Coordination variables are updated iteratively to reach consensus: In the formula, Let z be the coordination variable for region z at iteration time t. To coordinate the step size (value 0.1), For the region The set of neighboring areas, The adjacency matrix elements defined in the second step.

[0050] It should be noted that the update formula of C1 can gradually make the coordination variables of each region more consistent, thus achieving a global consensus.

[0051] C2: The updated coordination variables are used as Lagrange multipliers to introduce a local objective function adjustment scheme until the coordination variables of each agent tend to be consistent and satisfy the global constraints.

[0052] Specifically, in each iteration, the agent Based on the updated coordination variables The reconstruction decision for this region is adjusted. The coordination variable is introduced as a Lagrange multiplier into the local optimization objective function, allowing local decisions to consider global constraints. After several iterations (typically 5 to 10), the coordination variables of each agent converge to a consistent value. At this point, the restructuring decisions in each region have reached a state of global coordination.

[0053] After coordination is complete, check whether the global constraints are met. Global constraints include indicators such as total network loss and total voltage deviation. If the global constraints are met, the reconstruction scheme is confirmed to be effective, and the execution step is initiated; if the global constraints are still not met, return to S300, where the reinforcement learning module regenerates candidate schemes until a feasible solution is found or the maximum number of iterations is reached.

[0054] The convergence speed of distributed coordination depends on the coordination step size. And network topology. Too small a step size will lead to slow convergence, while too large a step size may cause oscillations. By adaptively adjusting the step size, the system can complete coordination within about 5 iterations. The coordination result (i.e., the switching operation instructions confirmed by each region) is passed to subsequent steps.

[0055] S500: Execute the reconstruction scheme, generate feedback information based on the execution result, and optimize the reinforcement learning model using the feedback information.

[0056] In this embodiment of the application, the execution of the reconstruction scheme and the generation of feedback information based on the execution result in step S500 includes the following steps D1-D2: D1: Perform switching operations in sequence to ensure the radial topology is maintained; It should be noted that the reconfiguration plan, confirmed through distributed coordination, is distributed to the corresponding switching equipment by the intelligent agents in each region. Switching operations are performed according to the principle of disconnecting before closing, ensuring that the distribution network always maintains a radiating state.

[0057] For example, the specific execution process of D1 is as follows: First, disconnect the switch that needs to be disconnected and wait for the circuit breaker to fully open (approximately 50 to 100 milliseconds); then close the switch that needs to be closed and wait for the circuit breaker to fully close (approximately 50 to 100 milliseconds). To avoid islanding or ringing in the distribution network due to simultaneous operation of multiple switches, the system sorts the switch operations, operating only one switch at a time, and executing the next action only after the previous action is completed and confirmed.

[0058] D2: After execution, re-collect operational and status data, calculate the network loss reduction rate after reconstruction, evaluate the reconstruction effect, and generate feedback information.

[0059] Specifically, after the switch operation is completed, new operating data is collected immediately to check the reconstruction effect.

[0060] For example, the collected data includes indicators such as node voltage, branch current, and network loss power, which are compared with the data before reconstruction. The network loss reduction rate after reconstruction is calculated and expressed as: In the formula, To reduce network loss rate, For the network loss of the system before reconstruction, This refers to the network loss of the reconstructed system. Simultaneously, it checks whether all constraints are met, including whether the voltage is within acceptable limits. to Within the range (as defined in S300), does the branch current not exceed... (As defined in S300).

[0061] In this embodiment of the application, the step S500, which involves optimizing the reinforcement learning model using the feedback information, includes step D3: D3: Store the current state, action, reward, and new state as an experience tuple for experience replay training to update the reinforcement learning model parameters, so as to achieve continuous adaptive optimization of the model.

[0062] Specifically, the reconstruction effect evaluation results serve as feedback information for the continuous optimization of the reinforcement learning model. The system will record the status of this reconstruction. ,action ,award (Calculated based on the S400 reward function), new state These are used to construct experience tuples, which are then stored in the experience pool.

[0063] During the background training process, samples are periodically drawn from the experience pool to update the neural network parameters, enabling the reinforcement learning model to continuously adapt to changes in the operating characteristics of the power distribution network and improve decision-making accuracy.

[0064] In one alternative implementation, if constraint violations persist after refactoring, an emergency handling procedure is initiated. First, an attempt is made to further utilize flexible control resources (returning to S300). If all flexible resources are utilized and the constraint still cannot be eliminated, load shedding or distributed generation power limiting control is executed to ensure the safe operation of the distribution network.

[0065] In another alternative implementation, based on the above implementation, the operating status of the distribution network is continuously monitored. When a new operating change or constraint violation is detected, a new round of reconfiguration decision-making process is automatically triggered, and S100-S500 are executed cyclically to achieve dynamic adaptive adjustment of the distribution network architecture.

[0066] In summary, this invention enables real-time adaptive adjustment of the distribution network topology, significantly improving system response speed. Through reinforcement learning online decision-making models and a distributed control architecture, reconfiguration decisions can be completed and executed within seconds, enhancing the system's ability to cope with distributed power output fluctuations, load changes, and faults. By constructing a collaborative optimization mechanism for network reconfiguration and flexible control, the combined benefits of various control resources can be fully utilized. A flexible control priority response mechanism is established, prioritizing the elimination of constraint violations using rapid response resources such as energy storage and flexible interconnection devices, and only performing network reconfiguration when flexible resources are insufficient, avoiding frequent switching operations. Combined with a distributed control architecture, the computational complexity and scalability issues of centralized optimization are addressed. Through regional division and multi-agent coordination, the large-scale distribution network optimization problem is decomposed into multiple smaller sub-problems, with each agent independently solving for the optimal solution in its region, achieving global coordination through a consensus protocol. Simultaneously, the distributed architecture improves system robustness, ensuring that communication failures in some areas do not affect the normal operation of other areas.

[0067] Example 3, referring to Tables 1-4, provides a simulation application scheme for an adaptive reconfiguration method for distribution network architecture based on distributed flexible control, to verify the feasibility and effectiveness of the present invention.

[0068] This example is an extended verification based on the IEEE 33-node distribution network standard example. This distribution network has a rated voltage of 10kV and includes 33 nodes, 32 branches, and 5 sectionalizing switches. It connects to 3 distributed photovoltaic power sources with a total capacity of 2MW, 2 energy storage systems with a total capacity of 1MWh / 0.5MW, and 1 flexible interconnection device with a capacity of 1MVA. The distribution network is divided into 3 control areas, with one control agent deployed in each area. A simulation model is built on the MATLAB / Simulink platform to simulate a 24-hour operating scenario, including changes in photovoltaic output from zero to rated power, load fluctuations from minimum to peak value, and 3 random fault conditions. The simulation step size is set to 1 second, generating 86,400 time points of operational data. The method of this invention is used for online adaptive reconfiguration, and compared with traditional timed reconfiguration methods and centralized optimization reconfiguration methods.

[0069] Table 1: Performance Comparison of Three Refactoring Methods

[0070] Regarding daily grid loss, traditional timed reconfiguration methods perform two reconfigurations daily at fixed times (6 AM and 6 PM), which cannot cope with the dynamic changes in daily load and photovoltaic output, resulting in higher grid losses during off-optimal periods. Centralized optimization reconfiguration methods perform full-network optimization and reconfiguration every hour, significantly reducing grid losses, but are computationally time-consuming and involve frequent switching actions. The method of this invention, through real-time monitoring and adaptive decision-making, performs reconfiguration only when necessary, significantly reducing the number of switching actions while maintaining low grid losses. The reduction in voltage limit violations is due to the priority response mechanism of flexible control, enabling the system to quickly adjust energy storage and flexible interconnection devices to cope with voltage fluctuations, avoiding 16 out of 18 voltage limit violations.

[0071] Table 2: Distributed Coordination Convergence Performance

[0072] As the number of regions increases, the number of iterations and convergence time of distributed coordination increase linearly, but remain within an acceptable range. Even when divided into 5 regions, the coordination time is only 3.8 seconds, much faster than the 156 seconds of centralized optimization. The coordination success rate decreases slightly due to the increased complexity caused by the increased number of regions, but it still remains above 95%. The maximum coordination deviation represents the maximum difference in coordination variables among regions; the smaller the value, the better the consistency. In the case of 5 regions, the deviation is 0.024, which is still within the acceptable range for engineering (less than 0.05).

[0073] Table 3: Effects of Reinforcement Learning Training

[0074] After 20,000 training rounds, the reinforcement learning model converged, with the average reward value increasing from an initial -125.3 to -12.1 (the negative reward value is due to the reward function being designed to minimize cost; the closer the value is to zero, the better the performance). The decision accuracy reached 97.8%, meaning the agent could choose the optimal or near-optimal action in 98% of scenarios. The constraint violation rate decreased to 1.4%, and these 1.4% violations were mainly concentrated in extreme conditions (simultaneous severe fluctuations in photovoltaic power and load), which were eliminated after a second adjustment. The exploration rate gradually decreased from 1.0 to 0.1; in the later stages of training, the agent mainly relied on the learned policy for decision-making, occasionally exploring to adapt to new scenarios.

[0075] Table 4: Typical Scenario Case Analysis

[0076] Case studies demonstrate that the system can flexibly select control strategies based on different scenarios. When a sudden surge in photovoltaic output at 08:15 caused voltage exceedance, the system resolved the issue simply by absorbing excess power through the energy storage system, with a response time of 0.4 seconds and no grid reconfiguration required. At 14:45, under complex conditions of peak load and low photovoltaic output, the system's flexible control capability was insufficient, and it automatically initiated grid reconfiguration, mitigating overload by changing the topology; the entire process was completed in 2.8 seconds. In a fault scenario at 18:20, the system quickly isolated the faulty branch and reconfigured the grid to achieve load transfer, with a response time of 1.2 seconds, significantly faster than the several minutes required by traditional methods. The grid loss optimization and reconfiguration during a light load period at 22:10 shows that the system can not only handle emergency situations but also continuously optimize economic operation.

[0077] Simulation results show that the method of the present invention is significantly superior to traditional methods in terms of reducing network loss, improving voltage quality, and increasing response speed, and has good practical value.

[0078] Example 4 illustrates a schematic scheme for an adaptive reconfiguration method for distribution network architecture based on distributed flexible control. It should be noted that the technical solution of this system for adaptive reconfiguration of distribution network architecture based on distributed flexible control is based on the same concept as the aforementioned method for adaptive reconfiguration of distribution network architecture based on distributed flexible control. Details not described in detail in this example can be found in the description of the aforementioned method for adaptive reconfiguration of distribution network architecture based on distributed flexible control.

[0079] This embodiment also provides an adaptive reconfiguration system for distribution network architecture based on distributed flexible control, including: The acquisition module is used to acquire real-time operating data and status data of the power distribution network to form system status information; The agent allocation module is used to divide the power distribution network into multiple control areas and deploy control agents in each control area; The detection module is used to detect whether the operating constraints are violated based on the system state information; if the operating constraints are violated, the flexible control device is used to make adjustments first; if the operating constraints are still violated after the flexible control adjustment, each control agent generates a candidate network reconstruction scheme based on the reinforcement learning model. The coordination module is used to perform distributed coordination of the candidate network reconstruction scheme through information interaction among multiple agents to obtain the coordinated reconstruction scheme. The execution feedback module is used to execute the reconstruction scheme, generate feedback information based on the execution results, and optimize the reinforcement learning model using the feedback information.

[0080] This embodiment also provides a computer device applicable to an adaptive reconfiguration of a distribution network architecture based on distributed flexible control, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the adaptive reconfiguration method for a distribution network architecture based on distributed flexible control as proposed in the above embodiment.

[0081] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements an adaptive reconfiguration method for a power distribution network architecture based on distributed flexible control, as proposed in the above embodiments.

[0082] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for adaptive reconfiguration of a power distribution network architecture based on distributed flexible control proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0083] From the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0084] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for adaptive reconfiguration of distribution network architecture based on distributed flexible control, characterized in that, include: Real-time acquisition of operation and status data of the power distribution network to form system status information; The power distribution network is divided into multiple control zones, and control agents are deployed in each control zone; Based on the system status information, it is detected whether the operating constraints are violated; if a violation of the operating constraints is detected, adjustment is preferentially made through flexible control equipment. If the operational constraints are still violated after the flexible control adjustment, each control agent generates candidate network reconstruction schemes based on the reinforcement learning model. The candidate network reconfiguration scheme is distributed and coordinated through information interaction among multiple agents to obtain the coordinated reconfiguration scheme. The reconstruction scheme is executed, and feedback information is generated based on the execution results. The reinforcement learning model is then optimized using the feedback information.

2. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 1, characterized in that, The power distribution network is divided into multiple control zones, and control agents are deployed in each control zone, including: The region is divided by clustering based on electrical distance, and the adjacency relationship between regions is established by using electrical distance as the clustering metric. A control agent responsible for local data processing and decision-making is deployed in each region. After the agents are deployed, the system establishes the communication topology between regions and defines the adjacency matrix. Indicates the connection relationship between regions: In the formula, For elements of the adjacency matrix, and This is the area code.

3. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 2, characterized in that, The preferred adjustment via flexible control equipment includes: Calculate the adjustment requirements for constraint violations and search for available flexible control resources; Select a flexible device and issue an adjustment command based on the priority of adjustment speed and cost. If the constraint is eliminated after adjustment, the current process ends.

4. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 3, characterized in that, If the operational constraints are still violated after adjustment by flexible control, each control agent generates candidate network reconstruction schemes based on a reinforcement learning model, including: State space information is generated based on system state information; Using the state space as input, actions including switching operations are selected using a pre-trained deep reinforcement learning strategy. A reward function that comprehensively considers network loss, voltage deviation, and switching operation costs; The agent calculates the Q-value of each action based on the state space information through a neural network and selects the action with the largest Q-value as the reconstruction scheme.

5. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 4, characterized in that, The candidate network reconfiguration schemes are distributed and coordinated through information interaction among multiple agents to obtain a coordinated reconfiguration scheme, including: The coordination variables are updated iteratively using a consensus protocol; The updated coordination variables are used as Lagrange multipliers to introduce a local objective function adjustment scheme until the coordination variables of each agent tend to be consistent and satisfy the global constraints.

6. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 5, characterized in that, The execution of the reconstruction scheme and the generation of feedback information based on the execution results include: Perform switching operations sequentially to ensure the radial topology is maintained; After execution, operational and status data are re-collected, the network loss reduction rate after reconstruction is calculated, the reconstruction effect is evaluated, and feedback information is generated.

7. The adaptive reconfiguration method for distribution network architecture based on distributed flexible control as described in claim 6, characterized in that, The step of optimizing the reinforcement learning model using the feedback information includes: The current state, action, reward, and new state are combined into an experience tuple for storage. This tuple is used for experience replay training to update the reinforcement learning model parameters, thereby achieving continuous adaptive optimization of the model.

8. An adaptive reconfiguration system for a distribution network architecture based on distributed flexible control, using the method described in any one of claims 1-7, characterized in that, include: The acquisition module is used to acquire real-time operating data and status data of the power distribution network to form system status information; The agent allocation module is used to divide the power distribution network into multiple control areas and deploy control agents in each control area; The detection module is used to detect whether the operating constraints are violated based on the system status information; if a violation of the operating constraints is detected, the flexible control device is used for adjustment first. If the operational constraints are still violated after the flexible control adjustment, each control agent generates candidate network reconstruction schemes based on the reinforcement learning model. The coordination module is used to perform distributed coordination of the candidate network reconstruction scheme through information interaction among multiple agents to obtain the coordinated reconstruction scheme. The execution feedback module is used to execute the reconstruction scheme, generate feedback information based on the execution results, and optimize the reinforcement learning model using the feedback information.

9. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the adaptive reconfiguration method for a power distribution network architecture based on distributed flexible control as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the adaptive reconfiguration method for a power distribution network architecture based on distributed flexible control as described in any one of claims 1 to 7.

Images