A method and device for collaborative optimization and control of multiple distribution areas in a distribution network adapting to topology changes

Abstract

This application provides a method and apparatus for collaborative optimization and control of multiple distribution substations in a distribution network that adapts to topology changes. The method includes: dividing the distribution network into clusters; constructing an optimization model based on the cluster division results; constructing a dynamic clustering Markov decision model for each substation as an agent; solving the dynamic clustering Markov decision model based on the optimization model to obtain the corresponding collaborative optimization and control strategy for multiple distribution substations, and controlling the distribution network according to the control strategy. This application combines dynamic cluster division and distributed autonomous optimization, uses the SLM algorithm to enhance the collaborative capability of substations and reduce the impact of cross-cluster power flow; this application adapts to topology changes by introducing a feature extraction network to optimize the generation of control strategies; and utilizes centralized training and decentralized execution to reduce communication dependencies, achieving efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

Description

Technical Field

[0001] This application relates to the field of power grid control, specifically to a method and device for collaborative optimization control of multiple distribution areas in a distribution network that adapts to topology changes. Background Technology

[0002] With the widespread integration of new energy sources and distributed resources, power distribution networks face numerous challenges. The randomness and volatility of intermittent distributed photovoltaic (PV) energy sources can easily lead to problems such as power flow reversal and voltage exceeding limits, threatening the stability and reliability of the power distribution network. Meanwhile, traditional voltage regulating equipment (such as grouped capacitors and on-load tap-changing transformers) cannot meet the real-time and high-precision requirements of modern power distribution networks due to their discrete regulation methods and long response times. In contrast, new reactive power resources such as photovoltaic inverters and static var compensators (SVCs), with their rapid response and high-precision regulation characteristics, have become ideal choices for optimizing voltage control in power distribution networks.

[0003] However, the dynamic characteristics of distribution networks (such as topology changes) and the lack of multi-source data in low-perception environments make it difficult for traditional centralized optimization methods to efficiently achieve voltage regulation. Existing mathematical optimization methods rely on accurate physical models and network parameters, such as using robust optimization and linearization techniques to solve non-convex problems and improve system voltage stability, but their requirement for model accuracy limits their application in practical low-perception scenarios. On the other hand, deep reinforcement learning, as a data-driven method, provides a more flexible solution through offline training and fast feedforward computation, which can address the problem of missing parameters to some extent. However, existing deep reinforcement learning algorithms have limited feature extraction capabilities in complex topologies and dynamically changing environments, making it difficult to generate high-precision and widely applicable regulation strategies. Summary of the Invention

[0004] To address the problems in existing technologies, this application provides a method and apparatus for collaborative optimization and control of multiple distribution areas in a distribution network that adapts to topology changes. It combines dynamic cluster partitioning and distributed autonomous optimization, uses the SLM algorithm to enhance the collaborative capability of distribution areas and reduce the impact of cross-cluster power flow, adapts to topology changes by introducing a feature extraction network, and optimizes the generation of control strategies. It reduces communication dependencies by using centralized training and decentralized execution, and achieves efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0005] To solve at least one of the above problems, this application provides the following technical solution:

[0006] According to a first aspect of the embodiments of this application, this application provides a method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, including:

[0007] Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm;

[0008] Based on the cluster partitioning results, an optimization model is constructed that includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0009] Each transformer area is treated as an intelligent agent, and a dynamic clustering Markov decision model for each transformer area is constructed based on the agent's observation space, action space, reward function, and state transition settings.

[0010] Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

[0011] According to any embodiment of this application, the step of dividing the distribution network into clusters based on electrical distance modularity and power balance using the SLM clustering algorithm includes:

[0012] Each transformer substation in the distribution network is treated as an independent initial cluster, and the node connection relationship and load power data of all transformer substations are obtained.

[0013] Using predefined electrical distance modularity and power balance as the indicators for cluster partitioning, each transformer area is attempted to be assigned to other clusters in each iteration. The change in electrical distance modularity after allocation is calculated, the allocation method that produces the maximum modularity gain is recorded, and this process is repeated to gradually optimize cluster partitioning until the modularity no longer changes.

[0014] The SLM cluster partitioning algorithm generates several sub-networks. Each sub-network is treated as a virtual node to participate in the next round of cluster partitioning, and the modularity is further optimized until it reaches the global optimum.

[0015] According to any embodiment of this application, the step of constructing an optimization model based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and power loss, includes:

[0016] Based on the cluster partitioning results, objective functions are defined with the optimization objectives of minimizing voltage deviation and minimizing power loss, respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value, while the power loss objective function is calculated through the branch resistances in the cluster and the flow of active and reactive power.

[0017] Based on the power balance principle within the cluster, the power balance relationship of each node is calculated based on the line admittance matrix and the voltage phase angle difference between nodes to constrain the power balance of the distribution network.

[0018] Based on a preset range, the voltage amplitude of all nodes is controlled within a safe threshold range, while the branch current is limited to not exceed the maximum allowable current.

[0019] The active power output of the photovoltaic power generation system is determined by the maximum power point tracking (MPPT) mode, and the reactive power adjustment range of the photovoltaic power generation system is set according to the reactive power capacity of the photovoltaic power generation system.

[0020] Based on a preset range, the current SOC state of the energy storage system must be within a safe threshold range, and the charging and discharging power range of the energy storage device is set in conjunction with the charging and discharging efficiency of the energy storage system.

[0021] According to any embodiment of this application, the step of treating each transformer area as an intelligent agent and constructing a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition includes:

[0022] Each transformer area is treated as an intelligent agent, and the observation space is defined as the real-time power flow information of all nodes in the transformer area based on the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes. The observation space is dynamically adjusted to adapt to the new cluster structure when the cluster boundary changes.

[0023] Based on the observation information of the intelligent agent, the action space is defined as the active and reactive power adjustment of the interaction between the transformer area and the feeder, including the power output or absorption achieved through photovoltaic and energy storage devices, and the power of the transformer area nodes is redistributed after the action is executed.

[0024] The reward function of the intelligent agent is constructed according to the control target, including a reward term based on the target function and a penalty term based on the constraint conditions. The reward term reflects the achievement of the control target by calculating the node voltage deviation and power loss, and the penalty term punishes the violation of the energy storage device constraint conditions.

[0025] From the initial state until the power flow has no feasible solution or the scheduling cycle ends, the agent calculates the reward function value after performing actions based on the current observation space and action space, and then enters the next time-instance state for state transition.

[0026] According to any embodiment of this application, the step of solving the dynamic grouping Markov decision model based on the optimization model using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-distribution network coordinated optimization and control strategy includes:

[0027] An Actor network is constructed for each agent, and based on the agent's observation input, the corresponding active power and reactive power adjustment values ​​are output.

[0028] A Critic network is constructed for each agent. The value of each action is evaluated based on the set of observations and actions of all agents in the current state. The output of the control Critic network is updated based on the parameters of the target Critic network.

[0029] For each agent, a target Actor network and a target Critic network are set up, and the parameters of the training network are periodically copied to the target network to maintain parameter stability during reinforcement learning;

[0030] In each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, the parameters of the Critic network are optimized by gradient descent to minimize the joint regression loss, and the parameters of the Actor network are optimized by gradient ascent to maximize the action value of the agent.

[0031] Based on the iterative training results, the corresponding multi-region collaborative optimization and control strategy of the distribution network is obtained, and the power regulation amount of the region is output based on the current observation space and the optimized action value.

[0032] According to any embodiment of this application, the step of constructing a Critic network for each agent and evaluating the value of each action based on the set of observations and actions of all agents in the current state, wherein controlling the output of the Critic network based on parameter updates of the target Critic network includes:

[0033] The Critic network input for each agent includes environmental and attention information about the power distribution network environment.

[0034] Weights are generated for each agent's observations and actions through an attention mechanism. The correlation between the current agent and other agents is calculated using a key-value query method. Weighted parameters are then generated to describe the degree of contribution of different agents to the value assessment.

[0035] The environmental and attention information of the agent is fused, and the fused features are input into the Critic network. Through activation functions and fully connected networks, the value output of the agent's actions is generated.

[0036] Specifically, a joint regression loss function is defined based on the outputs of the Critic network and the target Critic network, and the Critic network is updated by minimizing the loss function.

[0037] According to a second aspect of the embodiments of this application, this application provides a multi-area collaborative optimization and control device for a distribution network that adapts to topology changes, comprising:

[0038] The cluster partitioning module is used to partition the distribution network into clusters based on electrical distance modularity and power balance using the SLM cluster partitioning algorithm.

[0039] The optimization model construction module is used to: construct an optimization model containing objective functions of minimizing voltage deviation and minimizing power loss based on the cluster partitioning results. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster, and the constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0040] The agent modeling module is used to: treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0041] The optimization solution module is used to: solve the dynamic cluster Markov decision model based on the optimization model using the multi-agent deep deterministic policy gradient (MADDPG) algorithm, obtain the corresponding multi-area collaborative optimization control strategy of the distribution network, and control the distribution network according to the control strategy.

[0042] According to any embodiment of this application, the cluster partitioning module includes:

[0043] Initialize the cluster unit, which is used to: treat each transformer area in the distribution network as an independent initial cluster, and obtain the node connection relationship and load power data of all transformer areas;

[0044] The iterative optimization unit is used to: use predefined electrical distance modularity and power balance as indicators for cluster partitioning, attempt to assign each transformer area to other clusters in each iteration, calculate the change in electrical distance modularity after allocation, record the allocation method that produces the maximum modularity gain, and repeat this process to gradually optimize cluster partitioning until the modularity no longer changes.

[0045] The subnetwork merging unit is used to generate several subnetworks using the SLM cluster partitioning algorithm, treat each subnetwork as a virtual node to participate in the next round of cluster partitioning, and continue to optimize the modularity until it reaches the global optimum.

[0046] According to any embodiment of this application, the optimization model construction module includes:

[0047] The objective function definition unit is used to: define objective functions based on the cluster partitioning results, with minimizing voltage deviation and minimizing power loss as optimization objectives respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value. The power loss objective function is calculated through the branch resistance and the flow of active and reactive power in the cluster.

[0048] The power balance constraint unit is used to: calculate the power balance relationship of each node based on the line admittance matrix and the voltage phase angle difference between nodes according to the power balance principle within the cluster, so as to perform power balance constraint of the distribution network;

[0049] The safety constraint unit is used to: control the voltage amplitude of all nodes within a safe threshold range based on a preset range, while limiting the branch current to not exceed the maximum allowable current;

[0050] The photovoltaic regulation unit is used to: determine the active power output of the photovoltaic power generation system through the maximum power point tracking mode (MPPT), and set the reactive power regulation range of the photovoltaic system according to the reactive power capacity of the photovoltaic power generation system.

[0051] The energy storage management unit is used to: control the current SOC state of the energy storage system to be within a safe threshold range based on a preset range, and set the charging and discharging power range of the energy storage device in conjunction with the charging and discharging efficiency of the energy storage system.

[0052] According to any embodiment of this application, the intelligent agent modeling module includes:

[0053] The observation space definition unit is used to: treat each transformer area as an intelligent agent, and define the observation space as the real-time power flow information of all nodes in the transformer area according to the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes, and dynamically adjust the observation space to adapt to the new cluster structure when the cluster boundary changes.

[0054] The action space definition unit is used to: define the action space as the active and reactive power adjustment quantities of the interaction between the transformer area and the feeder based on the observation information of the intelligent agent, including the power output or absorption realized through photovoltaic and energy storage devices, and redistribute the power of the transformer area nodes after the action is executed.

[0055] The reward function construction unit is used to: construct a reward function for the intelligent agent according to the control target, including a reward term based on the target function and a penalty term based on the constraint conditions, wherein the reward term reflects the achievement of the control target by calculating the node voltage deviation and power loss, and the penalty term punishes the violation of the energy storage device constraint conditions;

[0056] The state transition unit is used to: calculate the reward function value after the agent performs actions based on the current observation space and action space from the initial state until the power flow has no feasible solution or the scheduling cycle ends, and then enter the next time step state to perform state transition.

[0057] According to any embodiment of this application, the optimization solution module includes:

[0058] The Actor network construction unit is used to: construct an Actor network for each agent, and output the corresponding active power and reactive power adjustment based on the agent's observation input;

[0059] The Critic network building unit is used to: build a Critic network for each agent, evaluate the value of each action based on the set of observations and actions of all agents in the current state, and control the output of the Critic network based on the parameter updates of the target Critic network.

[0060] The target network setting unit is used to: set up a target Actor network and a target Critic network for each agent, and periodically copy the parameters of the training network to the target network to maintain parameter stability during reinforcement learning;

[0061] The network parameter optimization unit is used to: in each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, optimize the parameters of the Critic network using gradient descent to minimize the joint regression loss, and optimize the parameters of the Actor network using gradient ascent to maximize the action value of the agent.

[0062] The control strategy generation unit is used to: obtain the corresponding multi-region collaborative optimization control strategy of the distribution network based on the iterative training results, and output the power regulation amount of the region based on the current observation space and the optimized action value.

[0063] According to any embodiment of this application, the Critic network building unit includes:

[0064] The environmental information input unit is used to input environmental and attention information about the power distribution network environment into the Critic network for each agent.

[0065] The attention weight generation unit is used to: generate weights for the combination of observations and actions of each agent through the attention mechanism, calculate the correlation between the current agent and other agents using the key-value query device, and generate weighted parameters to describe the degree of contribution of different agents to the value assessment.

[0066] The feature fusion unit is used to: fuse the environmental information and attention information of the agent, input the fused features into the Critic network, and generate the value output of the agent's actions through the activation function and the fully connected network;

[0067] The network update unit is used to: define a joint regression loss function based on the output of the Critic network and the output of the target Critic network, and update the Critic network by minimizing the loss function.

[0068] According to a third aspect of the embodiments of this application, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes.

[0069] According to a fourth aspect of the embodiments of this application, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes.

[0070] According to a fifth aspect of the embodiments of this application, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the aforementioned method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes.

[0071] As can be seen from the above technical solution, this application provides a method and device for collaborative optimization and control of multiple distribution areas in a distribution network that adapts to topology changes. This involves dividing the distribution network into clusters; constructing an optimization model based on the cluster division results; constructing a dynamic clustering Markov decision model for each distribution area as an agent; solving the dynamic clustering Markov decision model based on the optimization model to obtain the corresponding collaborative optimization and control strategy for multiple distribution areas; and controlling the distribution network according to the control strategy. This application combines dynamic cluster division and distributed autonomous optimization, using the SLM algorithm to enhance the collaborative capability of distribution areas and reduce the impact of cross-cluster power flow. It also introduces a feature extraction network to adapt to topology changes and optimize the generation of control strategies. Furthermore, it utilizes centralized training and decentralized execution to reduce communication dependencies, achieving efficient voltage control in low-perception scenarios in complex large-scale distribution network environments. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is one of the flowcharts illustrating the collaborative optimization and control method for multiple distribution areas in a distribution network that adapts to topology changes, as described in this application embodiment.

[0074] Figure 2 This is the second flowchart illustrating the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in this application.

[0075] Figure 3 This is the third flowchart illustrating the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in this application.

[0076] Figure 4 This is a schematic diagram of the feeder-multi-distribution-region collaborative structure of the distribution network multi-region collaborative optimization control method for adapting to topology changes in the embodiments of this application.

[0077] Figure 5 This is the fourth flowchart illustrating the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in the embodiments of this application.

[0078] Figure 6 This is a schematic diagram illustrating the cluster boundary changes under different topologies in the collaborative optimization and control method for multi-area distribution networks that adapts to topology changes in the embodiments of this application.

[0079] Figure 7 This is the fifth flowchart illustrating the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in the embodiments of this application.

[0080] Figure 8 This is the sixth flowchart illustrating the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in the embodiments of this application.

[0081] Figure 9 This is a schematic diagram of the AECN network structure of the collaborative optimization and control method for multiple distribution areas in a distribution network that adapts to topology changes, as described in the embodiments of this application.

[0082] Figure 10 This is a structural diagram of the multi-area collaborative optimization and control device for distribution networks that adapts to topology changes in the embodiments of this application;

[0083] Figure 11 This is a schematic diagram of the structure of the electronic device in the embodiments of this application. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0085] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0086] In view of the problems existing in the prior art, this application provides a method and device for collaborative optimization and control of multiple distribution areas in a distribution network that adapts to topology changes. It can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the collaborative capability of distribution areas and reduce the impact of cross-cluster power flow; it can adapt to topology changes by introducing a feature extraction network and optimize the generation of control strategies; and it can reduce communication dependence by using centralized training and decentralized execution, so as to achieve efficient voltage control in low-perception scenarios in the complex environment of large-scale distribution networks.

[0087] To combine dynamic cluster partitioning and distributed autonomous optimization, the SLM algorithm is used to enhance the coordination capability of distribution substations and reduce the impact of cross-cluster power flow. A feature extraction network is introduced to adapt to topology changes and optimize the generation of control strategies. Centralized training and distributed execution are used to reduce communication dependencies, achieving efficient voltage regulation in low-perception scenarios in complex large-scale distribution networks. This application provides an embodiment of a multi-distribution substation collaborative optimization control method for distribution networks that adapts to topology changes. See [link to relevant documentation]. Figure 1 The method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes specifically includes the following:

[0088] Step S101: Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm.

[0089] First, each transformer substation in the distribution network is treated as an independent initial cluster, and its node connectivity and load power data are obtained. Using electrical distance modularity and power balance as indicators, the SLM algorithm is used to dynamically partition the distribution network into clusters. Electrical distance modularity reflects the tight coupling between nodes between substations, while power balance assesses the supply-demand matching within a substation. In each iteration, substations are attempted to be assigned to other clusters, and the allocation method that maximizes modularity gain is recorded until the modularity no longer changes. Subsequently, further optimization is performed by generating virtual subnetworks, ultimately determining the globally optimal cluster partitioning result.

[0090] Step S102: Construct an optimization model based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0091] Based on the cluster partitioning results, an optimization model is constructed with the objectives of minimizing voltage deviation and power loss. The voltage deviation objective function measures the sum of squared deviations of all node voltages from the baseline value, while the power loss objective function is calculated using branch resistances within the cluster and the flow of active and reactive power. Simultaneously, the model sets various constraints, including node power balance constraints, voltage safety range constraints, photovoltaic reactive power adjustment range constraints, and SOC and charging / discharging power limits for the energy storage system, providing clear objectives and constraints for the subsequent generation of intelligent agent control strategies.

[0092] Step S103: Treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0093] Each transformer substation is abstracted as an intelligent agent and described using a dynamic clustering Markov decision model. The observation space includes real-time power flow information for all nodes within the substation, such as active power, reactive power, and voltage amplitude, and can dynamically adapt to changes in the cluster boundary. The action space is defined as the active and reactive power regulation quantities achieved through photovoltaic and energy storage devices. The reward function combines the objective function and constraints, including a reward term for achieving the objective and a penalty term for violating the constraints. The agent generates the state for the next time step based on the current observations and actions until the regulation cycle ends or the power flow cannot be solved, completing the state transition and dynamic modeling.

[0094] Step S104: Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

[0095] The MADDPG algorithm is used to solve the dynamic clustered Markov decision model. An Actor network is constructed for each agent to generate action outputs based on the observation space; simultaneously, a Critic network is constructed to evaluate the value of actions, and the network is trained stably using the target network. During training, gradient descent is used to optimize the Critic network to minimize regression loss, while gradient ascent is used to optimize the Actor network to maximize policy value. After training, the generated control policy is used for real-time power regulation of distribution network equipment, covering photovoltaic reactive power output and energy storage device charging and discharging, achieving multi-area coordinated optimization and control of the distribution network.

[0096] As can be seen from the above description, the multi-transformer area collaborative optimization and control method for distribution networks that adapts to topology changes provided in this application can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the collaborative capability of transformer areas and reduce the impact of cross-cluster power flow; adapt to topology changes by introducing a feature extraction network and optimize the generation of control strategies; reduce communication dependence by using centralized training and decentralized execution, and achieve efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0097] To enhance intra-cluster coordination, reduce the impact of cross-cluster power flow, dynamically adapt to topology changes, optimize modularity to achieve globally optimal partitioning, and reduce power loss and control complexity, in one embodiment of the topology-adaptive multi-distributor area collaborative optimization control method for distribution networks, see [link to relevant documentation]. Figure 2 The method of dividing the distribution network into clusters based on electrical distance modularity and power balance using the SLM clustering algorithm includes:

[0098] Step S101A: Treat each transformer substation in the distribution network as an independent initial cluster, and obtain the node connection relationship and load power data of all transformer substations;

[0099] Step S101B: Using the predefined electrical distance modularity and power balance as the indicators for cluster partitioning, each transformer area is attempted to be assigned to other clusters in each iteration. The change in electrical distance modularity after allocation is calculated, the allocation method that produces the maximum modularity gain is recorded, and this process is repeated to gradually optimize the cluster partitioning until the modularity no longer changes.

[0100] Step S101C: Generate several sub-networks using the SLM cluster partitioning algorithm, treat each sub-network as a virtual node to participate in the next round of cluster partitioning, and continue to optimize the modularity until it reaches the global optimum.

[0101] For example, clustering criteria serve as the basis for clustering distribution substations. The clusters should have high electrical coupling within each cluster and be approximately decoupled from each other. This paper focuses on distribution network voltage regulation and, to reduce power losses across clusters, primarily considers two clustering criteria: electrical distance modularity and power balance.

[0102] Electrical Distance Modularity Index: Modularity is used to measure the structural strength of a community network. A higher modularity value indicates a better cluster partitioning result. The modularity function ρ is defined as:

[0103]

[0104] In the formula: W ij K is the adjacency matrix in the modularity index. i Let be the sum of the weights of all edges connected to node i; let s be the sum of all network weights; C i Let λ represent the cluster where node i belongs, and let λ be a 0-1 function. λ = 1 indicates that node i and node j are assigned to the same cluster.

[0105] Power balance index: In order to achieve full absorption of DG and reduce the power transmission of the cluster, a power balance index e is defined to describe the power supply and demand balance within the cluster.

[0106] Combining modularity and power balance metrics, the overall cluster partitioning index λ is expressed as:

[0107] λ=a1ρ+a2e (2)

[0108] In the formula: a1 and a2 are the weight coefficients of the two cluster partitioning indicators.

[0109] The process of cluster partitioning a distribution network based on the SLM cluster partitioning algorithm is as follows: First, each transformer substation in the distribution network is considered a cluster. Then, each substation is sequentially assigned to the clusters of other substation nodes, and the difference in modularity before and after the assignment is calculated. The position of the node with the largest difference is recorded. This process is repeated until the modularity no longer changes. Next, the LMH algorithm is run to generate several subnetworks. Each subnetwork is considered a node and participates in the cluster allocation iteration process again until the modularity no longer increases, indicating that the optimal partitioning method has been found.

[0110] To achieve the goal of minimizing voltage deviation and power loss, an optimization model is constructed for the control strategy. Combined with constraints, this improves grid operating efficiency and control stability. In one embodiment of the multi-distributor area collaborative optimization control method for distribution networks adapting to topology changes, see [link to relevant documentation]. Figure 3 The optimization model constructed based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and power loss, includes:

[0111] Step S102A: Based on the cluster partitioning results, define objective functions with minimizing voltage deviation and minimizing power loss as optimization objectives respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value. The power loss objective function is calculated by the branch resistance and the flow of active and reactive power in the cluster.

[0112] Step S102B: Based on the power balance principle within the cluster, calculate the power balance relationship of each node based on the line admittance matrix and the voltage phase angle difference between nodes to constrain the power balance of the distribution network.

[0113] Step S102C: Based on a preset range, control the voltage amplitude of all nodes within the safe threshold range, while limiting the branch current to not exceed the maximum allowable current;

[0114] Step S102D: Determine the active power output of the photovoltaic power generation system through the maximum power point tracking mode (MPPT), and set the reactive power adjustment range of the photovoltaic system according to the reactive power capacity of the photovoltaic power generation system.

[0115] Step S102E: Based on a preset range, control the current SOC state of the energy storage system to be within a safe threshold range, and set the charging and discharging power range of the energy storage device in conjunction with the charging and discharging efficiency of the energy storage system.

[0116] For example, the connection relationship between the low-voltage distribution substation and the feeder is as follows: Figure 4 As shown, the active and reactive power of the entire transformer area at each moment interacts with the feeder layer.

[0117] The transformer area includes PV, energy storage system (ESS), and DC loads. The power balance equation for transformer area i is:

[0118]

[0119]

[0120] In the formula: and These are the active and reactive power flowing into transformer area i from DG, respectively; and These are the charging and discharging power of the energy storage, respectively. and These represent the active and reactive power of load i in transformer area; and These represent the active and reactive power flowing into the transformer area, respectively.

[0121] To fully exploit the adjustment capabilities of adjustable resources in distribution transformer substations, this application uses minimizing distribution network voltage deviation and active power loss as the objective function. Considering constraints such as power balance, adjustment resource output, and node voltage safety, a multi-distribution transformer substation collaborative optimization model is constructed. By optimizing the reactive power output of PV and the charging and discharging power of ESS within the substations, the economic efficiency and safety of the distribution network operation are improved. Based on the cluster partitioning results in the previous section, cluster C is constructed. m The objective function within is as follows:

[0122]

[0123]

[0124] In the formula: C is the cluster number, Ω clu For all cluster sets; C m Represents the set of nodes in cluster m; T represents the total scheduling period; U i,t Let U be the voltage magnitude at node i at time t. ref This is the voltage reference value; Ω represents the active power loss of cluster m at time t. br,m Let r be the set of branches within cluster m. ij Let P be the resistance of branch ij. ij,t and Q ij,t These represent the active power and reactive power flowing into branch ij, respectively.

[0125] The constraints in this application are set as follows:

[0126] Distribution network power balance constraints:

[0127]

[0128] Where: G ij and B ij Let θ be the real and imaginary parts of the line admittance ij. ij Let be the voltage phase angle difference between nodes i and j.

[0129] Operational safety constraints:

[0130]

[0131] In the formula: U max and U min These are the safe upper and lower limits for node voltage; Let be the maximum current allowed to flow through branch ij.

[0132] DG regulation performance constraints:

[0133] The photovoltaic active power adopts the Maximum Power Point Tracking (MPPT) mode. This paper considers the full absorption of distributed photovoltaic power and only uses its reactive power capacity for regulation. The reactive power regulation range of Hengshi in the power control mode is:

[0134]

[0135] In the formula: S PV,max For the maximum capacity of photovoltaic power, P PV Contributing to photovoltaic power, Q PV It contributes reactive power to photovoltaic power generation.

[0136] The current State of Charge (SOC) of the ESS is related to the SOC and charging / discharging power of the previous time step. The charging / discharging power constraint of the ESS is as follows:

[0137]

[0138] E i.t ≤E max (11)

[0139]

[0140] Where: E i.t η represents the amount of electricity stored in the ESS within area i at time t. ch and η dis E represents the charging and discharging efficiency of ESS, respectively. max This is the maximum storage capacity of ESS; This is an indicator of the ESS discharge status. This indicates that the ESS is in a discharging state; and These represent the maximum values ​​of the ESS charging and discharging power, respectively.

[0141] To adapt to changes in cluster boundaries, adaptive regulation is achieved through a reward function, enhancing the collaborative optimization capability of distribution areas. In one embodiment of the distribution network multi-distribution area collaborative optimization regulation method adapting to topology changes in this application, see [link to relevant documentation]. Figure 5 The process involves treating each transformer substation as an agent and constructing a dynamic clustering Markov decision model for each substation based on the agent's observation space, action space, reward function, and state transitions. This includes:

[0142] Step S103A: Treat each transformer area as an intelligent agent, and define the observation space as the real-time power flow information of all nodes in the transformer area according to the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes, and dynamically adjust the observation space to adapt to the new cluster structure when the cluster boundary changes.

[0143] Step S103B: Based on the observation information of the intelligent agent, the action space is defined as the active and reactive power adjustment of the interaction between the transformer area and the feeder, including the power output or absorption achieved through photovoltaic and energy storage devices, and the power of the transformer area nodes is redistributed after the action is executed.

[0144] Step S103C: Construct a reward function for the intelligent agent based on the control objective, including a reward term based on the objective function and a penalty term based on the constraint conditions. The reward term reflects the achievement of the control objective by calculating the node voltage deviation and power loss, and the penalty term penalizes violations of the energy storage device's constraint conditions.

[0145] Step S103D: From the initial state until the power flow has no feasible solution or the scheduling cycle ends, the agent calculates the reward function value after performing actions based on the current observation space and action space, and enters the next time step state for state transition.

[0146] For example, the collaborative optimization of multiple distribution network areas, which makes decisions based on the current and previous operating states, can be viewed as a sequential decision problem. Since the electrical coupling between groups is weak after cluster partitioning, autonomous control occurs within each group. Therefore, this paper uses a multi-agent Markov decision model to model the multi-distribution network clusters, with each distribution area as an agent and the distribution network as the training environment. The main components of the model are as follows:

[0147] 1) Intelligent agent observation space O i,t The observation space represents the observations upon which the agent i in the distribution area makes its decisions, and consists of power flow information for all nodes within the same cluster where the distribution area is located.

[0148]

[0149] In the formula: and The active and reactive power of the distribution network node connected to transformer substation i at time t; Ω m,SA This is the set of all station nodes in cluster m.

[0150] Considering topology generation, such as Figure 6 After the changes shown, the agent's observations also change accordingly with the cluster boundary. Taking the three agents i1, i2, and i3 in topology A as examples, their observation spaces are as follows:

[0151]

[0152] When the topology changes from A to B, the observation spaces of the three agents in the monitoring area become as follows:

[0153]

[0154] 2) Agent action space A i,t Action space representation of the intelligent agent i in the station area based on observation O i,t The decision-making quantity is as follows:

[0155]

[0156] In the formula: and These represent the active and reactive power of the interaction between transformer area i and the feeder layer, respectively, which are realized by the photovoltaic and energy storage inverters on the transformer area side. This indicates the active power output of the transformer substation. Indicates the absorption of active power. Similarly.

[0157] Action A i,t After execution, the power of the distribution network node connected to transformer area i is:

[0158]

[0159] 3) Reward function R i,t The reward function is used to measure the reward of an agent i based on a set of observations and actions (O). i,t A i,t The value of the reward includes the objective function reward and the constraint reward. A larger reward function indicates a higher current (O) value. i,t A i,t The closer the distance to the optimal goal, the more cooperative the relationship between platforms. The reward function is represented by a negative objective function and a penalty term.

[0160]

[0161]

[0162] In the formula: Let λ represent the energy storage constraint violated by agent j during its action decision-making process, and let λ be the weighting coefficient balancing the objective reward and the constraint reward. and All are coefficients of the ESS penalty function.

[0163] 4) State transition: Agent j changes state based on the current observation O i,t Perform action A j,t Then, the immediate reward function R is obtained. j,t And then enter the next state, obtaining the observation O at the next moment. i,t This process is a state transition. The entire process begins at the initial moment and proceeds with state transitions over time until there is no feasible solution for the power flow or a scheduling cycle ends.

[0164] To combine centralized training with distributed execution, reduce communication requirements, adapt to large-scale scenarios, and improve control efficiency and accuracy, in one embodiment of the distribution network multi-distributor area collaborative optimization control method adapted to topology changes in this application, see [reference needed]. Figure 7 Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization and control strategy for the distribution network, including:

[0165] Step S104A: Construct an Actor network for each agent, and output the corresponding active power and reactive power adjustment amounts based on the agent's observation inputs;

[0166] Step S104B: Construct a Critic network for each agent, evaluate the value of each action based on the set of observations and actions of all agents in the current state, wherein the output of the control Critic network is updated based on the parameters of the target Critic network;

[0167] Step S104C: Set up a target Actor network and a target Critic network for each agent, and periodically copy the parameters of the training network to the target network to maintain parameter stability during reinforcement learning;

[0168] Step S104D: In each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, the parameters of the Critic network are optimized by gradient descent to minimize the joint regression loss, and the parameters of the Actor network are optimized by gradient ascent to maximize the action value of the agent.

[0169] Step S104E: Obtain the corresponding multi-region collaborative optimization control strategy of the distribution network based on the iterative training results, and output the power regulation amount of the region based on the current observation space and the optimized action value.

[0170] For example, the actions of the transformer area agents are active or reactive power (DG and ESS), both of which are continuously adjustable. Therefore, the MADDPG algorithm, which uses a continuous action space in multi-agent reinforcement learning, is chosen to solve the problem. For each transformer area agent i, the MADDPG algorithm needs to maintain four neural networks: based on the observation O i,t Perform action A j,t Actor network π i,t (O i,t ξ i ) and the target Actor network; the Critic network Q evaluates the value of actions. i,t (S t A t η i) and the target Critic network, where S t and A t Let ξ be the set of observations and actions of agent i and the other agents in the same cluster at time t. i and η i These are the corresponding network parameters. The target network does not participate in network training. The network parameters of the Actor and Critic that are used in training are periodically copied to the corresponding target network to stabilize the network training process.

[0171] The output of the Actor network is action A. i,t , represented as:

[0172]

[0173] In the formula: σ is Gaussian noise, and These are the upper and lower limits of the operation, which are the upper and lower limits of the power interaction between the transformer substation and the distribution network; This represents a random number between the upper and lower limits of the action.

[0174] The Critic network output is for S t and A t Value assessment based on updating network parameters η using gradient descent. i Actor network training employs gradient ascent to maximize the current (O) value. i,t, A i,t The value of Q) j (O i,t A i,t ).

[0175] To optimize value assessment using environmental and attention information, enhance feature extraction capabilities, and dynamically adapt to complex environments, joint regression is used to optimize the stability and accuracy of the Critic network. In one embodiment of the collaborative optimization and control method for multi-distribution areas in a distribution network adapting to topology changes proposed in this application, see [link to relevant documentation]. Figure 8 The step involves constructing a Critic network for each agent, evaluating the value of each action based on the set of observations and actions of all agents in the current state, wherein controlling the output of the Critic network is based on parameter updates of the target Critic network, including:

[0176] Step S104B1: Input environmental information and attention information about the power distribution network environment into the Critic network for each agent;

[0177] Step S104B2: Generate weights for the combination of observations and actions of each agent through the attention mechanism, calculate the correlation between the current agent and other agents using the key-value query method, and generate weighted parameters to describe the degree of contribution of different agents to the value assessment.

[0178] Step S104B3: The environmental information and attention information of the agent are fused, and the fused features are input into the Critic network. Through the activation function and the fully connected network, the value output of the agent's action is generated.

[0179] Step S104B4: Wherein, a joint regression loss function is defined based on the output of the Critic network and the output of the target Critic network, and the Critic network is updated by minimizing the loss function.

[0180] The Critic network in the traditional MADDPG algorithm is composed of a fully connected neural network, and the number of neurons in its input layer is (S t A t The fixed dimension of the Critic network means that once the network is built, the input dimension cannot be changed, thus limiting it to scenarios with a fixed topology and a fixed number of agents. When the topology changes, the cluster boundary expands or contracts, leading to the problem of dynamic changes in the agent observation space as described in Section 3.2. To address this issue, this paper proposes an attention encoder-based Critic network (AECN), which maps dynamic observation inputs to a fixed-dimensional space through encoding, ensuring the network's dynamic scalability. The network structure diagram is shown below. Figure 9 As shown.

[0181] For example, for any intelligent agent i, the input of its AECN can be divided into two parts: one part is environmental information h. i,t =g i (O i,t A i,t This indicates the observation-action combination (O) of the user regarding the distribution network environment. i,t A i,t The other part is the attention information h of the observation-action combinations of the other agents in the cluster. j,t :

[0182]

[0183]

[0184]

[0185] Where: g j (O j,t A j,t W represents the encoding of agent j. k and W q For mapping g j (Oj,t A j,t ) to key-value and mapping g i (O i,t A i,t The parameters for the query, e ij,t To calculate the correlation coefficient of the observation-action combinations for agents i and j. ω ij,t To make e ij,t The attention weight coefficients obtained after normalization represent the degree to which agent i observes the importance of the actions of other agents to the task.

[0186] The Critic network of agent i concatenates the two inputs to obtain the output Q. i,t (S t A t ) is represented as:

[0187]

[0188] In the formula: F i For a fully connected neural network, V is the input to g. j (O j,t A j,t The linear transformation matrix of ), where h is the activation function.

[0189] The Critic network update is implemented by minimizing the joint regression function, with the loss function being:

[0190]

[0191] y i,t =R i,t +γQ i '(S t ',A t ') (26)

[0192] In the formula: E represents the expectation, Q i '(S t ',A t ') represents the output of the target Critic network.

[0193] As can be seen from the above embodiments, this application provides a new solution for voltage regulation in low-sensitivity and dynamic topology environments of distribution networks by combining dynamic cluster partitioning with distributed autonomous optimization. In addition, by using the SLM algorithm for cluster partitioning, the coordination capability of transformer substations within the cluster is enhanced, and the impact of cross-cluster power flow on overall regulation is reduced.

[0194] Introducing a feature extraction network into the reinforcement learning model enables the control strategy to adapt to frequent topology adjustments, while also optimizing the control strategy generation process between stations. Utilizing a centralized training and distributed execution approach reduces reliance on network-wide communication, helping to maintain control functionality even under data shortages.

[0195] This application can support large-scale distribution network scenarios, meet the requirements for real-time performance, adaptability and accuracy, and provide a foundation for collaborative optimization of multiple distribution areas in complex operating environments.

[0196] To combine dynamic cluster partitioning and distributed autonomous optimization, the SLM algorithm is employed to enhance the coordination capability of distribution substations and reduce the impact of cross-cluster power flow. A feature extraction network is introduced to adapt to topology changes and optimize the generation of control strategies. Centralized training and decentralized execution are used to reduce communication dependencies, achieving efficient voltage control in low-perception scenarios in complex large-scale distribution network environments. This application provides an embodiment of a topology-adaptive multi-distribution substation collaborative optimization control device for implementing all or part of the aforementioned topology-adaptive multi-distribution substation collaborative optimization control method. See [link to relevant documentation]. Figure 10 The aforementioned multi-area collaborative optimization and control device for distribution networks that adapts to topology changes specifically includes the following components:

[0197] Cluster partitioning module 1101 is used to: partition the distribution network into clusters based on electrical distance modularity and power balance degree using the SLM cluster partitioning algorithm;

[0198] The optimization model construction module 1102 is used to: construct an optimization model containing objective functions of minimizing voltage deviation and minimizing power loss based on the cluster partitioning results. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster, and the constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage equipment.

[0199] The agent modeling module 1103 is used to: treat each transformer area as an agent and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0200] The optimization solution module 1104 is used to: solve the dynamic cluster Markov decision model based on the optimization model using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy of the distribution network, and control the distribution network according to the control strategy.

[0201] According to any embodiment of this application, the cluster partitioning module includes:

[0202] Initialize the cluster unit, which is used to: treat each transformer area in the distribution network as an independent initial cluster, and obtain the node connection relationship and load power data of all transformer areas;

[0203] The iterative optimization unit is used to: use predefined electrical distance modularity and power balance as indicators for cluster partitioning, attempt to assign each transformer area to other clusters in each iteration, calculate the change in electrical distance modularity after allocation, record the allocation method that produces the maximum modularity gain, and repeat this process to gradually optimize cluster partitioning until the modularity no longer changes.

[0204] The subnetwork merging unit is used to generate several subnetworks using the SLM cluster partitioning algorithm, treat each subnetwork as a virtual node to participate in the next round of cluster partitioning, and continue to optimize the modularity until it reaches the global optimum.

[0205] According to any embodiment of this application, the optimization model construction module includes:

[0206] The objective function definition unit is used to: define objective functions based on the cluster partitioning results, with minimizing voltage deviation and minimizing power loss as optimization objectives respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value. The power loss objective function is calculated through the branch resistance and the flow of active and reactive power in the cluster.

[0207] The power balance constraint unit is used to: calculate the power balance relationship of each node based on the line admittance matrix and the voltage phase angle difference between nodes according to the power balance principle within the cluster, so as to perform power balance constraint of the distribution network;

[0208] The safety constraint unit is used to: control the voltage amplitude of all nodes within a safe threshold range based on a preset range, while limiting the branch current to not exceed the maximum allowable current;

[0209] The photovoltaic regulation unit is used to: determine the active power output of the photovoltaic power generation system through the maximum power point tracking mode (MPPT), and set the reactive power regulation range of the photovoltaic system according to the reactive power capacity of the photovoltaic power generation system.

[0210] The energy storage management unit is used to: control the current SOC state of the energy storage system to be within a safe threshold range based on a preset range, and set the charging and discharging power range of the energy storage device in conjunction with the charging and discharging efficiency of the energy storage system.

[0211] According to any embodiment of this application, the intelligent agent modeling module includes:

[0212] The observation space definition unit is used to: treat each transformer area as an intelligent agent, and define the observation space as the real-time power flow information of all nodes in the transformer area according to the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes, and dynamically adjust the observation space to adapt to the new cluster structure when the cluster boundary changes.

[0213] The action space definition unit is used to: define the action space as the active and reactive power adjustment quantities of the interaction between the transformer area and the feeder based on the observation information of the intelligent agent, including the power output or absorption realized through photovoltaic and energy storage devices, and redistribute the power of the transformer area nodes after the action is executed.

[0214] The reward function construction unit is used to: construct a reward function for the intelligent agent according to the control target, including a reward term based on the target function and a penalty term based on the constraint conditions, wherein the reward term reflects the achievement of the control target by calculating the node voltage deviation and power loss, and the penalty term punishes the violation of the energy storage device constraint conditions;

[0215] The state transition unit is used to: calculate the reward function value after the agent performs actions based on the current observation space and action space from the initial state until the power flow has no feasible solution or the scheduling cycle ends, and then enter the next time step state to perform state transition.

[0216] According to any embodiment of this application, the optimization solution module includes:

[0217] The Actor network construction unit is used to: construct an Actor network for each agent, and output the corresponding active power and reactive power adjustment based on the agent's observation input;

[0218] The Critic network building unit is used to: build a Critic network for each agent, evaluate the value of each action based on the set of observations and actions of all agents in the current state, and control the output of the Critic network based on the parameter updates of the target Critic network.

[0219] The target network setting unit is used to: set up a target Actor network and a target Critic network for each agent, and periodically copy the parameters of the training network to the target network to maintain parameter stability during reinforcement learning;

[0220] The network parameter optimization unit is used to: in each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, optimize the parameters of the Critic network using gradient descent to minimize the joint regression loss, and optimize the parameters of the Actor network using gradient ascent to maximize the action value of the agent.

[0221] The control strategy generation unit is used to: obtain the corresponding multi-region collaborative optimization control strategy of the distribution network based on the iterative training results, and output the power regulation amount of the region based on the current observation space and the optimized action value.

[0222] According to any embodiment of this application, the Critic network building unit includes:

[0223] The environmental information input unit is used to input environmental and attention information about the power distribution network environment into the Critic network for each agent.

[0224] The attention weight generation unit is used to: generate weights for the combination of observations and actions of each agent through the attention mechanism, calculate the correlation between the current agent and other agents using the key-value query device, and generate weighted parameters to describe the degree of contribution of different agents to the value assessment.

[0225] The feature fusion unit is used to: fuse the environmental information and attention information of the agent, input the fused features into the Critic network, and generate the value output of the agent's actions through the activation function and the fully connected network;

[0226] The network update unit is used to: define a joint regression loss function based on the output of the Critic network and the output of the target Critic network, and update the Critic network by minimizing the loss function.

[0227] As can be seen from the above description, the multi-area collaborative optimization and control device for distribution networks that adapts to topology changes provided in this application can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the collaborative capability of distribution areas and reduce the impact of cross-cluster power flow; adapt to topology changes by introducing a feature extraction network and optimize the generation of control strategies; reduce communication dependence by using centralized training and decentralized execution, and achieve efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0228] From a hardware perspective, in order to combine dynamic cluster partitioning and distributed autonomous optimization, the SLM algorithm is adopted to enhance the coordination capability of distribution areas and reduce the impact of cross-cluster power flow; by introducing a feature extraction network to adapt to topology changes, the generation of control strategies is optimized; by utilizing centralized training and decentralized execution to reduce communication dependencies, efficient voltage control in low-perception scenarios is achieved in the complex environment of large-scale distribution networks. This application provides an embodiment of an electronic device for implementing all or part of the above-described method for coordinated optimization control of multiple distribution areas in a distribution network that adapts to topology changes. The electronic device specifically includes the following:

[0229] The system comprises a processor, memory, a communication interface, and a bus; wherein the processor, memory, and communication interface communicate with each other via the bus; the communication interface is used to realize information transmission between the multi-distribution area collaborative optimization control device for adapting to topology changes and core business systems, user terminals, and related databases and other related equipment; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the multi-distribution area collaborative optimization control method for adapting to topology changes and the embodiments of the multi-distribution area collaborative optimization control device for adapting to topology changes in the embodiments, the contents of which are incorporated herein, and repeated details will not be described again.

[0230] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0231] In practical applications, parts of the multi-distributor area collaborative optimization control method for distribution networks adapting to topology changes can be executed on the electronic equipment side as described above, or all operations can be completed in the client equipment. The choice can be made based on the processing capacity of the client equipment and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client equipment, the client equipment may further include a processor.

[0232] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0233] Figure 11 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 11 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 11 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0234] In one embodiment, the function of the multi-area collaborative optimization control method for distribution networks adapting to topology changes can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0235] Step S101: Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm;

[0236] Step S102: Construct an optimization model based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0237] Step S103: Treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0238] Step S104: Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

[0239] As can be seen from the above description, the electronic equipment provided in this application embodiment can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the coordination capability of distribution areas and reduce the impact of cross-cluster power flow; adapt to topology changes by introducing a feature extraction network and optimize the generation of control strategies; reduce communication dependence by using centralized training and decentralized execution, and achieve efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0240] In another embodiment, the distribution network multi-area collaborative optimization control device that adapts to topology changes can be configured separately from the central processing unit 9100. For example, the distribution network multi-area collaborative optimization control device that adapts to topology changes can be configured as a chip connected to the central processing unit 9100, and the function of the distribution network multi-area collaborative optimization control method that adapts to topology changes can be realized through the control of the central processing unit.

[0241] like Figure 11 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 11All components shown; in addition, the electronic device 9600 may also include Figure 11 For components not shown, please refer to existing technologies.

[0242] like Figure 11 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0243] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0244] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0245] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0246] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device's communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0247] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module (transmitter / receiver) 9110 is coupled to the central processing unit 9100 to provide input signals and receive output signals, which can be the same as in a conventional mobile communication terminal.

[0248] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module (transmitter / receiver) 9110 is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored sound via the speaker 9131.

[0249] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the above-described method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes, where the execution subject is a server or client. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the above-described method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes, where the execution subject is a server or client. For example, when the processor executes the computer program, it implements the following steps:

[0250] Step S101: Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm;

[0251] Step S102: Construct an optimization model based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0252] Step S103: Treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0253] Step S104: Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

[0254] As can be seen from the above description, the computer-readable storage medium provided in the embodiments of this application can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the coordination capability of distribution areas and reduce the impact of cross-cluster power flow; by introducing a feature extraction network to adapt to topology changes, optimize the generation of control strategies; and by using centralized training and decentralized execution to reduce communication dependence, achieve efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0255] Embodiments of this application also provide a computer program product capable of implementing all steps in the above-described method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes, where the execution subject is a server or client. When executed by a processor, this computer program / instruction implements the steps of the method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes. For example, the computer program / instruction implements the following steps:

[0256] Step S101: Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm;

[0257] Step S102: Construct an optimization model based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices.

[0258] Step S103: Treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings.

[0259] Step S104: Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

[0260] As can be seen from the above description, the computer program product provided in this application embodiment can combine dynamic cluster partitioning and distributed autonomous optimization, use the SLM algorithm to enhance the coordination capability of distribution areas and reduce the impact of cross-cluster power flow; adapt to topology changes by introducing a feature extraction network and optimize the generation of control strategies; reduce communication dependence by using centralized training and decentralized execution, and achieve efficient voltage control in low-perception scenarios in complex environments of large-scale distribution networks.

[0261] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0262] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0263] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0264] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0265] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, characterized in that, The method includes: Based on electrical distance modularity and power balance, the distribution network is divided into clusters using the SLM clustering algorithm; Based on the cluster partitioning results, an optimization model is constructed that includes objective functions for minimizing voltage deviation and minimizing power loss. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster. The constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices. Treating each transformer substation as an agent, a dynamic clustering Markov decision model is constructed for each substation based on the agent's observation space, action space, reward function, and state transition settings. Specifically, this includes: Each transformer area is treated as an intelligent agent, and the observation space is defined as the real-time power flow information of all nodes in the transformer area based on the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes. The observation space is dynamically adjusted to adapt to the new cluster structure when the cluster boundary changes. Based on the optimization model, the dynamic cluster Markov decision model is solved using the multi-agent deep deterministic policy gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization control strategy for the distribution network, and the distribution network is controlled according to the control strategy.

2. The method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes according to claim 1, characterized in that, The method of dividing the distribution network into clusters based on electrical distance modularity and power balance using the SLM clustering algorithm includes: Each transformer substation in the distribution network is treated as an independent initial cluster, and the node connection relationship and load power data of all transformer substations are obtained. Using predefined electrical distance modularity and power balance as the indicators for cluster partitioning, each transformer area is attempted to be assigned to other clusters in each iteration. The change in electrical distance modularity after allocation is calculated, the allocation method that produces the maximum modularity gain is recorded, and this process is repeated to gradually optimize cluster partitioning until the modularity no longer changes. The SLM cluster partitioning algorithm generates several sub-networks. Each sub-network is treated as a virtual node to participate in the next round of cluster partitioning, and the modularity is further optimized until it reaches the global optimum.

3. The method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes according to claim 1, characterized in that, The optimization model constructed based on the cluster partitioning results, which includes objective functions for minimizing voltage deviation and power loss, includes: Based on the cluster partitioning results, objective functions are defined with the optimization objectives of minimizing voltage deviation and minimizing power loss, respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value, while the power loss objective function is calculated through the branch resistances in the cluster and the flow of active and reactive power. Based on the power balance principle within the cluster, the power balance relationship of each node is calculated based on the line admittance matrix and the voltage phase angle difference between nodes to constrain the power balance of the distribution network. Based on a preset range, the voltage amplitude of all nodes is controlled within a safe threshold range, while the branch current is limited to not exceed the maximum allowable current. The active power output of the photovoltaic power generation system is determined by the maximum power point tracking (MPPT) mode, and the reactive power adjustment range of the photovoltaic power generation system is set according to the reactive power capacity of the photovoltaic power generation system. Based on a preset range, the current SOC state of the energy storage system must be within a safe threshold range, and the charging and discharging power range of the energy storage device is set in conjunction with the charging and discharging efficiency of the energy storage system.

4. The method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes according to claim 1, characterized in that, The process involves treating each transformer substation as an agent and constructing a dynamic clustering Markov decision model for each substation based on the agent's observation space, action space, reward function, and state transitions. This includes: Based on the observation information of the intelligent agent, the action space is defined as the active and reactive power adjustment of the interaction between the transformer area and the feeder, including the power output or absorption achieved through photovoltaic and energy storage devices, and the power of the transformer area nodes is redistributed after the action is executed. The reward function of the intelligent agent is constructed according to the control target, including a reward term based on the target function and a penalty term based on the constraint conditions. The reward term reflects the achievement of the control target by calculating the node voltage deviation and power loss, and the penalty term punishes the violation of the energy storage device constraint conditions. From the initial state until the power flow has no feasible solution or the scheduling cycle ends, the agent calculates the reward function value after performing actions based on the current observation space and action space, and then enters the next time-instance state for state transition.

5. The method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes according to claim 1, characterized in that, Based on the optimization model, the dynamic cluster Markov decision model is solved using the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm to obtain the corresponding multi-area collaborative optimization and control strategy for the distribution network, including: An Actor network is constructed for each agent, and based on the agent's observation input, the corresponding active power and reactive power adjustment values ​​are output. A Critic network is constructed for each agent. The value of each action is evaluated based on the set of observations and actions of all agents in the current state. The output of the control Critic network is updated based on the parameters of the target Critic network. For each agent, a target Actor network and a target Critic network are set up, and the parameters of the training network are periodically copied to the target network to maintain parameter stability during reinforcement learning; In each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, the parameters of the Critic network are optimized by gradient descent to minimize the joint regression loss, and the parameters of the Actor network are optimized by gradient ascent to maximize the action value of the agent. Based on the iterative training results, the corresponding multi-region collaborative optimization and control strategy of the distribution network is obtained, and the power regulation amount of the region is output based on the current observation space and the optimized action value.

6. The method for coordinated optimization and control of multiple distribution areas in a distribution network adapting to topology changes according to claim 5, characterized in that, The step involves constructing a Critic network for each agent, evaluating the value of each action based on the set of observations and actions of all agents in the current state, wherein the output of the Critic network is controlled based on parameter updates of the target Critic network, including: The Critic network input for each agent includes environmental and attention information about the power distribution network environment. Weights are generated for each agent's observations and actions through an attention mechanism. The correlation between the current agent and other agents is calculated using a key-value query method. Weighted parameters are then generated to describe the degree of contribution of different agents to the value assessment. The environmental and attention information of the agent is fused, and the fused features are input into the Critic network. Through activation functions and fully connected networks, the value output of the agent's actions is generated. Specifically, a joint regression loss function is defined based on the outputs of the Critic network and the target Critic network, and the Critic network is updated by minimizing the loss function.

7. A multi-area collaborative optimization and control device for a distribution network that adapts to topology changes, characterized in that, The device includes: The cluster partitioning module is used to partition the distribution network into clusters based on electrical distance modularity and power balance using the SLM cluster partitioning algorithm. The optimization model construction module is used to: construct an optimization model containing objective functions of minimizing voltage deviation and minimizing power loss based on the cluster partitioning results. The objective functions of the optimization model include the calculation of node voltage deviation and cluster power loss within the cluster, and the constraints include power balance, node voltage safety, photovoltaic reactive power adjustment range, and preset range of SOC state of energy storage devices. The agent modeling module is used to: treat each transformer area as an agent, and construct a dynamic clustering Markov decision model for each transformer area based on the agent's observation space, action space, reward function, and state transition settings. The optimization solution module is used to: solve the dynamic grouping Markov decision model based on the optimization model using the multi-agent deep deterministic policy gradient (MADDPG) algorithm, obtain the corresponding multi-area collaborative optimization control strategy of the distribution network, and control the distribution network according to the control strategy; The agent modeling module includes: The observation space definition unit is used to: treat each transformer area as an intelligent agent, and define the observation space as the real-time power flow information of all nodes in the transformer area according to the current cluster boundary and the state of the transformer area nodes, including the active power, reactive power and voltage amplitude of the nodes, and dynamically adjust the observation space to adapt to the new cluster structure when the cluster boundary changes.

8. The multi-area collaborative optimization and control device for distribution networks adapting to topology changes according to claim 7, characterized in that, The cluster partitioning module includes: Initialize the cluster unit, which is used to: treat each transformer area in the distribution network as an independent initial cluster, and obtain the node connection relationship and load power data of all transformer areas; The iterative optimization unit is used to: use predefined electrical distance modularity and power balance as indicators for cluster partitioning, attempt to assign each transformer area to other clusters in each iteration, calculate the change in electrical distance modularity after allocation, record the allocation method that produces the maximum modularity gain, and repeat this process to gradually optimize cluster partitioning until the modularity no longer changes. The subnetwork merging unit is used to generate several subnetworks using the SLM cluster partitioning algorithm, treat each subnetwork as a virtual node to participate in the next round of cluster partitioning, and continue to optimize the modularity until it reaches the global optimum.

9. The multi-area collaborative optimization and control device for distribution networks adapting to topology changes according to claim 7, characterized in that, The optimization model construction module includes: The objective function definition unit is used to: define objective functions based on the cluster partitioning results, with minimizing voltage deviation and minimizing power loss as optimization objectives respectively. The voltage deviation objective function characterizes the sum of squares of the deviations between the voltage amplitudes of all nodes in the cluster and the voltage reference value. The power loss objective function is calculated through the branch resistance and the flow of active and reactive power in the cluster. The power balance constraint unit is used to: calculate the power balance relationship of each node based on the line admittance matrix and the voltage phase angle difference between nodes according to the power balance principle within the cluster, so as to perform power balance constraint of the distribution network; The safety constraint unit is used to: control the voltage amplitude of all nodes within a safe threshold range based on a preset range, while limiting the branch current to not exceed the maximum allowable current; The photovoltaic regulation unit is used to: determine the active power output of the photovoltaic power generation system through the maximum power point tracking mode (MPPT), and set the reactive power regulation range of the photovoltaic system according to the reactive power capacity of the photovoltaic power generation system. The energy storage management unit is used to: control the current SOC state of the energy storage system to be within a safe threshold range based on a preset range, and set the charging and discharging power range of the energy storage device in conjunction with the charging and discharging efficiency of the energy storage system.

10. The multi-area collaborative optimization and control device for distribution networks adapting to topology changes according to claim 7, characterized in that, The agent modeling module includes: The action space definition unit is used to: define the action space as the active and reactive power adjustment quantities of the interaction between the transformer area and the feeder based on the observation information of the intelligent agent, including the power output or absorption realized through photovoltaic and energy storage devices, and redistribute the power of the transformer area nodes after the action is executed. The reward function construction unit is used to: construct a reward function for the intelligent agent according to the control target, including a reward term based on the target function and a penalty term based on the constraint conditions, wherein the reward term reflects the achievement of the control target by calculating the node voltage deviation and power loss, and the penalty term punishes the violation of the energy storage device constraint conditions; The state transition unit is used to: calculate the reward function value after the agent performs actions based on the current observation space and action space from the initial state until the power flow has no feasible solution or the scheduling cycle ends, and then enter the next time step state to perform state transition.

11. The multi-area collaborative optimization and control device for distribution networks adapting to topology changes according to claim 7, characterized in that, The optimization solution module includes: The Actor network construction unit is used to: construct an Actor network for each agent, and output the corresponding active power and reactive power adjustment based on the agent's observation input; The Critic network building unit is used to: build a Critic network for each agent, evaluate the value of each action based on the set of observations and actions of all agents in the current state, and control the output of the Critic network based on the parameter updates of the target Critic network. The target network setting unit is used to: set up a target Actor network and a target Critic network for each agent, and periodically copy the parameters of the training network to the target network to maintain parameter stability during reinforcement learning; The network parameter optimization unit is used to: in each training iteration, based on the actions generated by the Actor network and the values ​​evaluated by the Critic network, optimize the parameters of the Critic network using gradient descent to minimize the joint regression loss, and optimize the parameters of the Actor network using gradient ascent to maximize the action value of the agent. The control strategy generation unit is used to: obtain the corresponding multi-region collaborative optimization control strategy of the distribution network based on the iterative training results, and output the power regulation amount of the region based on the current observation space and the optimized action value.

12. The multi-area collaborative optimization and control device for distribution networks adapting to topology changes according to claim 11, characterized in that, The Critic network building unit includes: The environmental information input unit is used to input environmental and attention information about the power distribution network environment into the Critic network for each agent. The attention weight generation unit is used to: generate weights for the combination of observations and actions of each agent through the attention mechanism, calculate the correlation between the current agent and other agents using the key-value query device, and generate weighted parameters to describe the degree of contribution of different agents to the value assessment. The feature fusion unit is used to: fuse the environmental information and attention information of the agent, input the fused features into the Critic network, and generate the value output of the agent's actions through the activation function and the fully connected network; The network update unit is used to: define a joint regression loss function based on the output of the Critic network and the output of the target Critic network, and update the Critic network by minimizing the loss function.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in any one of claims 1 to 6.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes, as described in any one of claims 1 to 6.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the method for coordinated optimization and control of multiple distribution areas in a distribution network that adapts to topology changes as described in any one of claims 1 to 6.

Images