A three-phase power distribution network voltage optimization control method based on vertical and horizontal flexible resource cooperation and large language model enhanced federal reinforcement learning

Abstract

The application belongs to the technical field of power system operation control, and relates to a three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource cooperation and large language model enhanced federal reinforcement learning. The method sets up a three-phase distribution network power flow model and defines a voltage optimization problem, sets a state space, an action space and an initial reward function of multi-agent reinforcement learning; each regional agent obtains a low-dimensional state vector by using a large language model and a data distillation algorithm, constructs a phase correlation matrix by using phase characteristics, and generates weighted features; each regional agent independently performs local reinforcement learning training, performs constraint solving based on the three-phase distribution network power flow model, generates a semantic abstract uploaded to a computing center, generates global strategy guidance information, and issues the global strategy guidance information to each agent; the trained agent is deployed to a distribution network control system, action instructions are generated according to real-time collected system states, and vertical and horizontal flexible resources are cooperatively controlled.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, specifically to a three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning. Background Technology

[0002] As the proportion of distributed photovoltaic (PV) power in three-phase distribution networks continues to increase, the strong uncertainty and rapid fluctuations in PV output lead to frequent voltage overruns and increased fluctuations in the distribution network, seriously affecting the safe operation and power quality of the distribution network.

[0003] In existing technologies, traditional voltage regulation methods for power distribution networks mainly rely on rigid mechanical voltage regulating equipment, such as on-load tap-changing transformers and capacitor banks, to achieve voltage regulation through tap switching or capacitor switching. Although this type of method can achieve base voltage control, it generally suffers from problems such as slow response speed, discrete regulation process, and equipment lifespan limitations due to frequent operation, making it difficult to adapt to the rapid voltage fluctuations brought about by high proportions of distributed photovoltaic (PV) grid integration.

[0004] To address the aforementioned issues, some studies have employed model-driven methods based on optimal power flow for voltage optimization control. However, these methods typically rely on precise network topology and line parameters, which are difficult and uncertain to obtain in actual distribution networks. Furthermore, under conditions of three-phase imbalance and large-scale nodes, the model dimensions are high, and the solution complexity is large, making it difficult to meet real-time control requirements.

[0005] In recent years, data-driven methods such as reinforcement learning have been increasingly applied to the field of distribution network voltage control. Centralized reinforcement learning methods, which gather state information from the entire network for unified decision-making, can reduce the dependence on accurate models to some extent. However, they have high requirements for communication bandwidth, computing resources, and measurement integrity. Furthermore, they require centralized transmission of operational data from each node, posing a risk of data privacy leakage. At the same time, as the system scales up, the state dimension increases dramatically, affecting the scalability of the algorithm.

[0006] To alleviate the shortcomings of centralized methods, federated reinforcement learning trains models locally in each region and aggregates them at the center, which reduces the privacy risks associated with data sharing to some extent. However, this type of method still requires frequent transmission of high-dimensional model parameters between each node and the center, which can easily lead to significant communication overhead and latency under limited communication bandwidth conditions, affecting the efficiency of collaborative training.

[0007] Furthermore, existing reinforcement learning methods for distribution network voltage control typically employ manually set static reward function weights, making it difficult to dynamically adjust the optimization objective based on actual operating conditions. This can lead to unstable control performance under multiple constraints and objectives. Simultaneously, in three-phase modeling, most methods simply concatenate the three-phase states, failing to effectively characterize inter-phase coupling relationships and making it difficult to adapt to three-phase unbalanced operating conditions, thus limiting further improvements in control performance. Therefore, this invention provides a three-phase distribution network voltage optimization control method based on longitudinal and lateral flexible resource collaboration and large language model-enhanced federated reinforcement learning to address the aforementioned shortcomings of existing technologies, which is highly necessary. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problem in the prior art that it is difficult to achieve efficient coordinated control of vertical and horizontal flexible resources under the conditions of three-phase imbalance, high photovoltaic penetration and limited measurement. The invention designs a three-phase distribution network voltage optimization control method based on the coordination of vertical and horizontal flexible resources and large language model enhanced federated reinforcement learning to solve the technical problems existing in the prior art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning, comprising the following steps: Step S1: Establish a three-phase power flow model of flexible resource coordination in the vertical and horizontal directions and define the voltage optimization problem. Transform the voltage optimization problem into a partially observable Markov decision process and set the state space, action space and initial reward function of multi-agent reinforcement learning. Step S2: Each region's intelligent agent uses a large language model to parse unstructured information, performs weighted processing on the original high-dimensional measurement features, and uses a data distillation algorithm to reduce the dimensionality of the features, obtaining low-dimensional state vectors and local low-dimensional observations; Step S3: Each region agent constructs a phase correlation matrix based on the low-dimensional state vector using phase features. During the feature extraction process, the intra-phase autocorrelation and inter-phase coupling effects in the state features are explicitly quantified to generate weighted features. Step S4: Each regional agent independently conducts local reinforcement learning training, interacts with the power distribution network simulation environment, receives low-dimensional state vectors, local low-dimensional observations and weighted features, performs constraint solving based on the three-phase power distribution network flow model, generates semantic summaries and uploads them to the computing center to replace model parameters for cross-regional information fusion analysis and policy reasoning, generates global policy guidance information and distributes it to each agent; Step S5: Deploy the trained agent to the power distribution network control system, generate control commands based on the real-time operating status, and send them to the local controllers of each device through the power distribution network communication system to perform coordinated voltage regulation of vertical and horizontal flexible resources.

[0010] The beneficial effects of this invention are as follows: This invention establishes a three-phase distribution network power flow model with flexible resource coordination in both vertical and horizontal directions, and transforms the voltage optimization problem into a partially observable Markov decision process. This enables intelligent modeling of voltage optimization under conditions of three-phase imbalance, high photovoltaic penetration, and limited measurement, improving the voltage control capability and regulation margin of the three-phase unbalanced distribution network under conditions of severe photovoltaic output fluctuations and uneven spatiotemporal load distribution. Each regional agent utilizes a large language model to parse unstructured information and generate semantic priors. Combined with a data distillation algorithm, weighted evaluation and dimensionality reduction are performed to obtain low-dimensional state vectors and local low-dimensional observations, improving the accuracy and robustness of state perception. The phase correlation matrix constructed based on the low-dimensional state vectors and the weighted features explicitly quantify intra-phase autocorrelation and inter-phase coupling effects, enhancing feature representation capabilities and improving control performance under unbalanced conditions, providing high-quality input for reinforcement learning. By enhancing the federated reinforcement learning collaborative training framework with a large language model that facilitates semantic information interaction, agents in each region can generate and upload semantic summaries to the computing center for cross-regional information fusion and policy reasoning while training independently locally. This forms global policy guidance information, compensating for the lack of global information caused by partial observability of agents in each region. It also achieves dual privacy protection for data and models, and efficient collaborative control of vertical and horizontal flexible resources. Finally, the trained agents are deployed to the power distribution network control system. Based on real-time system status data, action commands are generated and sent to the local controllers of each device, ensuring stable operation and voltage optimization of the three-phase power distribution network under complex conditions. This effectively solves the problem of efficient collaboration of vertical and horizontal flexible resources that is difficult to achieve in existing technologies.

[0011] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0013] Figure 1This is a flowchart of a three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0015] Example: like Figure 1 As shown in the figure, this embodiment provides a voltage optimization control method for a three-phase distribution network based on vertical and horizontal flexible resource coordination and large language model-enhanced federated reinforcement learning. This method, based on vertical and horizontal flexible resource coordination and large language model-enhanced federated reinforcement learning, aims to achieve efficient voltage optimization and regulation of the three-phase distribution network under complex operating conditions, and includes the following steps: Step S1: Establish a three-phase power flow model of flexible resource coordination in the vertical and horizontal directions and define the voltage optimization problem. Transform the voltage optimization problem into a partially observable Markov decision process and set the state space, action space and initial reward function of multi-agent reinforcement learning. The specific steps of step S1 are as follows: S11. Establish a three-phase power flow model for the distribution network that includes the hybrid transformer HT, the smart soft switch SOP, and the distributed inverter resource IBR; Step S11 specifically includes: Set up a three-phase unbalanced radial distribution network with N+1 nodes. The common coupling point is located at node 0 and connected to the main network. and Let p(j) be the set of distribution network nodes and lines, respectively, where p(j) represents the direct parent node of node j. , Indicates each phase.

[0016] S111. Hybrid modeling of the hybrid transformer HT is performed based on the equivalent voltage source model and the power injection model. The hybrid transformer HT includes a series inverter and a parallel inverter. The voltage level of a single feeder is longitudinally adjusted by injecting compensation voltage through the series inverter, and reactive power support is provided through the parallel inverter. The hybrid inverter HT connecting nodes i and j is constrained by the maximum adjustable voltage, internal active power exchange, and inverter capacity as shown in equation (1), so that the hybrid transformer HT has the ability to continuously adjust voltage along the feeder direction and perform longitudinal voltage control: (1) in and yes Phase node and nodes voltage amplitude, This is for the flow through the internal impedance of Phase current, and These represent the three-phase rated capacities of the parallel inverter and the series inverter, respectively. The parallel inverter is located at the node... supply The internal active power and reactive power It is by and Produced.

[0017] The series inverter described in step S111 is compensated by dynamically injecting a voltage. To perform longitudinal voltage regulation, the parallel inverter provides reactive power compensation at node j. To maintain voltage stability, the hybrid transformer HT, through the coordinated operation of series and parallel inverters, forms a continuous voltage regulation capability along the line direction within a single feeder, transforming voltage regulation from traditional discrete regulation into a continuous and controllable process, providing key support for vertically refined voltage regulation.

[0018] S112. Model the intelligent soft switch SOP between node i and node j. The intelligent soft switch SOP is equipped with a back-to-back voltage source converter structure, which enables it to have real-time control capability of active power and reactive power and to operate flexibly in four quadrants. On this basis, establish the power balance relationship, loss model and capacity constraint of the SOP, and apply the following constraint condition (2) to its operation process, so that the intelligent soft switch SOP has the ability to bidirectionally adjust active power and independently adjust reactive power between different feeders, and to perform horizontal power regulation: (2) in and It is an inflow Phase node and outflow nodes SOP active power and It's power loss. and It is the loss coefficient of a certain section of the converter. It is the three-phase rated capacity of SOP.

[0019] S113. Model the distributed inverter resource IBR. The distributed inverter resource IBR model defines the active and reactive power output capabilities and operating constraints. Connect photovoltaic devices and energy storage devices to node i, enabling them to have reactive power regulation capabilities and active power regulation capabilities respectively, and regulate the local voltage. On this basis, establish the constraint relationship between photovoltaic devices and inverter capacity and the constraint relationship between energy storage devices and inverter capacity and charge / discharge operation, and form a multi-level collaborative regulation system with hybrid transformer HT and smart soft switch SOP. The inverter capacity constraint mentioned in step S113 is:

[0020] The charging and discharging operation constraints mentioned in step S113 are:

[0021] in, and These are the active power and reactive power output by the photovoltaic inverter, respectively. It refers to the installed capacity of photovoltaic systems. and These are the discharge and charging power of the energy storage device, respectively. and These are the discharge and charging power limits for energy storage. It is the state of charge of the stored energy and is at its minimum value. With the maximum value between, Indicates charge / discharge efficiency. Indicates the time step interval. This represents the charged state at the initial time step.

[0022] S114. Establish a three-phase distribution network power flow model, and define a voltage optimization problem based on voltage amplitude constraints and line capacity constraints. Optimize with minimizing operating costs as the optimization objective. During the optimization process, consider the longitudinal voltage regulation effect of the hybrid transformer HT and the lateral power regulation effect of the intelligent soft switch SOP to perform coordinated optimization of voltage control and power distribution. The three-phase distribution network power flow model mentioned in step S114 is as follows:

[0023] The voltage amplitude constraint and line capacity constraint define the voltage optimization problem expression as follows:

[0024] The operating cost includes electricity purchase cost, energy storage charging and discharging loss cost, and smart soft switching operating loss cost; the expression for minimizing the operating cost is:

[0025] in, and Representing nodes respectively of The net injected active and reactive power of the phase. and For nodes in the node admittance matrix of AND node of Corresponding elements, For nodes of AND node of Voltage phase angle difference between phases, , , , The weight coefficients for each optimization objective are... and This indicates the maximum and minimum values ​​of the voltage amplitude. This indicates a line capacity limitation.

[0026] By modeling the hybrid transformer HT and the smart soft switch SOP, their longitudinal voltage regulation and lateral power regulation capabilities are clarified. Combined with various capacity and power constraints, reliable technical means are provided for the safe operation and efficient utilization of flexible resources in the distribution network. Based on the above modeling, the hybrid transformer HT and the smart soft switch SOP work together from two dimensions: voltage regulation within the feeder and power redistribution between feeders. This breaks through the limitations of the traditional single-device voltage regulation method and realizes the coupled optimization of longitudinal voltage control and lateral power regulation. This provides a structural foundation for subsequent reinforcement learning strategies to achieve global optimal control.

[0027] S12. The entire power distribution system is divided into multiple control areas according to electrical distance. Each area is managed by an agent m. Considering the measurement constraints, the voltage optimization problem is transformed into a partially observable Markov decision process of reinforcement learning. An observation space, action space, and initial reward function are configured for each agent. The reward function includes voltage over-limit penalty and economic operating cost penalty. The actions of each agent directly affect the hybrid transformer HT and the intelligent soft switch SOP to carry out coordinated control of internal voltage regulation and power interaction between feeders.

[0028] The observation space configuration process described in step S12 is as follows: Intelligent agent Get the responsible area The available measurement information includes the set of observable nodes. Each phase load power and reactive power Active power of photovoltaic inverters and reactive power Active power of SOP and reactive power The charging and discharging power of energy storage ( ) and state of charge HT's reactive power support and node voltage amplitude The collected data are combined in sequence to form the observation vector of agent m:

[0029] intelligent agent observation space Defined as the set of all possible observation vectors, i.e.

[0030] The observation vector The core data form constituting the original high-dimensional measurement features is output to step S2 for data distillation and feature dimensionality reduction processing.

[0031] The motion space configuration process described in step S12 is as follows: Within its designated area m, agent m acquires action information based on the actual output capabilities of the adjustable devices and incorporates it as part of its action space. Specifically, the agent... The operation includes the compensation voltage of the hybrid transformer HT. With reactive power support Active power at both ends of the intelligent soft switch SOP and reactive power , Reactive power of photovoltaic inverters and the charging and discharging power of energy storage devices The agent m combines the above actions in sequence to form an action vector.

[0032]

[0033] The feasible range of the action vector must satisfy all the physical constraints established in steps S111 to S114, including the capacity constraints, charge and discharge constraints, and power loss constraints of HT, SOP, photovoltaic inverter and energy storage device, i.e., constraint conditions (1)-(7), so as to ensure that the control decision generated based on the original high-dimensional measurement characteristics meets the physical feasibility of actual distribution network operation.

[0034] intelligent agent Action space Defined as the set of all possible action vectors

[0035] The reward function configuration process in step S12 is as follows: Each node is checked to see if the voltage amplitude of each observation node within the responsible region m exceeds the upper and lower limits. For each node exceeding the upper limit, the excess amount is accumulated as a penalty value; for each node below the lower limit, the shortfall amount is also accumulated, resulting in the region voltage exceeding the limit penalty. Based on constraint condition (6), the intelligent agent m reads the real-time active and reactive power output of each adjustable device in the region, multiplies it by the corresponding weighting coefficient β, and accumulates them to form the economic target penalty. , combined and Generate reward function :

[0036] in This indicates a penalty for exceeding voltage limits. This represents the penalty for economic goals. The agent's goal is to maximize rewards. ,and and The weighting coefficients are for the two optimization objectives; the reward function is calculated based on the voltage state and operating cost reflected by the original high-dimensional measurement features, and is used to guide the optimization direction of subsequent reinforcement learning strategies.

[0037] By dividing the power distribution network into multiple control areas and defining the observation space, action space, and reward function for each agent, distributed intelligent decision-making under locally observable conditions is realized, enabling the equipment regulation in each area to effectively maintain voltage safety while taking into account economic operation.

[0038] It should be noted that step S1, through detailed modeling of the hybrid transformer HT and the intelligent soft-switching SOP, fully utilizes the longitudinal voltage regulation and reactive power support capabilities of series and parallel inverters to achieve dynamic adjustment of the voltage of a single feeder. The back-to-back converter structure of the intelligent soft-switching SOP allows for flexible transfer of active power between different feeders, while simultaneously enabling independent adjustment of reactive power, thus enhancing the lateral power regulation capability of the distribution network. The modeling, combined with equipment capacity constraints and power loss constraints, ensures that flexible resources operate within a safe range and stabilizes voltage during system regulation. This provides a reliable technical foundation for the coordinated control of longitudinal and lateral flexible resources in the distribution network, and also provides an operable physical model for subsequent optimization and reinforcement learning control. By integrating the hybrid transformer HT and the smart soft switch SOP into the voltage optimization control framework in a coordinated manner, a multi-dimensional flexible regulation mechanism for three-phase unbalanced distribution networks is formed. The action decisions of each intelligent agent directly affect key flexible regulation equipment, including the hybrid transformer HT and the smart soft switch SOP, so that the learned control strategy can explicitly coordinate the longitudinal voltage regulation and the lateral power interaction process, thereby achieving coordinated voltage optimization between multiple regions.

[0039] Step S2: Each region's intelligent agent uses a large language model to parse unstructured information, performs weighted processing on the original high-dimensional measurement features, and uses a data distillation algorithm to reduce the dimensionality of the features, obtaining low-dimensional state vectors and local low-dimensional observations; The specific steps of step S2 are as follows: S21: The agents in each region use a large language model to perform semantic parsing on unstructured information related to the operation of the distribution network, extract key state variables that affect voltage stability and power flow distribution, and generate state variable importance assessment results. The specific steps of step S21 are as follows: S211. Each regional agent selects an open-source large language model from the Qwen or Deepseek series, utilizing its general logical reasoning capabilities and embedded prior physical knowledge related to the power system as a tool for unstructured information parsing. Preset prompts are constructed to guide the large language model in performing tasks, and these prompts are input along with the unstructured information into the large language model. The prompts guide the large language model in parsing distribution network operation-related information and outputting the importance assessment results of state variables. The prompts include the following: "You are an expert in the field of distribution network operation and optimization control, familiar with three-phase distribution network power flow, voltage control, distributed energy access characteristics, and measurement system configuration. Your task is to analyze unstructured information related to distribution network operation, including weather forecasts, equipment operation logs, and distribution network topology descriptions, and based on this information, determine which state variables have the most critical impact on voltage stability and power flow distribution."

[0040] Please output the following based on the input information: Identify key nodes, determine which state variables are most important, assign an importance weight (0~1) to each state variable, and recommend the set of state variables to be retained first for subsequent data distillation algorithms; output in JSON structured format. S212. Use for a period of time including Historical data at each time step or continuously collected numerical measurement data of the distribution network of the same length are used to construct a global original high-dimensional state vector. ,in As a characteristic dimension. To eliminate the differences between different electrical dimensions such as voltage per-unit values ​​and power in kilowatts, [the following is used]: Normalization is performed to obtain the standard high-dimensional state vector. .

[0041] S213. Synchronous Acquisition and... Unstructured information such as meteorological texts, equipment operation logs, and distribution network topology information for the corresponding time period is standardized and keywords are extracted, and then converted into structured natural language descriptions, including time, meteorological conditions, equipment operating status, and distribution network structure information. For example: "Time: 12:00 PM, August 20, 2025; Temperature: 35 degrees Celsius; Weather: Sunny; Light Intensity: Very High; Log: 'Voltage limit exceeded by photovoltaic access node X' 'State of charge of energy storage access node X is X'...", "This distribution network has a radial structure, with substation node Bus650. The main line is: Bus650 → Bus632 → Bus633 → Bus634. Branch lines branch off at Bus632: Bus632 → Bus671 → Bus684."

[0042] The structured text and the prompt words are concatenated or fused and input into the large language model. The large language model performs multi-dimensional semantic reasoning tasks based on the input, including identifying key nodes and key equipment that affect voltage stability and identifying state variables that affect the operation of the distribution network. It also sorts the importance of various state variables, assigns a quantitative importance weight in the range of 0 to 1 to each state variable, and selects a set of key state variables that are prioritized for retention.

[0043] The semantic reasoning results described above are organized into state variable importance assessment results, including the set of key state variables and their corresponding importance weight diagonal matrices. This is used for subsequent data distillation and feature weighting operations, thereby optimizing the low-dimensional state input of the reinforcement learning agent.

[0044] The above results are output in JSON structured form. By performing semantic parsing on the unstructured information through a large language model, the key state variables that have the greatest impact on voltage stability and power flow distribution can be identified. A quantitative importance weight is assigned to each state variable, providing reliable prior information for subsequent data distillation and feature weighting, and enhancing the reinforcement learning agent's ability to understand the complex power distribution network operating environment.

[0045] S22. Construct a weighted data distillation mechanism to map high-dimensional state vectors into low-dimensional compact state representations, generating low-dimensional observation inputs usable by the agent; The specific steps of step S22 are as follows: S221. Transform the standard high-dimensional state vector diagonal matrix of weights Combined, calculate the weighted covariance matrix. :

[0046] S222. To Perform eigenvalue decomposition to obtain ,in It is a diagonal matrix whose eigenvalues ​​are arranged in descending order. This represents the corresponding eigenvector matrix. Based on the set cumulative variance contribution rate, such as 95% or 90%, the top [values] are selected. The eigenvectors corresponding to the largest eigenvalues, where Construct the distillation-reduced projection matrix .

[0047] S223. Use a weighted diagonal matrix. With distillation dimensionality reduction projection matrix Build a data distillation model and store it offline; S224. During reinforcement learning training or actual operation, for any time step t, the agent calls the currently normalized high-dimensional observation vector. Real-time dimensionality reduction is performed using a data distillation model to obtain a low-dimensional state vector. ; combined with S1 Intelligent agents can be obtained Local low-dimensional observations .

[0048] By employing a weighted data distillation mechanism, high-dimensional state vectors are mapped to low-dimensional, compact representations, effectively reducing the dimensionality of the state space, improving the efficiency of reinforcement learning training, and accelerating convergence. Simultaneously, by combining the weight matrix and the dimensionality-reduced projection matrix, the low-dimensional state vectors retain key feature information, ensuring the accuracy and robustness of the agent's decisions.

[0049] It should be noted that by combining semantic parsing of a large language model with weighted data distillation, an efficient transformation from raw high-dimensional measurement features to low-dimensional compact state representations is achieved. This reduces the dimensionality of the reinforcement learning state space and the training complexity, while retaining the most important feature information for voltage regulation and power flow distribution, thus improving the training efficiency, convergence speed, and decision accuracy of the agent. Furthermore, the structured output of prior information about state variables can provide consistent and interpretable key state references for different regions or different agents, enhancing the overall performance of the distributed control system.

[0050] Step S3: Each region agent constructs a phase correlation matrix based on the low-dimensional state vector using phase features. During the feature extraction process, the intra-phase autocorrelation and inter-phase coupling effects in the state features are explicitly quantified to generate weighted features. The specific steps of step S3 are as follows: S31. The set of observable nodes within the control area of ​​agent m for each region. Each node in the process is assigned a distinct feature. This is used to identify and distinguish the A, B, and C phase information to which a node belongs. The phase characteristics are typically fixed and directly obtainable after the line construction is completed, and are represented using one-hot encoding.

[0051] Combine the phase characteristics of all nodes in sequence to construct the phase correlation matrix:

[0052] Will According to local low-dimensional observation The corresponding node numbers are transformed and concatenated to obtain the corresponding and observations with phase information .

[0053] By constructing a phase correlation matrix, the phase information of each node is encoded, enabling the agent to clearly distinguish different phases when processing low-dimensional observations, thus ensuring the integrity and consistency of multiphase control features.

[0054] S32. Will and The input is taken into an independent multilayer perceptron, and features are fused and mapped through a nonlinear activation function to obtain abstract features. The attention embedding module will The input is fed into a linear layer for attention calculation, using three independent learnable weight matrices. , , Calculate the query, key, and value vector: Calculate the dot product of q and k. This is used to assess the original correlation between different characteristic elements under the current state of the distribution network. Combined with the phase correlation matrix after transformation and splicing and phase weight The attention matrix is ​​generated by normalization using the SoftMax function:

[0055] in It is a scaling factor that prevents the gradient vanishing problem. It is the phase relationship coefficient. It is an identity matrix. Used to express learnable weights and physics knowledge The representation of in-phase autocorrelation and inter-phase cross-phase coupling provides a priori importance reference for attention generation.

[0056] Subsequently, the attention embedding module will... AND value vector Multiplication yields weighted features The phase attention embedding module explicitly quantifies the intra-phase autocorrelation and cross-phase coupling relationships between nodes through matrix operations, thereby improving the agent's strategy learning efficiency and coordinated voltage regulation capability in complex unbalanced distribution networks.

[0057] By using a multilayer perceptron for feature fusion and quantifying the intraphase autocorrelation and cross-phase coupling relationships between nodes through a phase attention embedding module, the weighted features can comprehensively reflect the operating status of complex unbalanced distribution networks, thereby improving the pertinence and accuracy of policy learning.

[0058] It should be noted that by combining phase-specific features and low-dimensional state vectors, the state information of multiple nodes and phases can be systematically captured during feature extraction, achieving a close integration of physical characteristics and intelligent features. The phase attention embedding module performs matrix quantization on the autocorrelation and cross-phase coupling between nodes, enabling the agent to identify key inter-phase coupling relationships and improving the adaptability of the control strategy to unbalanced distribution networks. The generated weighted features can effectively fuse and efficiently map multi-node and multi-phase information, providing technical support for the stability and rapid convergence of the subsequent reinforcement learning agent.

[0059] Step S4: Each regional agent independently conducts local reinforcement learning training, interacts with the power distribution network simulation environment, receives low-dimensional state vectors, local low-dimensional observations and weighted features, performs constraint solving based on the three-phase power distribution network flow model, generates semantic summaries and uploads them to the computing center to replace model parameters for cross-regional information fusion analysis and policy reasoning, generates global policy guidance information and distributes it to each agent; The specific steps of step S4 are as follows: S41. Based on the observation space and local low-dimensional observation of each regional intelligent agent m and weighted features Deploy an open-source large language model in the computing center and configure preset prompt words for cross-regional collaborative analysis; Establish data channels based on semantic information interaction between each region and the computing center, and replace the traditional model parameter transmission method with semantic summarization; define the federated communication round structure, including the local training phase and the semantic information interaction phase; and finally construct a federated reinforcement learning collaborative training framework based on semantic information interaction, so that agents in each region can achieve collaborative training through semantic information without sharing the original measurement data and network parameters.

[0060] By constructing a federated reinforcement learning collaborative training framework based on semantic information interaction, agents in different regions can conduct collaborative training without sharing raw measurement data and network parameters, which ensures data privacy and security while reducing communication overhead.

[0061] S42. A federated reinforcement learning collaborative training framework based on semantic information interaction, in which each regional agent m independently conducts local reinforcement learning training. During the training process, the three-phase power flow model is combined to solve the constraints of low-dimensional state vectors, local low-dimensional observations and weighted features, update the control strategy parameters, and generate a semantic summary reflecting the operating state and constraint satisfaction after local training is completed. The training adopts the Soft Actor Critic (SAC) training algorithm.

[0062] Step S42 specifically includes: S421. Input the semantic embedding vector issued in the previous round of federated communication. Action bias Exploration rate adjustment And the reward weight adjustment amount initialized to a zero vector during the first round of training. Local training steps Experience replay buffer capacity Batch size Soft update coefficient Discount factor Learning rate , , ; Initialize the Actor network, the dual Critic network, and the corresponding target Critic network parameters; initialize the empirical replay region and temperature coefficient. Initialize target entropy; Initialize bias decay coefficient ; According to the Federal Communications Commission Adjusting the Actor network According to the Federal Communications Commission Adjusting the reward function and Initialize the statistical variable to 0: cumulative reward Maximum voltage deviation Voltage over-limit count ; S422. The agent updates its policy based on local low-dimensional observations and weighted features, setting the current training step to... The intelligent agent interacts with the power distribution network simulation environment, and the weighted features are... Input Actor network sampling action , Integrating Federal Action Bias + Output actions to the simulation environment; use a three-phase power flow model to solve for constraints on low-dimensional state vectors, local low-dimensional observations, and weighted features to obtain the next step of the original observations; calculate the reward function according to formula (10). Update statistical variables , , ; Empirical tuples The data is stored in the experience replay area; then, batches are sampled from the experience replay area, and the dual-Critic network parameters are updated according to the principles of the SAC algorithm. , The Critic network input additionally includes To assist the local network in perceiving global collaborative situational information; and to update the Actor network parameters. Update temperature coefficient Simultaneously, a soft update is performed on the Critic network; the training steps are... Change to Repeat the training steps until the preset number of steps is reached. .

[0063] After training, calculate the average reward for this round. Average voltage over-limit rate ; Statistical action strategy trend and constraint satisfaction state .

[0064] Combining power flow calculation results with constraint satisfaction, the data is converted into a natural language description using a predefined template, and a semantic summary is output. This concludes the local training session.

[0065] S423. Output the updated Actor network parameters. Critic network parameters , Semantic summarization The semantic summary is output in structured JSON format.

[0066] Introducing local low-dimensional observations and weighted features into local reinforcement learning training enables agents to efficiently learn policies and generate structured semantic summaries, improving the convergence speed of local training and the interpretability of policy representations.

[0067] S43. Semantic summarization and uploading to the computing center replaces the traditional federated learning method of uploading a huge network parameter model, performs cross-regional information fusion analysis and policy reasoning, generates global policy guidance information, and distributes it to each agent; The specific steps of step S43 are as follows: S431. Each regional agent will submit the semantic summary generated in this round of training. The data is uploaded to the computing center, which receives semantic summary data from all regions and calls the deployed open-source large language model. The semantic summary is then concatenated with preset prompt words to construct a unified input text. The preset prompt text in the computing center is: "You are a scheduling and coordination expert for multi-regional coordinated voltage control in a distribution network. The following are the operational semantic summaries submitted by the agents in each region after this round of training. Please comprehensively analyze the voltage control effect, flexible resource utilization, and constraint satisfaction of each region, and complete the following tasks: (1) Identify the weak areas and bottleneck equipment of the current network voltage control; (2) Analyze the potential coordinated strategies for power mutual assistance between regions through SOP; (3) Generate strategy adjustment suggestions for each agent, including suggested exploration directions, action bias adjustments, and reward weight corrections; (4) Generate strategy guidance vectors for each agent. The output is given in JSON structured form." The constructed unified input text is sent to the large language model to perform forward inference calculation.

[0068] S432. The large language model performs semantic parsing and logical reasoning operations on the input multi-region semantic summary, and jointly analyzes the voltage control effect, flexible resource utilization status, and constraint satisfaction of each region; based on the analysis results, it identifies weak regions and bottleneck devices in the overall network voltage control; based on the topological associations and operating status between regions, it calculates the coordinated adjustment path for power mutual assistance through SOP; for each agent, it generates policy adjustment results, including the exploration rate adjustment direction, action bias correction amount, and reward weight correction amount, and outputs the corresponding policy guidance text. .

[0069] S433. During the parsing of the strategy guidance text, the large language model dynamically determines the focus of the current control objective based on a comprehensive semantic understanding and logical reasoning of the voltage over-limit situation, equipment operating status, and economic indicators in each region. It then adaptively provides a reward weight correction amount in the generated strategy adjustment information. Specifically: Strategy guidance text output by large language models The text is parsed and processed to extract policy adjustment information from the text using keyword extraction, which is then converted into numerical form to obtain the action bias. Exploration rate adjustment and reward weight adjustment Simultaneously, the strategy guides the text input into the Transformer encoding structure within the large language model: through the tokenizer... Convert to a token sequence, perform forward reasoning, and obtain the last hidden state matrix. ;right Performing pooling operations yields a fixed-dimensional vector. Low-dimensional semantic embedding vectors are obtained through learnable linear projection layer mapping. Among them, projection layer parameters and Joint training and updating with Critic network parameters.

[0070] S434. The semantic embedding vector obtained above... Action bias Exploration rate adjustment and reward weight adjustment Combined to form global strategy guidance information And distribute it to the corresponding regional agents; in each round of federated communication, A calculation is performed and remains unchanged during the local training round. This result is then concatenated with the local state features and action vectors and input into the Critic network, enabling it to incorporate cross-regional collaborative information during the state-action value evaluation process.

[0071] The local reinforcement learning training process is combined with the semantic summary uploading and policy guidance information distribution process to form a complete federated semantic communication round. After completing one local training, a semantic information exchange is performed. Multiple federated communication rounds are executed in a loop according to the above process until the preset convergence judgment condition is met. Throughout the process, each regional agent only uploads the semantic summary and does not transmit the original measurement data or network parameters.

[0072] During policy reasoning, when the large language model parses the semantic summary and identifies the average voltage limit violation rate of agent m... Higher, or maximum voltage deviation When the safety threshold is exceeded, an increased voltage over-limit penalty weight is generated. And reduce the weight of economic targets. The adjustment amount, and through Distribute the information; when the intelligent agent... of When the output has dropped to zero or near zero, and there is sufficient equipment capacity margin, the generation rate decreases. and increase The adjustment amount will shift the focus of optimization to economy; when both voltage control performance and economic indicators are within the set range, the system will maintain or fine-tune the settings. and This allows for dynamic updates of reward weights during subsequent local training.

[0073] By using large language models for cross-regional semantic information fusion and policy reasoning, we can identify weak links across the entire network, generate targeted policy adjustment suggestions, and automatically output global guidance information in numerical form, thus achieving effective collaboration and dynamic optimization among intelligent agents.

[0074] It should be noted that this step combines low-dimensional state features, local reinforcement learning training, and large language model inference to achieve cross-regional collaborative optimization and adaptive policy updates. By replacing traditional parameter transmission with semantic summarization, communication bandwidth consumption is significantly reduced while ensuring the privacy of operational data in each region. Utilizing the semantic understanding and logical reasoning capabilities of the large language model, the operational states of multiple regions can be comprehensively analyzed, automatically generating global policy guidance and achieving intelligent and automated training. The iterative, round-by-round federated semantic communication enables the agent to continuously optimize policies while ensuring security and economic constraints, improving the overall stability and flexibility of the system.

[0075] Step S5: Deploy the trained agent to the power distribution network control system, generate control commands based on the real-time operating status, and send them to the local controllers of each device through the power distribution network communication system to perform coordinated voltage regulation of vertical and horizontal flexible resources.

[0076] The specific steps of step S5 are as follows: S51. Deploy the trained agent to each regional control unit, establish a power distribution network communication link to connect the local controllers of each device, and collect real-time operating status data of the power distribution network control system, including the voltage of each node, injected power, and three-phase imbalance. Then, input the collected operating status data into the agent to generate control commands.

[0077] S52. The intelligent agent sends the HT series compensation voltage setting value and the parallel reactive power setting value to the HT local controller according to the control command. The HT series inverter injects compensation voltage into the feeder according to the setting value, dynamically adjusting the voltage of each downstream node to achieve longitudinal increase or decrease along the feeder; the HT parallel inverter sets the reactive power at the installation node to provide local voltage support. The active and reactive power setpoints at both ends of the SOP are sent to the local controller of the SOP. The SOP then performs lateral transfer of active power and independent adjustment of reactive power between different feeders according to the setpoints. The reactive power setting value of the photovoltaic inverter and the charging and discharging power setting value of the energy storage are sent to each distributed resource controller. Under the premise of ensuring that the active power output remains unchanged, the photovoltaic inverter uses the remaining capacity to provide or absorb reactive power. The energy storage device adjusts the charging and discharging power to change the net active power injected into the node and performs local voltage regulation.

[0078] S53. During the execution of action commands, the phase attention embedding module is used to generate differentiated control commands for phases A, B, and C. Under three-phase imbalance conditions, the agent takes different amplitude or direction adjustment actions for different phases. For example, it reduces the HT compensation voltage for phase A with high voltage and increases the compensation voltage for phase C with low voltage, effectively alleviating the three-phase imbalance problem.

[0079] In actual long-term operation, the system monitors changes in operating conditions. If the operating conditions of the distribution network change significantly, such as the addition of distributed photovoltaic access, network topology adjustment, or long-term load drift, update step S2 and trigger step S4 to maintain the adaptability and control performance of the strategy network.

[0080] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A voltage optimization control method for a three-phase distribution network based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning, characterized in that, Includes the following steps: Step S1: Establish a three-phase power flow model of flexible resource coordination in the vertical and horizontal directions and define the voltage optimization problem. Transform the voltage optimization problem into a partially observable Markov decision process and set the state space, action space and initial reward function of multi-agent reinforcement learning. Step S2: Each region's intelligent agent uses a large language model to parse unstructured information, performs weighted processing on the original high-dimensional measurement features, and uses a data distillation algorithm to reduce the dimensionality of the features, obtaining low-dimensional state vectors and local low-dimensional observations; Step S3: Each region agent constructs a phase correlation matrix based on the low-dimensional state vector using phase features. During the feature extraction process, the intra-phase autocorrelation and inter-phase coupling effects in the state features are explicitly quantified to generate weighted features. Step S4: Each regional agent independently conducts local reinforcement learning training, interacts with the power distribution network simulation environment, receives low-dimensional state vectors, local low-dimensional observations and weighted features, performs constraint solving based on the three-phase power distribution network flow model, generates semantic summaries and uploads them to the computing center to replace model parameters for cross-regional information fusion analysis and policy reasoning, generates global policy guidance information and distributes it to each agent; Step S5: Deploy the trained agent to the power distribution network control system, generate control commands based on the real-time operating status, and send them to the local controllers of each device through the power distribution network communication system to perform coordinated voltage regulation of vertical and horizontal flexible resources.

2. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning as described in claim 1, characterized in that, The specific steps of step S1 are as follows: S11. Establish a three-phase power flow model for the distribution network that includes the hybrid transformer HT, the smart soft switch SOP, and the distributed inverter resource IBR; S12. The entire power distribution system is divided into multiple control areas according to electrical distance. Each area is managed by an agent m. The voltage optimization problem is transformed into a partially observable Markov decision process of reinforcement learning. An observation space, action space and initial reward function are configured for each agent. The reward function includes voltage over-limit penalty and economic operating cost penalty. The actions of each agent directly affect the hybrid transformer HT and the intelligent soft switch SOP to carry out coordinated control of internal voltage regulation of feeders and power interaction between feeders.

3. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning as described in claim 2, is characterized in that... Step S11 specifically includes: S111. Hybrid modeling of the hybrid transformer HT is performed based on the equivalent voltage source model and power injection model. The hybrid transformer HT includes a series inverter and a parallel inverter. For the hybrid inverter HT connecting node i and node j, constraints are established based on the maximum adjustable voltage, internal active power exchange, and inverter capacity, so that the hybrid transformer HT has continuous voltage regulation capability along the feeder direction and performs longitudinal voltage control. S112. Model the intelligent soft switch SOP, configure back-to-back voltage source converters between node i and node j, establish the power balance relationship, loss model and capacity constraints of the SOP, so that the intelligent soft switch SOP has the ability to bidirectionally allocate active power and independently regulate reactive power between different feeders, and perform horizontal power regulation. S113. Model the distributed inverter resource IBR, connect photovoltaic devices and energy storage devices at node i, establish the constraint relationship between photovoltaic devices and inverter capacity and the constraint relationship between energy storage devices and inverter capacity and charge and discharge operation, and form a multi-level collaborative regulation system with hybrid transformer HT and smart soft switch SOP. S114. Establish a three-phase distribution network power flow model, and define a voltage optimization problem based on voltage amplitude constraints and line capacity constraints. Optimize with minimizing operating costs as the optimization objective. During the optimization process, consider the longitudinal voltage regulation effect of the hybrid transformer HT and the lateral power regulation effect of the intelligent soft switch SOP to perform coordinated optimization of voltage control and power distribution. The three-phase distribution network power flow model mentioned in step S114 is as follows: The observation space configuration process described in step S12 is as follows: Intelligent agent Get the responsible area The available measurement information forms the observation vector of agent m. observation space Defined as the set of all possible observation vectors, i.e. The motion space configuration process described in step S12 is as follows: Within the area of ​​responsibility m, the agent m acquires action information and forms action vectors based on the actual output capabilities of the adjustable devices. intelligent agent Action space Defined as the set of all possible action vectors The reward function mentioned in step S12 is:

4. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning as described in claim 3, is characterized in that... The specific steps of step S2 are as follows: S21: The agents in each region use a large language model to perform semantic parsing on unstructured information related to the operation of the distribution network, extract key state variables that affect voltage stability and power flow distribution, and generate state variable importance assessment results. S22. Construct a weighted data distillation mechanism to map high-dimensional state vectors into low-dimensional compact state representations, generating low-dimensional observation inputs usable by the agent.

5. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning according to claim 4, characterized in that, The specific steps of step S21 are as follows: S211. Each regional intelligent agent selects an open-source large language model as a tool for parsing unstructured information, constructs preset prompt words to guide the large language model to perform tasks, and inputs them into the large language model along with unstructured information as input conditions. The prompt words are used to guide the large language model to parse distribution network operation-related information and output the state variable importance evaluation results. S212. Use for a period of time including Historical data at each time step or continuously collected numerical measurement data of the distribution network of the same length are used to construct a global original high-dimensional state vector. ,in For feature dimension, Normalization is performed to obtain the standard high-dimensional state vector. ; S213. Synchronous Acquisition and... Unstructured information corresponding to the time period is standardized and keywords are extracted, and then converted into structured natural language descriptions, including time, weather conditions, equipment operating status and power distribution network structure information; The structured text and the prompt words are concatenated or fused and input into a large language model. The large language model performs a multidimensional semantic reasoning task based on the input and organizes the semantic reasoning results into state variable importance evaluation results, including a set of key state variables and their corresponding importance weight diagonal matrix. .

6. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning according to claim 5, characterized in that, The specific steps of step S22 are as follows: S221. Transform the standard high-dimensional state vector diagonal matrix of weights Combined, calculate the weighted covariance matrix. : S222. To Perform eigenvalue decomposition to obtain ,in It is a diagonal matrix whose eigenvalues ​​are arranged in descending order. This represents the corresponding eigenvector matrix; based on the set cumulative variance contribution rate, the top... The eigenvectors corresponding to the largest eigenvalues, where Construct the distillation-reduced projection matrix ; S223. Use a weighted diagonal matrix. With distillation dimensionality reduction projection matrix Build a data distillation model and store it offline; S224. During reinforcement learning training or actual operation, for any time step t, the agent calls the currently normalized high-dimensional observation vector. Real-time dimensionality reduction is performed using a data distillation model to obtain a low-dimensional state vector. ; combined with S1 Intelligent agents can be obtained Local low-dimensional observations .

7. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning as described in claim 6, is characterized in that... The specific steps of step S3 are as follows: S31. The set of observable nodes within the control area of ​​agent m for each region. Each node in the process is assigned a distinct feature. The phase-specific features are typically fixed and directly obtainable after the line construction is completed. These phase-specific features are represented using one-hot encoding. Combine the phase characteristics of all nodes in sequence to construct the phase correlation matrix: Will According to local low-dimensional observation The corresponding node numbers are transformed and concatenated to obtain the corresponding and observations with phase information ; S32. Will and The input is taken into an independent multilayer perceptron, and features are fused and mapped through a nonlinear activation function to obtain abstract features. ; The attention embedding module will The input is fed into a linear layer for attention calculation, using three independent learnable weight matrices. , , Calculate the query, key, and value vector: Calculate the dot product of q and k. , Combined with the phase correlation matrix after transformation and splicing and phase weight The attention matrix is ​​generated by normalization using the SoftMax function: in It is a scaling factor. It is the phase relationship coefficient. It is the identity matrix; Used to express learnable weights and physics knowledge This represents the in-phase autocorrelation and interphase cross-phase coupling relationships; The attention embedding module will AND value vector Multiplication yields weighted features .

8. The three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning according to claim 7, characterized in that, The specific steps of step S4 are as follows: S41. Based on the observation space and local low-dimensional observation of each regional intelligent agent m and weighted features Deploy an open-source large language model in the computing center and configure preset prompt words for cross-regional collaborative analysis; establish a data channel based on semantic information interaction between each region and the computing center, and replace the traditional model parameter transmission method with semantic summarization; define a federated communication round structure, including a local training phase and a semantic information interaction phase, and build a federated reinforcement learning collaborative training framework based on semantic information interaction. S42. A federated reinforcement learning collaborative training framework based on semantic information interaction, in which each regional agent m independently conducts local reinforcement learning training. During the training process, the three-phase power flow model is combined to solve the constraints of low-dimensional state vectors, local low-dimensional observations and weighted features, update the control strategy parameters, and generate a semantic summary reflecting the operating state and constraint satisfaction after local training is completed; the training adopts the soft actor-commentator training algorithm. S43. Semantic summarization is uploaded to the computing center for cross-regional information fusion analysis and policy reasoning, generating global policy guidance information, which is then distributed to each agent.

9. A three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning as described in claim 8, characterized in that, The specific steps of step S43 are as follows: S431. Each regional agent will submit the semantic summary generated in this round of training. The data is uploaded to the computing center, which receives semantic summary data from all regions and calls the deployed open-source large language model. The semantic summary is then concatenated with preset prompts to construct a unified input text, which is then fed into the large language model to perform forward inference computation. S432. The large language model performs semantic parsing and logical reasoning operations on the input multi-region semantic summary. Based on the topological associations and operational states between regions, it calculates the coordinated adjustment path for power mutual assistance through SOP and outputs the corresponding policy guidance text. ; S433. During the parsing of the strategy guidance text, the large language model dynamically determines the focus of the current control objective based on a comprehensive semantic understanding and logical reasoning of the voltage over-limit situation, equipment operating status, and economic indicators in each region. It then adaptively provides a reward weight correction amount in the generated strategy adjustment information. Specifically: Strategy guidance text output by large language models The text is parsed and processed to extract policy adjustment information from the text using keyword extraction, which is then converted into numerical form to obtain the action bias. Exploration rate adjustment and reward weighting adjustment ; The strategy-guided text is input into the Transformer encoding structure inside the large language model: it is then processed by a tokenizer... Convert to a token sequence, perform forward reasoning, and obtain the last hidden state matrix. ;right Performing pooling operations yields a fixed-dimensional vector. Low-dimensional semantic embedding vectors are obtained through learnable linear projection layer mapping. Among them, projection layer parameters and Joint training and updating with Critic network parameters; S434. The semantic embedding vector obtained above... Action bias Exploration rate adjustment and reward weighting adjustment Combined to form global strategy guidance information And distribute it to the corresponding regional agents; In each round of federal communications Perform a calculation, which remains unchanged during the local training round, and then concatenate it with the local state features and action vectors before inputting it into the Critic network.

10. A three-phase distribution network voltage optimization control method based on vertical and horizontal flexible resource collaboration and large language model-enhanced federated reinforcement learning according to claim 9, characterized in that, The specific steps of step S5 are as follows: S51. Deploy the trained agent to each regional control unit, establish a power distribution network communication link to connect the local controllers of each device, collect the operating status data of the power distribution network control system in real time, input the collected operating status data into the agent, and generate action commands. S52. The intelligent agent sends the HT series compensation voltage setpoint and parallel reactive power setpoint to the HT local controller according to the control command. The HT series inverter injects compensation voltage into the feeder according to the setpoint, dynamically adjusting the voltage of each downstream node to achieve longitudinal increase or decrease along the feeder. The HT parallel inverter sets the reactive power at the installation node to provide local voltage support. The active power and reactive power setpoints at both ends of SOP are sent to the SOP local controller. SOP performs lateral transfer of active power and independent adjustment of reactive power between different feeders according to the setpoint. The reactive power setpoint of the photovoltaic inverter and the charging and discharging power setpoint of the energy storage are sent to each distributed resource controller. The photovoltaic inverter uses its remaining capacity to provide or absorb reactive power, and the energy storage device adjusts the charging and discharging power to change the net injected active power at the node, thereby regulating the local voltage. S53. During the execution of action instructions, the phase attention embedding module is used to generate differentiated control instructions for the three phases A, B, and C. Under the three-phase unbalanced working condition, the agent takes adjustment actions of different amplitudes or directions for different phases. In actual long-term operation, the system monitors changes in operating conditions. If the operating conditions of the distribution network change significantly, update step S2 and trigger step S4.

Images