A multi-target coordinated control method and related device for multi-area distributed photovoltaic systems

By constructing a hybrid reinforcement learning training model that integrates prior guidance and environmental interaction experience, the problem of insufficient control strategy sensitivity of multi-station distributed photovoltaic coordinated control methods under unconventional operating conditions is solved, and efficient multi-objective coordinated control is achieved.

CN123121841APending Publication Date: 2026-11-17ZHONGSHAN POWER SUPPLY BUREAU OF GUANGDONG POWER GRID
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611321915.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multi-station distributed photovoltaic coordinated control methods are not sensitive enough to rare high-risk states under unconventional operating conditions, and traditional data augmentation methods are difficult to meet the high-quality training requirements of reinforcement learning strategies in complex operating scenarios.

Method used

By selecting and expanding the dataset, a hybrid reinforcement learning training model is constructed, which integrates prior guidance experience and environmental interaction experience, iteratively updates the network parameters of multiple agents, and generates a multi-objective coordinated control model to achieve efficient control of unconventional working conditions.

Benefits of technology

It significantly enhances the control robustness and risk response capability under unconventional operating conditions, stably outputs optimized control commands for adjustable resources in each distribution area, and realizes efficient multi-objective coordinated control of distributed photovoltaic systems in multiple distribution areas.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121841A_ABST
    Figure CN123121841A_ABST
Patent Text Reader

Abstract

This invention provides a multi-objective coordinated control method and related device for multi-area distributed photovoltaic systems, belonging to the field of distributed photovoltaic control. The method first uses a target dataset from unconventional operating scenarios as a guide to selectively select source datasets for normal operating scenarios. Then, it combines the selected datasets with the target dataset to perform targeted sample expansion to construct an enhanced training set. Finally, it employs a hybrid reinforcement learning model that integrates prior guidance experience and environmental interaction experience to iteratively train the multi-agent decision-making architecture. The resulting multi-objective coordinated control model significantly enhances control robustness and risk response capabilities under unconventional operating conditions, stably outputs control commands for adjustable resources in each area, and achieves efficient multi-objective coordinated control of multi-area distributed photovoltaic systems. This solves the problem of insufficient sensitivity of existing control methods to rare high-risk states and fully meets the needs of high-quality training and reliable control of reinforcement learning strategies in complex operating scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of control technology for multi-area distributed photovoltaic power in power system distribution networks, specifically relating to a multi-objective coordinated control method and related devices for multi-area distributed photovoltaic power. Background Technology

[0002] With the large-scale integration of distributed photovoltaic, energy storage, electric vehicles and flexible loads into the distribution network, the traditional distribution system is gradually transforming from a one-way power supply and passive regulation operation mode into a new operation mode with multi-entity participation, multi-energy coordination and interaction between source, grid, load and storage.

[0003] In multi-distribution distribution systems, distributed photovoltaics typically have characteristics such as dispersed access points, strong power output fluctuations, multiple ownership entities, and obvious local interest linkages. Their operating status not only affects the voltage level and network loss level of their own distribution area, but also affects the operational safety of adjacent distribution areas through feeder power flow, voltage coupling, and reactive power support relationships.

[0004] The operational risks of distribution networks with a high proportion of distributed photovoltaic (PV) grid integration often concentrate under unconventional operating conditions. If reinforcement learning control strategies are trained solely on historical data from normal operating conditions, the resulting strategies may lack sensitivity to rare, high-risk states, potentially leading to decreased control effectiveness or even strategy failure in some scenarios. Furthermore, traditional data augmentation methods have limited applicability in power system scenarios, making it difficult to support high-quality training of reinforcement learning strategies under complex operating conditions. Summary of the Invention

[0005] Based on this, it is necessary to provide a multi-objective coordinated control method and related device for multi-area distributed photovoltaic systems to address the above-mentioned technical problems, so as to overcome the problems that existing multi-area distributed photovoltaic coordinated control methods are not sensitive enough to rare high-risk states under unconventional operating conditions, and that traditional data augmentation methods are difficult to meet the high-quality training requirements of reinforcement learning strategies in complex operating scenarios.

[0006] In a first aspect, the present invention provides a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas, comprising the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0007] Furthermore, based on the source dataset corresponding to the normal operating scenario and the target dataset corresponding to the unconventional operating scenario, the source dataset is filtered to obtain a filtered dataset covering the unconventional operating conditions, including: By combining quantifiable and non-quantifiable prior knowledge, a hybrid prior constraint system is constructed. Quantifiable prior knowledge is used to ensure the physical compliance of the selected samples, while non-quantifiable prior knowledge is used to guide the selected samples to match the distribution characteristics of unconventional operating conditions. Using a hybrid prior constraint system as the screening criterion, the source dataset is subjected to distribution alignment screening to obtain a filtered dataset that matches the distribution characteristics of the target dataset and satisfies the physical operation rules.

[0008] Furthermore, based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set, including: Using the union of the filtered dataset and the target dataset as a reference basis, an initial augmented sample set is generated using a large language model driven by structured cue words; Verify the distribution consistency between the initial expanded sample set and the reference base. If the verification fails, adjust the structured prompt words until the samples generated by the large language model meet the distribution consistency requirements. The validated initial augmented sample set is merged with the filtered dataset to obtain the enhanced training set.

[0009] Furthermore, the hybrid reinforcement learning training is based on a control framework that integrates a large language model and reinforcement learning; the training process includes: The action suggestions generated by combining the large language model with the power grid operation state reasoning and the corresponding environmental interaction results are integrated to obtain the first state-action-next state-reward quadruple, which is then stored in the prior knowledge experience pool. Based on the operational data generated by the interaction between the multi-agent decision-making architecture and the actual power grid environment, a second state-action-next state-reward quadruple is obtained and stored in the interaction experience pool. Batch experience data are sampled from the prior knowledge experience pool and the interaction experience experience pool respectively, and loss is calculated to obtain the total policy loss; With minimizing the total policy loss as the optimization objective, the network parameters of the multi-agent decision architecture are iteratively updated to complete a single round of hybrid reinforcement learning training iteration.

[0010] Furthermore, batches of experience data are sampled from the prior knowledge experience pool and the interaction experience experience pool, respectively, and loss calculations are performed to obtain the total policy loss. With minimizing the total policy loss as the optimization objective, the network parameters of the multi-agent decision-making architecture are iteratively updated, including: Based on the batch of experience data sampled from the prior knowledge experience pool, calculate the imitation learning loss; Based on the batch experience data sampled from the interaction experience pool, the multi-agent reinforcement learning policy loss and value function loss are calculated respectively. The total policy loss is obtained by weighting and fusing the imitation learning loss and the multi-agent reinforcement learning policy loss by balancing hyperparameters. The decision network parameters of the multi-agent decision architecture are updated with the goal of minimizing the total policy loss, and the evaluation network parameters of the multi-agent decision architecture are updated with the goal of minimizing the value function loss.

[0011] Furthermore, the pre-construction steps of the multi-agent decision-making architecture include: Various adjustable resources within a multi-area distribution network are abstracted into independent intelligent agents, forming a multi-agent set; By setting a local observation range for each agent, the local observation space of each agent is obtained. By combining the local observation spaces of each intelligent agent, a global state space of the system is constructed; Aggregate the ranges of decision variables from all agents to form a globally distributed action space; Based on the operation optimization objectives of multi-regional distribution networks, a multi-objective reward function is set. Based on a multi-agent set, a local observation space, a system global state space, a global distributed action space, and a multi-objective reward function, a decentralized partially observable multi-agent decision-making architecture is constructed.

[0012] Secondly, the present invention provides a multi-area distributed photovoltaic multi-objective coordinated control device, comprising: The data filtering module is used to filter the source dataset based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, so as to obtain the filtered dataset covering the non-normal operation conditions. The data augmentation module is used to augment the dataset based on the filtered dataset and the target dataset to obtain an enhanced training set. The model building module is used to perform hybrid reinforcement learning training on a pre-built multi-agent decision-making architecture based on the enhanced training set, so as to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. The coordination control module is used to output control commands for adjustable resources in each transformer area using a multi-objective coordination control model.

[0013] Thirdly, the present invention provides a computer device, the device including a processor and a memory: The memory is used to store computer programs and send the instructions of the computer programs to the processor; The processor executes the following steps according to the instructions of the computer program: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0014] Fourthly, the present invention provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0015] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, performs the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0016] In summary, this invention provides a multi-objective coordinated control method and related device for multi-area distributed photovoltaic systems. The method first uses a target dataset of unconventional operating scenarios as a guide to selectively filter the source dataset of normal operating scenarios, extracting effective samples covering unconventional operating conditions from existing normal operating condition data. This addresses the problem of scarce high-risk operating condition samples and unbalanced data distribution in the original training set from the data source. Furthermore, it combines the filtered dataset with the target dataset to perform targeted sample expansion to construct an enhanced training set, effectively expanding the sample size and scenario coverage of unconventional operating scenarios. Finally, it employs a hybrid reinforcement learning model that integrates prior guidance experience and environmental interaction experience. The method iteratively trains a multi-agent decision-making architecture. With the support of full-scenario data in the enhanced training set, it not only leverages prior guidance experience to rapidly improve the policy's perception sensitivity and response adaptability to rare high-risk states, but also continuously optimizes the accuracy of multi-objective coordinated control through environmental interaction and iteration. The resulting multi-objective coordinated control model can significantly enhance the control robustness and risk response capability under unconventional operating conditions, stably output optimized control commands for adjustable resources in each transformer area, and realize efficient multi-objective coordinated control of distributed photovoltaic systems in multiple transformer areas. This effectively solves the problem of insufficient sensitivity of existing control methods to rare high-risk states and fully meets the needs of high-quality training and reliable control of reinforcement learning strategies in complex operating scenarios. Attached Figure Description

[0017] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas, as described in one embodiment of the present invention. Figure 2 This is a control framework diagram for integrating a large language model and reinforcement learning in one embodiment of the present invention; Figure 3 This is a topology diagram of an improved IEEE 33-node multi-station distribution system in one example of the present invention; Figure 4 This is a training curve diagram of the MAPPO-LLM model in one example of the present invention; Figure 5 This is a typical daily network loss curve under four methods in a calculation example of the present invention; Figure 6 This is a typical daily node average voltage deviation rate curve under four methods in a calculation example of the present invention; Figure 7 This is a comparison chart of MMD index results under different data construction methods in a calculation example of the present invention; Figure 8 This is a graph showing the comparison of physical constraint satisfaction rates under different data augmentation methods in a case study of the present invention. Figure 9 This is a comparison chart of the average coverage rate of unconventional working conditions under different data construction methods in a calculation example of the present invention; Figure 10 This is a heatmap showing the coverage of different data construction methods under various key unconventional operating conditions in a calculation example of the present invention; Figure 11 This is a structural block diagram of a multi-target coordinated control device for distributed photovoltaic systems in multiple areas, according to one embodiment of the present invention. Figure 12 This is an internal structural diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] It should be noted that the user information (including but not limited to user images, user portrait information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0022] To better understand the technical solution and innovation of this invention, the background technology of this invention and the problems existing in the prior art will be described in detail below.

[0023] With the large-scale integration of distributed photovoltaic (PV), energy storage, electric vehicles, and flexible loads into distribution networks, traditional distribution systems are gradually shifting from a unidirectional power supply and passive regulation mode to a new operational model involving multiple stakeholders, multi-energy coordination, and interaction between power sources, grids, loads, and storage. Particularly in multi-distribution-area distribution systems, distributed PV typically exhibits characteristics such as dispersed connection points, strong output fluctuations, diverse ownership entities, and significant local interest linkages. Its operational status not only affects the voltage and network loss levels of its own area but also impacts the operational safety of adjacent areas through feeder power flow, voltage coupling, and reactive power support relationships. Therefore, achieving coordinated control of multiple resources such as distributed PV, energy storage, wind turbines, and static var compensators across multiple distribution areas has become a crucial issue for improving the economic efficiency, safety, and power quality of high-proportion renewable energy distribution networks.

[0024] Existing methods for coordinated control of multi-area distributed photovoltaic (PV) systems mainly include droop control, optimized power flow control, and intelligent control methods based on reinforcement learning. Droop control offers advantages such as fast response and simple implementation, but it relies heavily on local measurement information, making it difficult to simultaneously address power flow coupling between multiple areas and global operational objectives. Traditional optimized power flow methods can reduce network losses and improve voltage distribution to some extent, but they typically rely on relatively complete system models and global information. When facing rapid fluctuations in distributed PV systems, random load changes, and decentralized decision-making by multiple agents, their temporal adaptability and online decision-making capabilities remain limited. In recent years, multi-agent deep reinforcement learning has provided a new approach to solving distributed coordinated control problems, enabling multi-resource coordinated optimization through interactive learning between agents. However, in complex distribution network scenarios, it still faces challenges such as large training sample requirements, low initial exploration efficiency, and insufficient generalization ability for unconventional operating conditions.

[0025] It is particularly important to note that the operational risks of distribution networks with a high proportion of distributed photovoltaic (PV) grid integration often concentrate under unconventional operating conditions, such as typhoons, rapid cloud cover, abnormal load changes during holidays, peak-valley load mismatch across distribution areas, and the superposition of high PV loads and localized low loads. These scenarios are characterized by low frequency of occurrence, high disturbance intensity, tight operating boundaries, and scarce historical samples. If reinforcement learning control strategies are trained solely on historical data from normal operating conditions, the resulting strategies may lack sensitivity to rare, high-risk states, potentially leading to decreased control effectiveness or even strategy failure under scenarios such as voltage exceeding limits, power flow overload, or sudden changes in PV output. Therefore, constructing a high-quality training dataset that covers unconventional operating states, meets physical constraints, and closely approximates the distribution of target operating conditions is a crucial foundation for improving the robustness and generalization ability of multi-agent control strategies.

[0026] Meanwhile, traditional data augmentation methods have limitations in their applicability to power system scenarios. While random perturbations, interpolation oversampling, or generative models can expand the sample size, they often fail to fully embed the physical constraints of the power system, such as node voltage, branch power flow, upper and lower limits of photovoltaic output, and energy storage operation boundaries. They also struggle to accurately represent non-quantifiable scenario patterns such as "typhoons causing sustained low photovoltaic output," "cloud cover causing short-term severe fluctuations," and "cold waves driving a continuous increase in residential load." In other words, relying solely on numerical statistical features to construct training samples can easily lead to problems such as a mismatch between sample distribution and unconventional operating conditions, a high proportion of physically abnormal samples, and insufficient coverage of key scenarios, making it difficult to support high-quality training of reinforcement learning strategies in complex operating scenarios.

[0027] Large language models possess strong semantic understanding, knowledge reasoning, and condition generation capabilities, enabling them to integrate and express power system operation experience, scenario descriptions, numerical constraints, and control objectives in the form of natural language or structured prompts. Introducing large language models into a multi-agent reinforcement learning framework can, on the one hand, assist in constructing augmented datasets for unconventional operating conditions that integrate quantifiable and non-quantifiable priors, ensuring that generated samples possess statistical rationality, physical feasibility, and scenario relevance; on the other hand, it can provide action suggestions with engineering experience characteristics in the early stages of reinforcement learning training, reducing ineffective trial and error caused by random exploration and improving policy convergence speed and the quality of initial actions. Therefore, integrating the prior knowledge of large language models with multi-agent reinforcement learning methods provides a new technical approach for solving the multi-agent collaborative control problem in high-proportion distributed photovoltaic distribution networks.

[0028] Based on the above background, this patent addresses the need for coordinated operation of distributed photovoltaic (PV) systems in multiple areas. It proposes a multi-objective coordinated control method and related devices for distributed PV systems in multiple areas based on a large language model and multiple agents. By addressing issues such as the scarcity of samples for unconventional operating conditions, low initial exploration efficiency of reinforcement learning, and the difficulty of multi-agent collaborative optimization, it constructs a dataset generation mechanism that integrates hybrid prior knowledge, designs a MAPPO-LLM multi-agent reinforcement learning framework assisted by a large language model, and takes network loss reduction and voltage deviation suppression as core optimization objectives. This enables coordinated control of multiple resources, including distributed PV, energy storage, wind turbines, and static var compensators, in multi-area scenarios. This invention helps improve the operational economy, power quality, and control robustness of multi-area distribution systems under complex operating conditions, providing technical support for the safe and stable operation of new distribution networks under conditions of high-proportion distributed renewable energy access.

[0029] In one embodiment, such as Figure 1 As shown, a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas is provided, including the following steps: S101: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions.

[0030] The source dataset refers to the complete historical record set of photovoltaic power generation output and node load under normal operating conditions; the target dataset refers to the historical data subset corresponding to unconventional power grid scenarios such as typhoons, rapid cloud cover, and abnormal load changes during holidays, representing the core operating range where the control strategy needs to achieve the highest robustness and generalization capability; the filtered dataset refers to the sample set obtained from the source dataset, whose distribution characteristics are aligned with the target dataset and can cover unconventional operating conditions.

[0031] It should be noted that the coverage of operating conditions in the training data directly determines the generalization ability of the control strategy. Conventional techniques often directly use all historical data of normal operating conditions as the training set, which can easily lead to the failure of the strategy in boundary scenarios due to the lack of rare high-risk operating condition samples. This step aims to selectively filter the source dataset based on the distribution characteristics of the target dataset, extracting effective samples from the existing normal operating condition data that are similar in distribution to those of unconventional operating conditions. This reduces the distribution offset between normal and unconventional operating conditions at the data level, providing scenario-specific basic data support for strategy training.

[0032] S102: Based on the filtered dataset and the target dataset, the dataset is expanded to obtain an enhanced training set.

[0033] Among them, the enhanced training set refers to the training data set formed by merging the selected dataset with the expanded unconventional working condition samples, which is used to provide sample support covering multiple operating scenarios for the training of reinforcement learning models.

[0034] It should be noted that this step uses both the filtered dataset and the target dataset as a reference basis, and expands the sample size and scene coverage of unconventional working conditions through targeted sample expansion to solve the problem that the sample size after filtering is insufficient to support high-quality training.

[0035] S103: Based on the enhanced training set, perform hybrid reinforcement learning training on the pre-constructed multi-agent decision-making architecture to obtain a multi-objective coordinated control model adapted to unconventional working conditions; hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agent to generate a multi-objective coordinated control model.

[0036] Among them, the multi-agent decision-making architecture refers to a distributed decision-making framework composed of multiple agents with autonomous decision-making capabilities, built for the scenario of multi-area distributed photovoltaic collaborative control; the hybrid reinforcement learning training refers to a training mode that integrates two types of experience sources, namely prior guidance experience and environmental interaction experience, and iteratively updates the network parameters of the agents through dual-path optimization; the multi-objective coordinated control model refers to a distributed photovoltaic coordinated control model that can simultaneously achieve multiple operational optimization objectives such as power quality disturbance suppression and network loss reduction after training.

[0037] It should be noted that this step is based on an augmented training set covering unconventional working conditions. The model is trained using a hybrid reinforcement learning mode that integrates dual-source experience. This mode not only improves the quality of actions and convergence speed in the early stages of training by leveraging prior experience, but also continuously optimizes the environmental adaptability of the strategy through environmental interaction, ultimately resulting in a control model that can adapt to unconventional working conditions.

[0038] S104: Using a multi-objective coordinated control model, output control commands for adjustable resources in each transformer area.

[0039] Among them, the control command refers to the operating setting parameters of adjustable resources in each transformer area output by the multi-objective coordinated control model, which is used to guide adjustable resources such as distributed photovoltaics to perform coordinated control actions.

[0040] It should be noted that this step deploys the trained multi-objective coordinated control model to the operating environment. By inputting the real-time grid operating status, the corresponding control commands can be output, realizing online multi-objective coordinated control of distributed photovoltaic power stations in multiple areas.

[0041] The method provided in this embodiment first uses the target dataset of unconventional operating scenarios as a guide to selectively filter the source dataset of normal operating scenarios. It extracts effective samples covering unconventional operating conditions from existing normal operating condition data, improving the problem of scarce high-risk sample data and unbalanced data distribution in the original training set from the data source. Then, it combines the filtered dataset with the target dataset to perform targeted sample expansion to construct an enhanced training set, effectively expanding the sample size and scenario coverage of unconventional operating scenarios. Finally, it employs a hybrid reinforcement learning model that integrates prior guidance experience and environmental interaction experience to iteratively apply the multi-agent decision-making architecture. The training method, supported by full-scenario data in the enhanced training set, not only leverages prior guidance experience to rapidly improve the policy's sensitivity and response adaptability to rare high-risk states, but also continuously optimizes the accuracy of multi-objective coordinated control through environmental interaction and iteration. The resulting multi-objective coordinated control model can significantly enhance the control robustness and risk response capability under unconventional operating conditions, stably output optimized control commands for adjustable resources in each distribution area, and realize efficient multi-objective coordinated control of distributed photovoltaic systems in multiple distribution areas. This effectively solves the problem of insufficient sensitivity of existing control methods to rare high-risk states and fully meets the needs of high-quality training and reliable control of reinforcement learning strategies in complex operating scenarios.

[0042] To construct a complete training dataset that covers unconventional operating conditions, this invention first clarifies the rules for dividing the basic data domain: the complete historical records of photovoltaic (PV) power output and node loads under normal operating conditions are defined as the source dataset, denoted as . The target dataset is defined as a subset of historical data corresponding to unconventional power grid scenarios (such as typhoons, rapid cloud cover, and abnormal load changes during holidays). This represents the core operating range within which the control strategy needs to achieve the highest robustness and generalization capability. However, this is achieved by sampling the source dataset only randomly. Directly selecting training samples cannot effectively reduce the size of the source dataset. With the target dataset The distribution shift between these parameters results in the final training dataset being insufficiently sensitive to rare but crucial operating states for system safety. This deficiency significantly reduces the generalization performance of the policy network, particularly limiting its control effectiveness under boundary conditions or high-stress operating conditions.

[0043] To mitigate this problem, in one exemplary embodiment, prior knowledge is introduced. (The following will be uniformly referred to as) (This is represented as a guiding mechanism for the data selection and augmentation process.) Accordingly, based on the source dataset corresponding to the normal operating scenario and the target dataset corresponding to the unconventional operating scenario, the source dataset is filtered to obtain a filtered dataset covering the unconventional operating conditions, including: S201: Combine quantifiable and non-quantifiable prior knowledge to construct a hybrid prior constraint system; quantifiable prior knowledge is used to ensure the physical compliance of the selected samples, while non-quantifiable prior knowledge is used to guide the selected samples to match the distribution characteristics of unconventional operating conditions.

[0044] Among them, the hybrid prior constraint system refers to the set of data screening constraints composed of quantifiable and non-quantifiable prior knowledge, denoted as: (1) In the formula, This represents quantifiable prior knowledge, including the probability distribution of photovoltaic output and load, grid parameters, operating constraints, etc. All such knowledge can be characterized through explicit mathematical models, statistical analysis, or probability-based formulas. This indicates non-quantifiable prior knowledge, such as the impact of unconventional weather on sudden fluctuations in photovoltaic power output and load. This type of knowledge is difficult to model directly using formal mathematical expressions.

[0045] For example, when constructing a hybrid prior constraint system, the physical rules and statistical characteristics of the distribution network operation are first sorted out, and the contents that can be accurately represented by mathematical formulas, such as the probability distribution of photovoltaic power output, the upper and lower limits of node voltage, and the transmission capacity of lines, are extracted to form quantifiable prior knowledge. Furthermore, the evolution patterns of unconventional scenarios summarized from engineering operation experience, such as the sharp drop in photovoltaic output when clouds rapidly obstruct the view and the time-specific shifts in load during holidays, are incorporated into the system as constraints to form non-quantifiable prior knowledge. Together, these two constitute the constraint benchmark for the entire data screening process.

[0046] S202: Using a hybrid prior constraint system as the screening basis, the source dataset is subjected to distribution alignment screening to obtain a filtered dataset that matches the distribution characteristics of the target dataset and satisfies the physical operation rules.

[0047] Distribution alignment screening refers to a sample screening process with the core objective of reducing the distribution offset between the source data subset and the target dataset; the filtered dataset refers to the output of the screening process, which is a subset of the source dataset, whose distribution characteristics are as close as possible to the target dataset and satisfy the prior knowledge constraints.

[0048] For example, for precise quantization of source datasets With the target dataset The degree of distribution shift in the feature space is assessed using the Maximum Mean Discrepancy (MMD) as the core evaluation metric. The core principle of MMD is to measure the Euclidean distance between the mean embedding vectors of two probability distributions in the Reproducing Kernel Hilbert Space (RKHS), thereby achieving an unbiased estimate of the difference between distributions. Its specific expression is defined as follows: (2) In the formula, It is the square of the maximum mean difference, used to measure the degree of distribution shift between two datasets; and The source datasets are respectively and target dataset The probability distribution; , To distribute from the source dataset The feature vector of the sample sample that is independently sampled in the middle. , To distribute from the target dataset The feature vector of the sample independently sampled in the middle; This is the mathematical expectation operator, used to calculate the statistical mean of a random variable; is a kernel function used to map the original low-dimensional features to a high-dimensional Hilbert space to capture nonlinear distribution features.

[0049] During the sample selection process, it is necessary to ensure that the distribution of the selected dataset approximates the target dataset as closely as possible. The distribution of satisfies the following optimization objective: (3) In the formula, This is the filtered dataset, i.e., the output of the sample filtering process. This is the minimum point operator, which indicates the value of the independent variable that makes the objective function reach its minimum value.

[0050] Due to non-quantifiable prior knowledge Since it is difficult to quantify precisely, this embodiment uses it as an additional constraint in the sample selection process to ensure that the selected samples conform to the prior distribution constraint, specifically as follows: (4) In the formula, To balance the maximum mean difference (MMD) with prior knowledge Hyperparameters of constraint weights; This is a priori regularization term used to ensure that the selected samples meet the requirements of prior knowledge. Its specific expression is defined as follows: (5) In the formula, Specifically, it is the KL divergence (relative entropy) between the dataset distribution and the quantifiable prior distribution, which measures the degree of difference between the two probability distributions. It takes a non-negative value, and the larger the value, the greater the difference between the two distributions. When the two distributions are exactly the same, the value is 0. It represents the probability distribution corresponding to quantifiable prior knowledge, covering domain knowledge that can be characterized through explicit mathematical models and statistical analysis, such as the probability distribution of photovoltaic power output and load, and the operating constraints of power grid parameters. This is a sample feature vector in the dataset, containing operational features such as photovoltaic output, load level, and node voltage; To filter sample features under the dataset distribution x The probability of its occurrence; By employing the sample selection process described above, the reduction in the number of samples after selection can be effectively achieved under the constraint of prior knowledge. With the target dataset The distribution differences between them can improve the quality of the screened samples and their practicality in engineering.

[0051] It should be noted that MMD is a classic metric in the field of machine learning for measuring the difference between two probability distributions. It can measure the difference of nonlinear distributions without making prior assumptions about the data distribution, and is suitable for the characteristics of power grid operation data, which are mostly nonlinear and non-Gaussian distributions. KL divergence is often used to measure the degree of difference between two probability distributions. It takes a non-negative value, and the larger the value, the greater the difference between the two distributions. In this embodiment, it is used as a prior regularization term, which can effectively prevent the selected samples from deviating from the physical operation rules of the power grid.

[0052] This embodiment constructs a hybrid prior constraint system and conducts distribution alignment screening. Under the joint guidance of quantifiable and non-quantifiable prior knowledge, it effectively reduces the distribution difference between the screened dataset and the target dataset, and improves the scenario matching degree and engineering practicality of the screened samples.

[0053] Although a training subset that is distributionally aligned and physically valid has been constructed during the sample selection phase, the selected dataset... The sample size is still relatively limited, insufficient to support the training requirements of robust strategies in highly dynamic scenarios. To address this issue, in an exemplary embodiment, a data augmentation strategy based on a Large Language Model (LLM) is proposed. The LLM is used as a condition generator to synthesize high-quality samples under the joint guidance of both quantifiable and non-quantifiable prior knowledge. Based on this data augmentation strategy, the dataset is augmented using the selected dataset and the target dataset to obtain an augmented training set, including: S301: Using the union of the filtered dataset and the target dataset as a reference basis, generate an initial expanded sample set using a large language model driven by structured cue words.

[0054] The reference basis refers to the union of the target dataset and the filtered source dataset. The initial expanded sample set serves as the data reference foundation for the large language model to generate samples; structured prompt words refer to structured query instructions that embed domain-specific goals and constraints, used to guide the large language model to generate power grid operation samples that meet the requirements; This refers to the first set of synthetic samples generated by the large language model that has not yet undergone distribution consistency verification.

[0055] For example, to achieve controllable and scene-aware sample generation, this embodiment designs a prompt-driven generation mechanism: the prompt is a structured query instruction that embeds domain-specific goals and constraints, and the sample generation process can be formally represented as follows: (6) In the formula, Sample sets generated for large language models; This serves as a cognitive reasoning chain (or reasoning context) to enhance the interpretability of the generation process. For the conditional generation function of the large language model, this embodiment uses a general large language model with domain knowledge embedding capability as the generator; The target dataset consists of historical data from unconventional power grid scenarios such as typhoons, rapid cloud cover, and abnormal load changes during holidays. This is a subset of the source dataset obtained after distribution alignment filtering. The target dataset and the filtered source dataset are the union of the target dataset and the source dataset, serving as the reference basis for generating samples for the large language model; the structured prompt words are used to embed the task objectives and constraints. This refers to prior knowledge of the domain.

[0056] In practice, structured prompt words The design comprises three core components: a) Task description module: Define the objective of sample generation, such as simulating the operation status of photovoltaic / load under unconventional weather conditions; b) Sample constraint module: based on domain prior knowledge The generated constraints include reasonable numerical ranges for operating parameters such as voltage, power flow, and nodal loads; c) Non-quantitative prior description module Non-quantifiable prior knowledge represented in natural language, such as the fluctuation pattern of photovoltaic power output under cloud cover and the transition characteristics of peak load during holidays.

[0057] Through the above hybrid prompt word design, the samples generated by the large language model have statistical rationality and meet the requirements of physical interpretability and scenario specificity.

[0058] S302: Verify the distribution consistency between the initial expanded sample set and the reference base. If the verification fails, adjust the structured prompt words until the samples generated by the large language model meet the distribution consistency requirements.

[0059] Among them, distribution consistency verification refers to the process of measuring the degree of distribution deviation between the initial expanded sample set and the reference basis by using the maximum mean difference index, and judging whether the generated samples conform to the statistical characteristics of the target scene; preset threshold It refers to pre-set The upper limit is used to determine whether the distribution consistency meets the standard.

[0060] For example, the specific process of distribution consistency verification is as follows: calculate the initial expanded sample set. With reference substrate Between Value, determine whether it satisfies: (7) If the calculation yields If the value exceeds the preset threshold, the distribution consistency check is deemed to have failed. In this case, the structured prompt words are adaptively adjusted according to the offset direction. The constraint parameters and scenario descriptions in the code, for example, if the photovoltaic output fluctuation of the generated sample is too small, then the prompt words will be strengthened. The feature description of the sharp drop in output caused by mid-cloud occlusion is then used to regenerate samples and validate them again, forming an iterative feedback loop until the generated samples are... The value is below the threshold.

[0061] S303: Merge the validated initial augmented sample set with the filtered dataset to obtain the enhanced training set.

[0062] The enhanced training set refers to the complete training dataset formed by merging the validated generated sample set with the filtered dataset, which contains both real historical samples and synthetic unconventional working condition samples.

[0063] It should be noted that after completing the distribution consistency verification and physical validity verification, the valid augmented samples generated by the large language model will be... Compared with the real samples obtained through screening After merging and removing duplicate and abnormal samples, the final enhanced training set is formed. This training set contains samples from real operating scenarios and is supplemented with a large number of synthetic samples of rare and unconventional operating conditions, which greatly expands the coverage of key operating conditions.

[0064] This embodiment uses structured prompt words to drive a large language model to generate samples and combines iterative verification of distribution consistency to form a closed-loop optimization, which breaks through the applicability limitations of traditional data augmentation methods in power system scenarios. The generated samples have both physical rationality and scenario relevance. The augmented training set formed by merging the filtered datasets significantly expands the coverage of rare critical operating conditions.

[0065] Building upon prior knowledge-guided sample augmentation, in an exemplary embodiment, an LLM (Language Model) is integrated as a knowledge controller into the reinforcement learning framework. This constructs a control framework that combines a large language model with reinforcement learning. Leveraging its large-scale knowledge base and semantic reasoning capabilities, it provides high-quality action suggestions to the policy network in the early stages of training, constructing a hybrid optimization mechanism that combines knowledge-driven and experience-driven approaches. Hybrid reinforcement learning training is then performed based on this control framework. The structure of the control framework integrating a large language model and reinforcement learning is as follows: Figure 2 As shown, the training process includes: S401: Integrate the action suggestions generated by the large language model in the power grid operation state reasoning, and the corresponding environmental interaction results to obtain the first state-action-next state-reward quadruple, and store it in the prior knowledge experience pool.

[0066] Among them, the prior knowledge experience pool This refers to a replay buffer used to store prior guidance experience generated by a large language model, belonging to the first type of experience storage unit in the fusion control framework; the state-action-next state-reward quadruple. It is the standard storage format for experiential data in reinforcement learning, in which s Given the current environmental state, a For the action to be performed, s' The next environmental state after the action is performed. r The reward obtained for performing an action.

[0067] For example, during training, the current power grid operating state is first input into the large language model. The large language model combines internal domain knowledge and control rules to infer and output corresponding action suggestions. After applying this action to the power grid environment, the next operating state and corresponding reward value are obtained. These elements are then integrated into... All formatted quadruples are stored in the prior knowledge pool. In this experience pool, the data represents control experience guided by prior knowledge, which can be used to provide high-quality initial action guidance for policy networks.

[0068] S402: Based on the operational data generated by the interaction between the multi-agent decision-making architecture and the actual power grid environment, the second state-action-next state-reward quadruple is obtained and stored in the interaction experience pool.

[0069] Among them, the interactive experience pool This refers to a replay buffer used to store real-world experience data generated by the interaction between MAPPO (Multi-Agent Proximal Policy Optimization) policies and the environment. It belongs to the second type of experience storage unit in the fusion control framework. MAPPO policies are a mainstream reinforcement learning algorithm for multi-agent collaborative decision-making scenarios. Its core architecture adopts a centralized training and distributed execution paradigm. During the training phase, policy optimization can be completed with the help of global system information, and during the execution phase, each agent can autonomously output decisions based only on its own local observations.

[0070] For example, during training, the multi-agent decision-making architecture receives the state of the power grid environment, outputs control actions from the current policy network, and generates the next state and reward value after the actions are applied to the environment. The corresponding data is then organized into... The formatted quadruple is then stored in the interactive experience pool. The data in this experience pool reflects the learning dynamics during the interaction between the strategy and the real environment, and is the core data source for the iterative optimization of reinforcement learning strategies.

[0071] It should be noted that interactive experience comes from the direct interaction between the strategy and the real environment, and can truly reflect the dynamic response characteristics of the environment. It is the core source of experience to ensure that the strategy is ultimately adapted to the actual running scenario. It is maintained in parallel with the prior knowledge experience pool, which can realize the independent management and flexible sampling of the two types of experience.

[0072] S403: Sample batches of experience data from the prior knowledge experience pool and the interaction experience experience pool respectively, and calculate the loss to obtain the total policy loss.

[0073] The total policy loss refers to the comprehensive loss value formed by integrating the loss terms corresponding to prior guidance experience and the loss terms corresponding to environmental interaction experience, which is used to drive the parameter update of the policy network.

[0074] It should be noted that during training iterations, batches of experience data are randomly extracted from the prior knowledge experience pool and the interaction experience experience pool according to the set sampling ratio. The corresponding loss terms are calculated based on the two types of batch data, and then the total policy loss is obtained by weighted fusion, which serves as the optimization target for updating the policy network parameters.

[0075] In practice, the sampling ratio of the two experience pools can be dynamically adjusted according to the training stage. In the early stage of training, the sampling ratio of the prior knowledge experience pool is increased to accelerate convergence, and in the later stage of training, the sampling ratio of the interactive experience pool is increased to improve environmental adaptability.

[0076] S404: With minimizing the total policy loss as the optimization objective, iteratively update the network parameters of the multi-agent decision architecture to complete a single round of hybrid reinforcement learning training iteration.

[0077] Among them, network parameter iterative update refers to the process of adjusting the weight parameters of the neural network along the direction of gradient descent of the loss function based on the backpropagation algorithm in order to reduce the loss value and improve the policy performance.

[0078] It should be noted that, based on the calculated total policy loss, the gradients of the parameters of each layer of the policy network are calculated using the gradient backpropagation algorithm, and the network parameters are updated using the gradient descent optimizer to complete a single round of training iteration; the process of experience collection, loss calculation and parameter update is repeated until the model converges, and finally a multi-objective coordinated control model adapted to unconventional working conditions is obtained.

[0079] Understandably, by integrating two types of experience into a hybrid training mode, the policy network can quickly establish basic control capabilities with the help of prior knowledge, and continuously optimize control accuracy and environmental adaptability through environmental interaction. Compared with policies trained on a single experience, it has stronger robustness and convergence efficiency.

[0080] This embodiment constructs a hybrid reinforcement learning training framework with dual experience pools, organically integrating the prior guidance experience of the large language model with the environmental interaction experience of reinforcement learning. This forms a dual-path optimization mechanism that combines knowledge-driven and experience-driven approaches, effectively improving the convergence speed and action quality in the early stages of model training, and enhancing the robustness and generalization ability of the final control strategy under unconventional conditions.

[0081] In an exemplary embodiment, batches of experience data are sampled from the prior knowledge experience pool and the interaction experience experience pool, respectively, and loss calculations are performed to obtain the total policy loss. The network parameters of the multi-agent decision-making architecture are iteratively updated with the goal of minimizing the total policy loss, including: S501: Calculate the imitation learning loss based on the batch of experience data sampled from the prior knowledge experience pool.

[0082] Among them, imitation learning loss refers to the loss term that measures the distribution difference between the output action of the current reinforcement learning policy and the suggested action generated by the large language model, and is used to guide the policy network to learn control experience from prior knowledge.

[0083] For example, since MAPPO is implemented based on a stochastic policy framework and its output is a probability distribution in the action space, large language models need to synchronously output alignable probabilistic results to achieve loss calculation. Based on this requirement, the imitation learning loss is defined as follows: (8) In the formula, To mimic learning loss, it is used to measure the distributional difference between the current policy output action and the suggested action generated by the large language model; These are the parameters of the policy network, and are the core objects that need to be updated during the reinforcement learning training process; This refers to the training batch size; The policy distribution output by the large language model controller. For the current reinforcement learning policy network in state The probability distribution of the output action; For the first n The global state of the system corresponding to each sample is formed by the aggregation of local observation information from all agents. Indicates action a It is sampled from the policy distribution output by the large language model.

[0084] S502: Based on the batch experience data sampled from the interaction experience pool, calculate the multi-agent reinforcement learning policy loss and value function loss respectively.

[0085] Among them, the multi-agent reinforcement learning policy loss refers to the policy loss of the MAPPO algorithm, which is used to measure the optimization objective of policy updates; the value loss refers to the value function loss of the evaluation network, which is used to measure the accuracy of the value function in estimating the cumulative reward.

[0086] For example, after sampling batch data from the interactive experience pool, the MAPPO policy loss is first calculated, and the expression is: (9) In the formula, This is the policy loss of the MAPPO algorithm, used to measure the optimization objective of policy updates. During training, policy performance is improved by minimizing this loss. These are the parameters of the policy network, and are the core objects that need to be updated during the reinforcement learning training process; This refers to the training batch size; This represents the probability ratio between the current strategy and the old strategy. The dominant function; This is a clipping function used to limit the probability ratio to a certain value. Within the range; This is a pruning parameter used to limit the step size of policy updates.

[0087] S503: The total policy loss is obtained by weighting and fusing the imitation learning loss and the multi-agent reinforcement learning policy loss by balancing hyperparameters.

[0088] Among them, the total policy loss refers to the comprehensive policy loss after weighted fusion, which is used to simultaneously drive the policy network to learn prior experience and environmental interaction experience; the balancing hyperparameter refers to the adjustable parameter used to adjust the weight ratio of the two types of loss terms.

[0089] For example, the expression for calculating the total policy loss is as follows: (10) In the formula, The total policy loss is used to drive the model to complete the reinforcement learning process based on real-world interactions. To mimic loss, it is used to measure the distributional difference between the current policy output action and the suggested action generated by the large language model; Let be the balancing hyperparameter for the two types of loss terms. When When a larger value is taken, the strategy tends to learn from the prior experience of the large language model. When taking smaller values, the strategy focuses more on learning and optimizing from environmental interactions.

[0090] S504: Update the decision network parameters of the multi-agent decision architecture with the optimization objective of minimizing the total policy loss, and update the evaluation network parameters of the multi-agent decision architecture with the optimization objective of minimizing the value function loss.

[0091] In this context, the decision network (actor) refers to the network used to output action policies in reinforcement learning, corresponding to the Actor part in the Actor-Critic architecture; the evaluation network (critic) refers to the network used to evaluate state value and estimate the output value function, corresponding to the Critic part in the Actor-Critic architecture.

[0092] It should be noted that the Actor-Critic architecture is a classic architecture for deep reinforcement learning. The Actor is responsible for making decisions, while the Critic is responsible for evaluating and providing feedback. The two work together iteratively to achieve continuous optimization of the policy, balancing training stability and convergence efficiency.

[0093] like Figure 2As shown, this control framework employs a joint training mechanism driven by dual experience pools: The LLM outputs an initial policy to provide initialization guidance for the MAPPO policy network, and simultaneously stores the control experience generated by inference into a replay buffer. The experience pool is used to calculate the imitation learning loss; the real-world operational experience generated by MAPPO's interaction with the environment is stored in the replay buffer. The experience pool is used to calculate the MAPPO policy loss and the value function loss respectively; the imitation learning loss and the MAPPO loss are fused into the total loss to update the actor decision network, and the value loss is used to update the critic evaluation network separately; finally, the final multi-agent coordinated control policy adapted to the scenario is obtained by weighted fusion of the initial LLM policy and the iteratively optimized MAPPO policy.

[0094] This embodiment designs a joint loss function that integrates imitation learning loss and MAPPO policy loss, clarifies the network parameter update rules driven by dual experience pools, and realizes the quantitative fusion of prior guidance experience and environmental interaction experience. It provides a feasible optimization objective and parameter update mechanism for hybrid reinforcement learning training, and ensures the training stability and convergence performance of the control model.

[0095] In an exemplary embodiment, considering the interrelationships and collaborative operation requirements of distributed photovoltaic (PV) systems in multiple areas, the collaborative control problem is modeled as a decentralized partially observable Markov Decision Process (Dec-POMDP). The framework matches the interrelationships of multiple PV entities through a distributed information interaction mechanism, balancing the need to suppress power quality disturbances with the need to reduce network losses in the optimization objective. Specifically, the pre-construction steps of the multi-agent decision-making architecture include: S601: Abstract various adjustable resources in the distribution network of multiple distribution areas into independent intelligent agents, forming a multi-agent set.

[0096] Among them, an intelligent agent refers to a decision-making entity with autonomous perception, decision-making, and execution capabilities; a collection of multiple intelligent agents It refers to the set of decision-making entities composed of intelligent agents corresponding to all adjustable resources.

[0097] For example, each type of energy resource is modeled as an independent intelligent agent with autonomous decision-making capabilities. Considering the property rights and interest boundaries of distributed photovoltaic power generation in different areas, the set of intelligent agents is divided according to equipment type as follows: (11) In the formula, This is a collection of distributed photovoltaic units from all areas of the entire network. A collection of energy storage systems, A collection of wind turbine generator sets. This is a collection of static var compensators (SVCs). Each photovoltaic intelligent entity can achieve collaborative decision-making through neighborhood information exchange.

[0098] S602: Set a local observation range for each agent to obtain the local observation space of each agent.

[0099] The local observation space refers to the set of all operational state information that each independent agent can obtain, and it is the only information input for the agent to make autonomous decisions.

[0100] For example, to meet the architectural requirements of distributed control and simultaneously match the information interaction boundaries under the interrelationship of interests among multiple stakeholders, the first... i Each intelligent agent (photovoltaic, energy storage, wind turbine, or static var compensator SVC) can only observe the operating status of itself and adjacent devices. This avoids the privacy and cost issues associated with global communication while ensuring collaborative information sharing among related entities. The observation vector is defined as follows: (12) In the formula: For the first i The local observation vector of an agent is the only information input for the agent to make autonomous decisions. For the first j The operational state feature vector of an agent contains the core operational parameters of that agent; , , The first i Active power, reactive power, and node voltage of each intelligent agent; In order to be with the first i A set of neighboring intelligent agents that are topologically connected and have collaborative interests.

[0101] S603: Combine the local observation spaces of each agent to construct the global state space of the system.

[0102] The global state space refers to the collection of all operational state information of the system, which is composed of the aggregation of local observation information of all agents, and provides a globally consistent state benchmark for multi-agent collaborative decision-making.

[0103] For example, the global state of the system can be represented as: (13) In the formula: For the global state space; For the first i Local observation vectors of an agent.

[0104] S604: Aggregates the range of decision variables of all agents to form a global distributed action space.

[0105] The action space refers to the set of values ​​for all control actions that can be executed by all agents; the global action set is composed of the control variables of all agents.

[0106] For example, the control commands of each agent are continuous variables, and the global action set consists of the control variables of all agents. Each photovoltaic entity can independently output control actions based on its own operating goals and neighborhood information, achieving distributed collaborative regulation. The global action space can be represented as: (14) In the formula, It is a global set of actions for all intelligent agents in the entire network, including control commands for all adjustable resources in the distribution network of multiple distribution areas; For the first i The reactive power reference setpoint of each agent is the core control variable in this scenario. The agent tracks this reference value to achieve dynamic adjustment of reactive power output, thereby supporting node voltage stability and reducing network losses.

[0107] S605: Based on the operation optimization objectives of multi-regional distribution networks, a multi-objective reward function is set.

[0108] Among them, the reward function refers to the quantitative indicator used in reinforcement learning to evaluate the effect of action execution. The agent optimizes the policy by maximizing the cumulative reward; the multi-objective reward function refers to the reward calculation function that takes into account multiple operational optimization objectives at the same time.

[0109] For example, the reward function is designed around the core optimization objective of distribution network operation, taking into account multiple objectives such as power quality disturbance suppression and network loss reduction, while implicitly including the synergistic benefit orientation of multiple photovoltaic entities, and is defined as: (15) In the formula: The reward is the immediate reward obtained by the agent after performing a control action. During the reinforcement learning training process, the strategy is optimized by maximizing the cumulative reward. The higher the reward value, the better the overall performance of the current control action. This represents the total active power network loss of the system at the current moment, expressed in MW. It reflects the economic level of the distribution network operation; the higher the loss, the higher the operating cost. The negative sign indicates that the lower the network loss, the higher the reward value, guiding the strategy to reduce the active power loss of the entire network. This is a voltage barrier function used to assess the degree of deviation of node voltage from its rated value; generally, the greater the deviation of the voltage from the rated value, the higher the output value of this function, and the stronger the penalty for power quality disturbances.

[0110] S606: Based on a multi-agent set, local observation space, system global state space, global distributed action space, and multi-objective reward function, a decentralized partially observable multi-agent decision-making architecture is constructed.

[0111] Among them, the Decentralized Partially Observable Markov Decision Process (Dec-POMDP) ​​refers to a Markov decision model applicable to multi-agent distributed decision-making scenarios, in which each agent makes autonomous decisions based only on local observation information and achieves the global optimization goal through collaboration.

[0112] It should be noted that by integrating the above-defined set of agents, observation space, global state space, action space, and reward function, a multi-agent decision-making architecture based on Dec-POMDP is constructed. This architecture matches the distributed operation characteristics and multi-agent interest boundaries of multi-regional distribution networks and can support the implementation of hybrid reinforcement learning training.

[0113] This embodiment uses Dec-POMDP modeling to model the distributed photovoltaic collaborative control problem in multiple distribution areas. It clearly defines five core elements: agent set, observation space, state space, action space, and multi-objective reward function. It constructs a multi-agent decision-making architecture adapted to the characteristics of distributed operation, providing a standardized decision model foundation for hybrid reinforcement learning training and ensuring the adaptability of the control strategy to the actual distribution network operation scenario.

[0114] The implementation of the present invention will be further described below with reference to a specific example.

[0115] (1) Example system and parameter settings: This example uses an improved IEEE 33-node multi-distribution area distribution system as the simulation test scenario. The system topology includes four distribution areas, configured with four diesel distributed power sources with rated capacities of 0.8~1.4MW, a total distributed photovoltaic (PV) unit with a capacity of 1.8MW, two wind turbines with rated capacities of 0.5MW, a 2MW / 4MWh shared energy storage system, and four 0.3MVar static var compensators. The distributed PV penetration rate reaches 35%, making it suitable for distribution network operation scenarios with a high proportion of distributed power sources. The improved IEEE 33-node multi-distribution area distribution system topology is as follows: Figure 3 As shown.

[0116] The core training parameters of reinforcement learning are fused with the hyperparameters of the large language model and dynamically configured according to the phased requirements of collaborative control: discount factor. Set to 0.99 to balance the immediate rewards and long-term benefits of the control strategy, adapting to the timing optimization needs of 24-hour multi-period coordinated control; the experience replay buffer capacity is set to... It can store interaction samples for at least 1000 complete control cycles, supporting the data requirements for large-batch training; the training batch size is set to 256, and a single iteration samples state-action samples from different operating scenarios in multiple zones, improving the scenario generalization ability of the policy; the optimizer is AdamW, and the initial learning rate is set to... During training, a cosine annealing strategy is used to dynamically decay the learning rate, ensuring rapid convergence in the early stage and improving the accuracy of strategy fine-tuning in the later stage.

[0117] (2) Analysis of model training and control results: The training curve of the MAPPO-LLM model is as follows: Figure 4 As shown, in the early stages of training (approximately the first 100 steps), the reward value increases rapidly, indicating that the policy can quickly learn effective control behavior. In the middle stage (approximately 100–300 steps), the reward fluctuates to some extent and rises slowly, reflecting the dynamic balance between exploration and exploitation in reinforcement learning. As training continues (after approximately 300 steps), the reward gradually stabilizes and reaches convergence at approximately 400–500 steps, demonstrating that the proposed MAPPO-LLM method can obtain stable and high-quality control policies within a limited number of training steps. Furthermore, the smoothed reward curve further verifies the overall convergence trend of the training process, indicating that the method has good convergence and stability.

[0118] To verify the effectiveness of the proposed method, this example compares and analyzes droop control, traditional OPF control, MAPPO control, and the proposed MAPPO-LLM control. Droop control, as a basic local control method, primarily relies on local measurement information to achieve rapid response. Traditional OPF control optimizes power flow distribution based on system operating constraints and can serve as a benchmark for conventional optimization scheduling methods. MAPPO control utilizes multi-agent reinforcement learning to achieve collaborative decision-making for resources across multiple transformer areas. The proposed MAPPO-LLM control, based on the MAPPO framework, introduces a large language model-assisted strategy understanding and control enhancement mechanism to improve global coordination capabilities and operational economy in complex operating scenarios.

[0119] Typical daily network loss curves under the four methods are as follows: Figure 5As shown in the typical daily network loss curves, the network loss levels of the four methods, from high to low, are generally as follows: droop control, traditional OPF control, MAPPO control, and the proposed MAPPO-LLM control. Droop control mainly relies on local response and lacks global coordination capabilities, thus resulting in the highest network loss during peak load periods. Traditional OPF control can optimize power flow distribution to some extent and reduces network loss compared to droop control. MAPPO control further improves system operation, but it still does not reach the optimal level overall. In contrast, the proposed MAPPO-LLM control maintains the lowest network loss for most of the day, especially effectively suppressing peak network loss during heavy load periods such as the evening peak. This indicates that the method can more fully coordinate distributed resources and network operating status, thereby improving the economic efficiency of distribution network operation.

[0120] Typical daily node average voltage deviation rate curves under the four methods are as follows: Figure 6 As shown in the figure, the average voltage deviation rate curves of typical daily nodes reveal a clear hierarchy of differences in voltage regulation performance among the four control methods. Overall, droop control exhibits the highest voltage deviation rate, with a significant increase during the evening peak period (approximately 18–21 hours), indicating that relying solely on local response is insufficient to effectively support the terminal voltage. Traditional OPF control reduces voltage deviation to some extent by optimizing power flow distribution, but its steady-state optimization capability is limited, and it still exhibits some voltage offset during periods of rapid load increase. MAPPO control further improves voltage regulation, with a smoother overall curve, indicating better dynamic response capabilities in multi-region coordinated regulation. In contrast, the proposed MAPPO-LLM control maintains the lowest voltage deviation rate throughout the entire period, especially effectively suppressing the expansion of voltage deviation during midday distributed power output fluctuations and evening peak load concentration periods, demonstrating its stronger ability to coordinate active and reactive power resources under complex operating scenarios. Meanwhile, compared with the network loss curve, the voltage deviation rate changes more smoothly and shows a local rise in some periods (such as when the photovoltaic output is high at noon). This indicates that the voltage operating status is not only affected by the load level, but also closely related to the access of distributed power sources and reactive power support capacity. The overall trend is more in line with the actual operation law of the distribution network.

[0121] (3) The impact of prior knowledge and large language models on datasets: Figure 7The results comparing the MMD (Missing Data Meaning) index under different data construction methods are presented. It can be seen that there is a significant distributional shift between the original source dataset and the target unconventional working condition dataset, with an MMD value reaching 0.286. While random selection can slightly reduce the distributional difference, the improvement is limited. Using only MMD selection, the MMD value drops to 0.163, indicating that sample selection based on distributional distance can improve the matching degree between the samples and the target working condition. Building on this, introducing quantifiable priors and mixed priors further reduces the MMD value to 0.118 and 0.083, respectively. Using "mixed prior + LLM enhancement," the MMD value is the lowest, at only 0.052, indicating that the proposed method can effectively reduce the distributional shift between the source dataset and the unconventional working condition target dataset, providing a more representative sample basis for control strategy training.

[0122] Figure 8 The paper presents a comparison of physical constraint satisfaction rates under different data augmentation methods. The comprehensive physical feasibility rate is used to evaluate the engineering usability of samples generated by different data augmentation methods. The comprehensive physical feasibility rate is defined as the proportion of generated samples that simultaneously satisfy voltage constraints, branch power flow constraints, and upper and lower limits of photovoltaic output constraints out of the total number of samples. Unlike the single constraint satisfaction rate, this indicator requires samples to remain feasible under all major operational constraints, thus more comprehensively reflecting whether the generated data conforms to the actual operating rules of the distribution network. A higher indicator indicates a lower proportion of physically anomalous samples in the generated samples, and a more suitable dataset for training reinforcement learning control strategies. It can be seen that although methods such as random perturbation, SMOTE, and GAN can expand the sample size, due to the lack of sufficient embedding of power system operating constraints, a certain proportion of anomalous samples still exist in terms of voltage constraints, power flow constraints, and photovoltaic output constraints. Specifically, the comprehensive physical feasibility rate of the random perturbation method is only 82.3%, while SMOTE and GAN methods improve it to 86.8% and 90.0%, respectively. In comparison, the LLM generation method improves the overall physical feasibility rate to 95.5% after introducing quantifiable priors; after further integrating non-quantifiable priors, the rate reaches 97.8%, indicating that hybrid prior knowledge can effectively constrain the sample generation process, enabling the generated samples to simultaneously meet the requirements of statistical rationality and power system physical feasibility.

[0123] Figure 9The paper presents a comparison of the average coverage rates of unconventional operating conditions under different data construction methods. It can be seen that the original source dataset and the random selection method have weak coverage capabilities for unconventional operating scenarios, with average coverage rates of only 36.8% and 39.8%, respectively, which is insufficient to support robust training of control strategies under low-frequency, high-risk operating conditions. After MMD selection alone, the average coverage rate increases to 56.0%, indicating that distribution alignment can enhance the sample's representation ability of the target operating conditions. Further introducing quantifiable priors and hybrid priors increases the average coverage rates to 69.0% and 82.0%, respectively. Using a "hybrid prior + LLM enhancement" approach, the average coverage rate reaches 94.5%, demonstrating that the proposed method can effectively supplement key unconventional scenarios such as typhoons, cold waves, cloud cover, and cross-regional load mismatches that are missing in traditional datasets.

[0124] Figure 10 Coverage heatmaps for different data construction methods under various key unconventional operating conditions are presented. It can be seen that the original source dataset suffers from insufficient coverage in all scenarios from S1 to S6 (covering typhoon conditions, rapid cloud cover conditions, abnormal load changes during holidays, cross-regional load peak-valley mismatch, high photovoltaic power generation coupled with local low load, and peak-hour load surge coupled with low SOC of energy storage). Coverage is particularly low in low-frequency, complex scenarios such as low and fluctuating photovoltaic output during typhoons and peak-hour load surge coupled with low SOC of energy storage. With the gradual introduction of MMD screening, quantifiable priors, hybrid priors, and LLM enhancement mechanisms, the coverage of all scenarios shows a continuous upward trend. In particular, after adopting the "hybrid prior + LLM enhancement" method, the coverage of scenarios from S1 to S6 all exceed 90%, indicating that this method not only improves the overall average coverage level but also provides a more balanced sample supplement for different types of unconventional operating conditions, thereby enhancing the training dataset's ability to fully describe complex operating states.

[0125] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0126] Based on the same inventive concept, this application also provides a multi-area distributed photovoltaic (PV) multi-objective coordinated control device for implementing the multi-area distributed PV multi-objective coordinated control method described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in the embodiments of the multi-area distributed PV multi-objective coordinated control device provided below can be found in the limitations of the multi-area distributed PV multi-objective coordinated control method described above, and will not be repeated here.

[0127] Please see Figure 11 This invention provides a multi-area distributed photovoltaic multi-objective coordinated control device, comprising: The data filtering module is used to filter the source dataset based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, so as to obtain the filtered dataset covering the non-normal operation conditions. The data augmentation module is used to augment the dataset based on the filtered dataset and the target dataset to obtain an enhanced training set. The model building module is used to perform hybrid reinforcement learning training on a pre-built multi-agent decision-making architecture based on the enhanced training set, so as to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. The coordination control module is used to output control commands for adjustable resources in each transformer area using a multi-objective coordination control model.

[0128] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0129] Reference Figure 12This invention also provides a computer device, including: a memory and a processor, and a computer program stored in the memory. When the computer program is executed on the processor, it performs the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0130] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 12 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0131] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0132] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0133] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0134] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0135] This invention provides a computer program product, including a computer program that, when executed by a processor, performs the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained by hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. Hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate a multi-objective coordinated control model. Using a multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

[0136] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0137] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0138] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0139] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-objective coordinated control method for distributed photovoltaic systems in multiple areas, characterized in that, Includes the following steps: Based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, the source dataset is filtered to obtain a filtered dataset covering the non-normal operation conditions. Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set; Based on the enhanced training set, the pre-constructed multi-agent decision-making architecture is trained using hybrid reinforcement learning to obtain a multi-objective coordinated control model adapted to unconventional working conditions. The hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agents in order to generate the multi-objective coordinated control model. Using the aforementioned multi-objective coordinated control model, control commands for adjustable resources in each transformer area are output.

2. The multi-objective coordinated control method for distributed photovoltaic systems in multiple areas according to claim 1, characterized in that, Based on the source dataset corresponding to the normal operating scenario and the target dataset corresponding to the unconventional operating scenario, the source dataset is filtered to obtain a filtered dataset covering the unconventional operating conditions, including: A hybrid prior constraint system is constructed by combining quantifiable and non-quantifiable prior knowledge; the quantifiable prior knowledge is used to ensure the physical compliance of the selected samples, and the non-quantifiable prior knowledge is used to guide the selected samples to match the distribution characteristics of unconventional operating conditions. Using the hybrid prior constraint system as the screening criterion, the source dataset is subjected to distribution alignment screening to obtain a filtered dataset that matches the distribution characteristics of the target dataset and satisfies the physical operation rules.

3. The multi-objective coordinated control method for distributed photovoltaic systems in multiple areas according to claim 1, characterized in that, Based on the filtered dataset and the target dataset, the dataset is augmented to obtain an enhanced training set, including: Using the union of the filtered dataset and the target dataset as a reference basis, an initial expanded sample set is generated using a large language model driven by structured cue words; Verify the distribution consistency between the initial expanded sample set and the reference base. If the verification fails, adjust the structured prompt words until the samples generated by the large language model meet the distribution consistency requirements. The validated initial augmented sample set is merged with the filtered dataset to obtain the enhanced training set.

4. The multi-objective coordinated control method for distributed photovoltaic systems in multiple areas according to claim 1, characterized in that, The hybrid reinforcement learning training is based on a control framework that integrates a large language model and reinforcement learning; the training process includes: The action suggestions generated by combining the large language model with the power grid operation state reasoning and the corresponding environmental interaction results are integrated to obtain the first state-action-next state-reward quadruple, which is then stored in the prior knowledge experience pool. Based on the operational data generated by the interaction between the multi-agent decision-making architecture and the actual power grid environment, a second state-action-next state-reward quadruple is obtained and stored in the interaction experience pool. Batch experience data are sampled from the prior knowledge experience pool and the interaction experience experience pool respectively, and loss is calculated to obtain the total policy loss; With minimizing the total policy loss as the optimization objective, the network parameters of the multi-agent decision architecture are iteratively updated to complete a single round of hybrid reinforcement learning training iteration.

5. The multi-objective coordinated control method for distributed photovoltaic systems in multiple areas according to claim 4, characterized in that, Batch experience data are sampled from the prior knowledge experience pool and the interaction experience experience pool respectively, and loss is calculated to obtain the total policy loss; With minimizing the total policy loss as the optimization objective, the network parameters of the multi-agent decision-making architecture are iteratively updated, including: Based on the batch of experience data sampled from the prior knowledge experience pool, the imitation learning loss is calculated; Based on the batch experience data sampled from the interaction experience pool, the multi-agent reinforcement learning policy loss and value function loss are calculated respectively. The total policy loss is obtained by weighting and fusing the imitation learning loss and the multi-agent reinforcement learning policy loss by balancing hyperparameters. The decision network parameters of the multi-agent decision architecture are updated with the goal of minimizing the total policy loss, and the evaluation network parameters of the multi-agent decision architecture are updated with the goal of minimizing the value function loss.

6. The multi-objective coordinated control method for distributed photovoltaic systems in multiple areas according to claim 1, characterized in that, The pre-construction steps of the multi-agent decision-making architecture include: Various adjustable resources within a multi-area distribution network are abstracted into independent intelligent agents, forming a multi-agent set; By setting a local observation range for each agent, the local observation space of each agent is obtained. By combining the local observation spaces of each intelligent agent, a global state space of the system is constructed; Aggregate the ranges of decision variables from all agents to form a globally distributed action space; Based on the operation optimization objectives of multi-regional distribution networks, a multi-objective reward function is set; Based on the multi-agent set, the local observation space, the system global state space, the global distributed action space, and the multi-objective reward function, a decentralized partially observable multi-agent decision-making architecture is constructed.

7. A multi-area distributed photovoltaic multi-objective coordinated control device, characterized in that, include: The data filtering module is used to filter the source dataset based on the source dataset corresponding to the normal operation scenario and the target dataset corresponding to the non-normal operation scenario, so as to obtain a filtered dataset covering the non-normal operation conditions. The data augmentation module is used to augment the dataset based on the filtered dataset and the target dataset to obtain an augmented training set. The model building module is used to perform hybrid reinforcement learning training on the pre-built multi-agent decision-making architecture based on the enhanced training set to obtain a multi-objective coordinated control model adapted to unconventional working conditions. The hybrid reinforcement learning training refers to the training process of integrating prior guidance experience and environmental interaction experience to iteratively update the network parameters of the multi-agent to generate the multi-objective coordinated control model. The coordination control module is used to output control commands for adjustable resources in each transformer area using the multi-objective coordination control model.

8. A computer device, characterized in that, The device includes a processor and a memory: The memory is used to store computer programs and send the instructions of the computer programs to the processor; The processor executes a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas as described in any one of claims 1-6, according to the instructions of the computer program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements a multi-objective coordinated control method for distributed photovoltaic systems in multiple areas as described in any one of claims 1-6.