A large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions

Abstract

This invention discloses a large-model-driven multi-objective photovoltaic-storage distribution network control method under extreme weather conditions, belonging to the field of distribution network control technology. The method includes: collecting real-time electrical data and extreme weather data from the distribution network; identifying topology change events and generating structured natural language scene descriptions; generating a dynamic reward function with four objectives—voltage stability, frequency recovery, etc.—adapted to the scenario using a lightweight large-language model fine-tuned with power sector corpora; extracting global high-dimensional features using a topology and feature dual-attention graph convolutional network; using a vector concatenated from features and reward weights as state input, generating multi-device collaborative control instructions through a near-end policy optimization algorithm, and outputting the instructions after verification. This invention achieves dynamic perception of the distribution network topology, dynamic adaptation of control objectives, and collaborative optimization control of multiple devices under extreme weather conditions, comprehensively improving the operational resilience and renewable energy utilization efficiency of high-proportion renewable energy distribution networks under extreme weather conditions.

Description

Technical Field

[0001] This invention relates to the field of power distribution network control technology, and more specifically, to a large-model-driven method for multi-objective photovoltaic-storage-power distribution network regulation under extreme weather conditions. Background Technology

[0002] The penetration rate of distributed energy, with distributed photovoltaic (PV) and electrochemical energy storage at its core, in power distribution networks is increasing year by year. PV-energy storage distribution networks have become a core form for achieving energy structure transformation and improving power supply capacity and energy utilization efficiency in the construction of new power systems. At the same time, global climate change has led to frequent extreme weather events such as heat waves, cold waves, and strong storms. The multi-dimensional and highly random disturbances they bring can cause drastic fluctuations in PV output, sharp increases in load, and abnormal line parameters, easily leading to safety hazards such as voltage and frequency exceeding limits and equipment overload. This poses a severe challenge to the safe and stable operation of PV-energy storage distribution networks with a high proportion of new energy integration.

[0003] For the operation and control of photovoltaic-storage-distribution networks, existing technologies mostly employ intelligent control methods based on reinforcement learning. While these methods can achieve autonomous optimization and control of multiple devices in the distribution network, they suffer from a core inherent flaw directly corresponding to the inventiveness of this invention in extreme weather scenarios: rigid reward function design. Existing reinforcement learning control schemes all rely on manually preset fixed-weight reward functions, which cannot adapt to the real-time changing priority requirements of the four core control objectives—voltage stability, frequency recovery, sustainable energy storage, and photovoltaic absorption—under extreme weather conditions. For example, in storm fault scenarios, priority should be given to ensuring voltage safety and frequency stability; during peak photovoltaic power generation periods, priority should be given to improving the renewable energy absorption rate and reducing curtailment. Fixed-weight reward functions cannot achieve dynamic adaptation of multiple objective priorities, easily leading to multi-objective control conflicts, control lags, or even failures in extreme scenarios. This has become a core technical bottleneck restricting the safe and stable operation of photovoltaic-storage-distribution networks and the efficient utilization of renewable energy under extreme weather conditions. Therefore, there is an urgent need to provide a large-scale model-driven multi-objective photovoltaic-storage-distribution network control method under extreme weather conditions to solve the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions. This method enables dynamic perception of the distribution network topology, dynamic adaptation of regulation targets, and collaborative optimization regulation of multiple devices under extreme weather conditions, thereby comprehensively improving the operational resilience and renewable energy utilization efficiency of high-proportion renewable energy distribution networks under extreme weather conditions.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] This invention provides a large-model-driven method for multi-objective photovoltaic-storage-distribution network regulation under extreme weather conditions, comprising the following steps:

[0007] S1. Collect real-time electrical data and extreme weather data of the power distribution network, identify topology change events based on voltage and current mutations, dynamically update the adjacency matrix and calculate the topology change confidence, and generate a structured natural language scene description.

[0008] S2. Input the structured natural language scene description into a lightweight large language model fine-tuned by power field corpus. Based on preset prompt templates, output format constraints and parsing rules, generate a dynamic multi-objective reward function containing four objectives: voltage stability, frequency recovery, sustainable energy storage and photovoltaic consumption. Extract weight parameters and scoring rules.

[0009] S3. Construct a node feature matrix that integrates distribution network electrical features, extreme weather features, and topological association features. Based on the adjacency matrix, use a topological and feature dual attention graph convolutional network to extract features, output a global high-dimensional feature vector, and verify the accuracy of the features by reconstructing the error.

[0010] S4. Set the input state of the policy network of the near-end policy optimization algorithm to the concatenation vector of the global high-dimensional feature vector and the weight parameters, set the action space to the combination of energy storage charging and discharging power, photovoltaic flexible curtailment power and transferable load adjustment power, set the reward function to the dynamic multi-objective reward function, and generate photovoltaic, energy storage and controllable load coordinated control instructions; perform multi-level verification of the control instructions on energy balance constraints and equipment rated constraints, and output standardized control instructions after the verification is passed;

[0011] S5. Train the graph neural network and the near-end policy optimization network based on the extreme weather dataset, fine-tune the lightweight large language model, and deploy the trained algorithm to the distribution network control device to achieve real-time closed-loop control.

[0012] As a preferred embodiment of the present invention, step S1 specifically includes:

[0013] S11. Data Acquisition: Collect real-time electrical operation data of the power distribution network, as well as raw meteorological observation data related to extreme weather, to provide basic input for subsequent modeling and sensing processing.

[0014] S12. Dynamic evolution modeling of extreme weather: For different types of extreme weather such as high temperature, cold wave, and storm, dynamic evolution models of meteorological elements corresponding to temperature, wind speed, and light intensity are constructed respectively to simulate the real-time fluctuation patterns of various meteorological elements under extreme weather conditions, and at the same time, the impact characteristics of extreme weather on power distribution network load and photovoltaic output are quantitatively characterized.

[0015] S13. Distribution network topology dynamic perception: Identify topology change events corresponding to line on / off, equipment switching, and switch actions by real-time mutations in electrical quantities of distribution network nodes. Update the distribution network dynamic adjacency matrix based on the identified topology change events. At the same time, perform confidence verification on the topology change events and determine whether to update the adjacency matrix based on the confidence verification results to avoid topology discrimination errors caused by false data.

[0016] S14. Multi-source data fusion and scene description generation: Extreme weather-related parameters, topology change events and corresponding verification results, and real-time operating status data of distribution network equipment are fused from multiple sources and transformed into standardized structured natural language scene descriptions, which serve as input prompts for subsequent large language models.

[0017] As a preferred embodiment of the present invention, step S2 specifically includes:

[0018] S21. Standardized input prompt construction: Design fixed prompt templates adapted to distribution network control scenarios, integrate current extreme weather types and levels, distribution network topology change events, real-time system operating status, and future weather forecast information, while clarifying the coverage target, output format, and industry standard constraints of the reward function;

[0019] S22. Dynamic reward function generation and structured parsing: Based on the input prompt template, a pre-trained generative large language model is used to generate a multi-objective dynamic reward function that is adapted to the current scenario. Then, natural language processing tools are used to extract the weight allocation logic and the segmented scoring rules of each control objective in the reward function and transform them into executable quantitative evaluation rules.

[0020] S23. Reward function rationality verification: The parsed reward function is compared and verified with the national and industry safety standards for distribution network operation. If the content does not meet the requirements of the standard, a secondary prompt for correction is triggered until the generated reward function fully complies with the power engineering safety standard.

[0021] As a preferred embodiment of the present invention, step S3 specifically includes:

[0022] S31. Construction of distribution network node feature matrix: Integrate the electrical operation characteristics of each node in the distribution network, the extreme weather characteristics of the area where the node is located, and the topological association characteristics of the node to construct a multi-dimensional node feature matrix adapted to graph neural network processing.

[0023] S32. Topology and Feature Dual Attention Graph Convolution Operation: A dual attention mechanism is introduced for the dynamic adjacency matrix of the distribution network that is updated in real time. The weights of feature correlation and topology connection strength between nodes are adaptively allocated, and the deep extraction and global fusion of node features are completed through multi-layer graph convolution operation.

[0024] S33. Feature Output and Consistency Verification: Pooling is performed on the hidden layer features output by the graph convolution to obtain a global feature vector that integrates dynamic topology information, extreme weather features, and node operating status. At the same time, the feature reconstruction error of the graph neural network is calculated, and a topology secondary verification is triggered based on a preset error threshold to avoid interference from erroneous topology information on subsequent control decisions.

[0025] As a preferred embodiment of the present invention, step S4 specifically includes:

[0026] S41. Construction of a reinforcement learning control decision framework: Based on the near-end policy optimization algorithm, a reinforcement learning decision framework adapted to the collaborative control of multiple devices in the photovoltaic-storage distribution network is constructed, clarifying the state space, action space, reward function, and policy update rules of the framework; the state space adopts the global feature vector output in step S3 and the weight parameters of the reward function generated in step S2, the reward function directly adopts the dynamic reward function generated in step S2, and the action space covers the adjustable commands of three types of control objects in the distribution network: energy storage, photovoltaic, and controllable load;

[0027] S42. Multi-device collaborative control decision generation: Based on the current system state, the Actor network of the reinforcement learning framework outputs multi-device collaborative control action commands for energy storage charging and discharging, photovoltaic flexible consumption, and controllable load time-series transfer within the distribution network. At the same time, the Critic network completes the state value assessment and completes the iterative optimization of the strategy based on preset rules.

[0028] S43. Safety constraint verification of control actions: Perform multi-level safety constraint verification on the output control action commands. The verification dimensions include at least the active power balance constraint of the distribution network, the rated operation constraint of each control device, and the transmission capacity constraint of the distribution network line, to ensure the engineering feasibility of the control actions and the safety of the distribution network operation.

[0029] As a preferred embodiment of the present invention, step S5 specifically includes:

[0030] S51. Construction of training and testing datasets: Based on the standard distribution network topology model, combined with historical meteorological datasets of various types of extreme weather and distribution network full-scenario operation datasets, an algorithm training and testing dataset covering different extreme weather, different topology states, and different operating conditions is constructed.

[0031] S52. Multi-network phased training and joint optimization: The large language model, graph neural network, and reinforcement learning policy network are trained in phases, and then the joint optimization of the multiple networks is completed. Among them, the large language model is supervised and fine-tuned using professional corpus in the power field, the graph neural network is trained with the ability to extract topological features, and the reinforcement learning network is trained iteratively with the control strategy and updated with parameters.

[0032] S53. Model Validation and Engineering Deployment: After training and optimization, the overall control model is tested and validated under multiple scenarios and operating conditions. Once the validation is successful, the trained control strategy is deployed in the distribution network field control device.

[0033] As a preferred embodiment of the present invention, the fixed prompt template in S21 also includes mandatory constraint clauses of the distribution network safety regulations. The mandatory constraint clauses include at least the legally qualified operating range of distribution network voltage and frequency, the minimum priority guarantee threshold of each control target, and the normalization constraint rules of the reward function weight. The secondary prompt correction in S23 adopts a targeted correction mode, which only makes local adjustments to the content that does not comply with the mandatory constraint clauses, without changing the core logic of the weight allocation of the reward function to adapt to the current extreme weather scenario.

[0034] As a preferred embodiment of the present invention, the pre-trained generative large language model in S22 is a lightweight large model that has been supervised and fine-tuned using professional corpus in the power field. The professional corpus includes national and industry standards for distribution network operation, historical cases of emergency control of distribution networks under extreme weather conditions, and technical specifications for photovoltaic and energy storage grid connection. The natural language processing tool has built-in fixed field extraction rules, which can directly convert the unstructured text output by the large language model into structured weight parameters and segmentation scoring rules that can be directly called by reinforcement learning algorithms without manual intervention.

[0035] As a preferred embodiment of the present invention, the dual attention mechanism in S32 includes two types: feature correlation attention and topology connection strength attention. Feature correlation attention is used to calculate the feature correlation degree between nodes, identify key nodes that are significantly affected by extreme weather and adaptively increase their feature weights. Topology connection strength attention is used to quantify the topology connection strength according to the real-time operating status of the line and adjust the connection weights between nodes accordingly.

[0036] As a preferred embodiment of the present invention, in step S33, when the feature reconstruction error exceeds the preset error threshold, a secondary verification of the distribution network topology data and a node feature outlier removal operation are triggered simultaneously to avoid erroneous data interfering with subsequent control decisions.

[0037] The beneficial technical effects of this invention are:

[0038] By employing a lightweight large language model fine-tuned from power sector corpora and combining it with fixed prompt templates that incorporate distribution network safety regulations, a multi-objective reward function is dynamically generated to adapt to real-time scenarios. This function adaptively adjusts the weight allocation of four objectives—voltage stability, frequency recovery, sustainable energy storage, and photovoltaic absorption—based on extreme weather types and distribution network operating conditions, completely resolving the pain point of traditional fixed-weight reward functions failing to adapt to dynamic changes in scenarios. Experimental verification shows that under extreme storm scenarios, the voltage deviation rate of this invention remains consistently below the 1% safety threshold, approximately 40% lower than existing LLM-PPO methods; frequency fluctuation amplitude is far below the 0.05Hz safety threshold, and the control accuracy is significantly superior to existing comparative algorithms, effectively avoiding distribution network voltage and frequency exceeding limits under extreme weather conditions and ensuring the safe and stable operation of the distribution network.

[0039] This invention introduces a dual attention mechanism of topology and features into graph neural networks, constructing a node feature matrix that integrates distribution network electrical features, extreme weather features, and topological correlation features. This matrix can respond in real time to dynamic topological changes caused by line faults and switching actions under extreme weather conditions, accurately identifying key nodes significantly affected by extreme weather and improving the accuracy of feature extraction. Simultaneously, through a feature reconstruction error verification mechanism, when the error exceeds a preset threshold, a secondary topology verification and outlier removal are triggered simultaneously, further ensuring the reliability of feature input. This design effectively solves the problems of feature extraction distortion and mismatch between control commands and actual operating conditions in traditional fixed topology models, providing accurate feature support for subsequent collaborative control decisions and indirectly improving the stability of the overall control effect.

[0040] Based on a near-end strategy optimization algorithm, a multi-device collaborative decision-making framework covering energy storage charging and discharging, flexible photovoltaic curtailment, and controllable load time-series transfer is constructed. Simultaneously, multi-level constraint verifications are set, including energy balance, equipment rating, and line transmission capacity, to ensure the engineering feasibility of control commands and the safety of equipment operation. Experimental results show that the photovoltaic absorption rate of the method of this invention is consistently above 95%, far exceeding existing comparative methods, effectively reducing the photovoltaic curtailment rate under extreme weather conditions; the energy storage SOC maintenance rate is consistently maintained above the safety threshold of 85%, demonstrating optimal stability and sustainability of energy storage control, achieving the dual goals of distribution network operation safety and efficient utilization of new energy.

[0041] By organically combining the dynamic rule generation capability of large language models, the precise feature extraction capability of graph neural networks, and the autonomous decision-making capability of reinforcement learning, the steps are closely connected and logically closed-loop. This approach not only leverages the advantages of large language models in scene understanding and dynamic rule generation, but also ensures the safety and controllability of the control process through standardized prompt templates, targeted correction, and multi-layer constraint verification. This promotes the upgrade of photovoltaic-storage distribution network control from fixed rule-driven to scene-adaptive intelligent-driven, and comprehensively enhances the operational resilience of high-proportion new energy distribution networks under extreme weather conditions.

[0042] Training and validation were completed using the IEEE 33-node system and the NREL2023 extreme storm dataset. The dataset covers different extreme weather conditions, topological states, and operating conditions. The trained model can be directly deployed to the distribution network field control device to achieve real-time closed-loop control. At the same time, the supervised fine-tuning of the large language model and the phased training and joint optimization of the graph neural network and reinforcement learning network give the model strong scene adaptability. It can be flexibly applied to various extreme weather scenarios such as high temperature, cold wave, and storm, without the need for separate debugging for a single operating condition. This significantly reduces the cost of engineering implementation and provides a reliable technical solution for extreme weather prevention and control in distribution networks with a high proportion of new energy access. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the present invention.

[0044] Figure 2 This is a graph showing the change in the reward function of this invention.

[0045] Figure 3 This is a comparison curve of voltage deviation rates for different control methods.

[0046] Figure 4 This is a comparison curve of frequency fluctuations using different control methods.

[0047] Figure 5 This is a comparison curve of photovoltaic absorption rates using different control methods.

[0048] Figure 6 This is a comparison curve of the SOC maintenance rate of energy storage using different control methods. Detailed Implementation

[0049] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0050] Combination Figure 1-6 The present invention provides the following embodiments:

[0051] Example 1:

[0052] A large-model-driven method for multi-objective photovoltaic-storage-distribution network regulation under extreme weather conditions includes the following steps:

[0053] S1. Collect real-time electrical data and extreme weather data of the power distribution network, identify topology change events based on voltage and current mutations, dynamically update the adjacency matrix and calculate the topology change confidence, and generate a structured natural language scene description.

[0054] S2. Input the structured natural language scene description into a lightweight large language model fine-tuned by power field corpus. Based on preset prompt templates, output format constraints and parsing rules, generate a dynamic multi-objective reward function containing four objectives: voltage stability, frequency recovery, sustainable energy storage and photovoltaic consumption. Extract weight parameters and scoring rules.

[0055] S3. Construct a node feature matrix that integrates distribution network electrical features, extreme weather features, and topological association features. Based on the adjacency matrix, use a topological and feature dual attention graph convolutional network to extract features, output a global high-dimensional feature vector, and verify the accuracy of the features by reconstructing the error.

[0056] S4. Set the input state of the policy network of the near-end policy optimization algorithm to the concatenation vector of the global high-dimensional feature vector and the weight parameters, set the action space to the combination of energy storage charging and discharging power, photovoltaic flexible curtailment power and transferable load adjustment power, set the reward function to the dynamic multi-objective reward function, and generate photovoltaic, energy storage and controllable load coordinated control instructions; perform multi-level verification of the control instructions on energy balance constraints and equipment rated constraints, and output standardized control instructions after the verification is passed;

[0057] S5. Train the graph neural network and the near-end policy optimization network based on the extreme weather dataset, fine-tune the lightweight large language model, and deploy the trained algorithm to the distribution network control device to achieve real-time closed-loop control.

[0058] As a preferred embodiment of the present invention, step S1 specifically includes:

[0059] S11. Data Acquisition: Collect real-time electrical operation data of the power distribution network, as well as raw meteorological observation data related to extreme weather, to provide basic input for subsequent modeling and sensing processing.

[0060] S12. Dynamic evolution modeling of extreme weather: For different types of extreme weather such as high temperature, cold wave, and storm, dynamic evolution models of meteorological elements corresponding to temperature, wind speed, and light intensity are constructed respectively to simulate the real-time fluctuation patterns of various meteorological elements under extreme weather conditions, and at the same time, the impact characteristics of extreme weather on power distribution network load and photovoltaic output are quantitatively characterized.

[0061] S13. Distribution network topology dynamic perception: Identify topology change events corresponding to line on / off, equipment switching, and switch actions by real-time mutations in electrical quantities of distribution network nodes. Update the distribution network dynamic adjacency matrix based on the identified topology change events. At the same time, perform confidence verification on the topology change events and determine whether to update the adjacency matrix based on the confidence verification results to avoid topology discrimination errors caused by false data.

[0062] S14. Multi-source data fusion and scene description generation: Extreme weather-related parameters, topology change events and corresponding verification results, and real-time operating status data of distribution network equipment are fused from multiple sources and transformed into standardized structured natural language scene descriptions, which serve as input prompts for subsequent large language models.

[0063] Through a progressive design involving data acquisition, modeling, topology sensing, and fusion generation, basic raw data is first acquired. Then, through dynamic evolution modeling, coarse-grained meteorological data is transformed into precise data strongly correlated with distribution network load and photovoltaic output. At the same time, topology sensing and confidence verification ensure the authenticity of topology information. Finally, multi-source data fusion transforms scattered meteorological, topology, and electrical data into standardized scene descriptions. This solves the technical pain points of messy raw data, susceptibility to interference in topology recognition, and inability to directly adapt scene information to large language models under extreme weather conditions. It provides comprehensive, accurate, and usable scene input for the subsequent generation of dynamic reward functions, laying the foundation for the entire control process.

[0064] Furthermore, the specific steps of S2 include:

[0065] S21. Standardized input prompt construction: Design fixed prompt templates adapted to distribution network control scenarios, integrate current extreme weather types and levels, distribution network topology change events, real-time system operating status, and future weather forecast information, while clarifying the coverage target, output format, and industry standard constraints of the reward function;

[0066] S22. Dynamic reward function generation and structured parsing: Based on the input prompt template, a pre-trained generative large language model is used to generate a multi-objective dynamic reward function that is adapted to the current scenario. Then, natural language processing tools are used to extract the weight allocation logic and the segmented scoring rules of each control objective in the reward function and transform them into executable quantitative evaluation rules.

[0067] S23. Reward function rationality verification: The parsed reward function is compared and verified with the national and industry safety standards for distribution network operation. If the content does not meet the requirements of the standard, a secondary prompt for correction is triggered until the generated reward function fully complies with the power engineering safety standard.

[0068] By standardizing prompt templates to constrain the input and output of the large language model, the generated reward function is ensured to fit the distribution network control scenario and have a uniform format, avoiding uncontrollable output of the large model. By utilizing a large language model adapted to power industry corpus, dynamic matching between the scenario and the reward function is achieved, solving the core pain point that traditional fixed reward functions cannot adapt to the dynamic changes in the priority of multiple control targets under extreme weather conditions. Furthermore, through rationality verification and secondary correction, the reward function is ensured to comply with power safety regulations, which not only leverages the scenario reasoning advantages of the large language model but also ensures the engineering feasibility of the reward function, providing a scientific and compliant decision-making guide for subsequent reinforcement learning control.

[0069] Furthermore, the specific steps of S3 include:

[0070] S31. Construction of distribution network node feature matrix: Integrate the electrical operation characteristics of each node in the distribution network, the extreme weather characteristics of the area where the node is located, and the topological association characteristics of the node to construct a multi-dimensional node feature matrix adapted to graph neural network processing.

[0071] S32. Topology and Feature Dual Attention Graph Convolution Operation: A dual attention mechanism is introduced for the dynamic adjacency matrix of the distribution network that is updated in real time. The weights of feature correlation and topology connection strength between nodes are adaptively allocated, and the deep extraction and global fusion of node features are completed through multi-layer graph convolution operation.

[0072] S33. Feature Output and Consistency Verification: Pooling is performed on the hidden layer features output by the graph convolution to obtain a global feature vector that integrates dynamic topology information, extreme weather features, and node operating status. At the same time, the feature reconstruction error of the graph neural network is calculated, and a topology secondary verification is triggered based on a preset error threshold to avoid interference from erroneous topology information on subsequent control decisions.

[0073] By constructing a multi-dimensional feature matrix, the system achieves comprehensive integration of three core information categories: distribution network electrical systems, extreme weather, and topological connections. This addresses the problem of traditional feature extraction focusing only on a single dimension and providing incomplete information. The introduction of a dual attention mechanism adaptively focuses on key nodes and topological connection strengths significantly affected by extreme weather, improving the targeting and accuracy of feature extraction and adapting to the dynamic changes in distribution network topology under extreme weather conditions. Feature reconstruction error verification further filters out erroneous data, ensuring that the output global feature vector is true and reliable. This provides high-quality feature input for subsequent reinforcement learning decisions and avoids control decision errors caused by feature distortion.

[0074] Furthermore, the specific steps of S4 include:

[0075] S41. Construction of a reinforcement learning-based control decision framework: Based on the near-end policy optimization algorithm, a reinforcement learning-based decision framework adapted to the collaborative control of multiple devices in the photovoltaic-storage distribution network is constructed, clarifying the state space, action space, reward function, and policy update rules of the framework. The state space adopts the global feature vector output by S3 and the weight parameters of the reward function generated by S2. The reward function directly adopts the dynamic reward function generated by S2. The action space covers the adjustable commands of three types of control objects in the distribution network: energy storage, photovoltaic, and controllable load.

[0076] S42. Multi-device collaborative control decision generation: Based on the current system state, the Actor network of the reinforcement learning framework outputs multi-device collaborative control action commands for energy storage charging and discharging, photovoltaic flexible consumption, and controllable load time-series transfer within the distribution network. At the same time, the Critic network completes the state value assessment and completes the iterative optimization of the strategy based on preset rules.

[0077] S43. Safety constraint verification of control actions: Perform multi-level safety constraint verification on the output control action commands. The verification dimensions include at least the active power balance constraint of the distribution network, the rated operation constraint of each control device, and the transmission capacity constraint of the distribution network line, to ensure the engineering feasibility of the control actions and the safety of the distribution network operation.

[0078] By combining the weights of the dynamic reward function with the global feature vector to construct the state space, the reinforcement learning framework can perceive dynamic changes in the scene and the priority of the control target in real time, solving the problems of unreasonable state space design and inability to adapt to extreme weather scenarios in traditional reinforcement learning. The Actor-Critic dual network structure enables the simultaneous generation of control instructions and state value evaluation, improving decision-making efficiency and optimization accuracy. At the same time, it covers the action space of three types of control objects, realizing deep collaboration among multiple devices and solving the pain points of poor coordination and limited control effect of single device control. Multi-level constraint verification avoids control risks from the perspectives of energy balance, equipment safety, and line load capacity, ensuring that the generated control instructions are engineering-applicable and achieving the dual goals of safe operation of the distribution network and multi-objective optimization.

[0079] Furthermore, the specific steps of S5 include:

[0080] S51. Construction of training and testing datasets: Based on the standard distribution network topology model, combined with historical meteorological datasets of various types of extreme weather and distribution network full-scenario operation datasets, an algorithm training and testing dataset covering different extreme weather, different topology states, and different operating conditions is constructed.

[0081] S52. Multi-network phased training and joint optimization: The large language model, graph neural network, and reinforcement learning policy network are trained in phases, and then the joint optimization of the multiple networks is completed. Among them, the large language model is supervised and fine-tuned using professional corpus in the power field, the graph neural network is trained with the ability to extract topological features, and the reinforcement learning network is trained iteratively with the control strategy and updated with parameters.

[0082] S53. Model Validation and Engineering Deployment: After training and optimization, the overall control model is tested and validated under multiple scenarios and operating conditions. Once the validation is successful, the trained control strategy is deployed in the distribution network field control device.

[0083] By constructing a dataset covering all scenarios, the comprehensiveness and generalization of model training are ensured, avoiding the problem that the model cannot adapt to different extreme weather and topological conditions due to a single training data. Phased specialized training can specifically improve the core performance of each network. Fine-tuning of the large language model enhances its understanding of distribution network scenarios, graph neural network training improves feature extraction accuracy, and reinforcement learning network training optimizes control strategies. Then, joint optimization achieves the coordinated adaptation of each network, solving the problem of insufficient coordination and poor overall control effect caused by independent training of multiple networks. Model validation and engineering deployment ensure that the trained model has reliable control performance and can be directly applied to achieve real-time closed-loop control of photovoltaic storage distribution networks under extreme weather conditions, reducing engineering implementation costs.

[0084] Furthermore, the fixed prompt template in S21 also includes mandatory constraint clauses of the distribution network safety regulations. These mandatory constraint clauses include at least the legally acceptable operating range of distribution network voltage and frequency, the minimum priority guarantee threshold for each control target, and the normalization constraint rules for the reward function weights. The secondary prompt correction in S23 adopts a targeted correction mode, which only makes local adjustments to the content that does not comply with the mandatory constraint clauses, without changing the core logic of the weight allocation of the reward function to adapt to the current extreme weather scenario.

[0085] By embedding mandatory constraints of power distribution network safety regulations into fixed prompt templates, the output of the large language model is constrained from the source, ensuring that the generated reward function naturally conforms to power industry safety standards. This avoids non-compliant weight allocation and scoring rules, addressing the pain points of uncontrollable output and easy deviation from power safety requirements in large models. The targeted correction mode only makes partial adjustments to non-compliant content, ensuring the compliance of the reward function while retaining its core weight logic adapted to current extreme weather scenarios. This avoids the reward function losing its scenario adaptability due to comprehensive correction, further improving the reliability and applicability of the dynamic reward function.

[0086] Furthermore, the pre-trained generative large language model described in S22 is a lightweight large model that has undergone supervised fine-tuning using professional corpus data in the power sector. This professional corpus includes national and industry standards for distribution network operation, historical cases of emergency control of distribution networks under extreme weather conditions, and technical specifications for photovoltaic and energy storage grid connection. The natural language processing tool has built-in fixed field extraction rules, which can directly convert the unstructured text output by the large language model into structured weight parameters and segmentation scoring rules that can be directly called by reinforcement learning algorithms without manual intervention.

[0087] Supervised fine-tuning of a lightweight large language model using specialized corpora in the power sector enhances the model's understanding of distribution network control scenarios, extreme weather response logic, and power safety regulations. This addresses the issues of insufficient adaptability of general-purpose large language models to distribution network scenarios and the generation of reward functions that do not conform to engineering realities. The design of fixed field extraction rules enables the automatic conversion of unstructured outputs of the large language model to structured inputs for reinforcement learning, avoiding conversion errors and inefficiencies caused by manual intervention. This improves the efficiency of dynamic reward function generation and parsing, ensuring that the reward function can be quickly and accurately applied to subsequent reinforcement learning control processes.

[0088] Furthermore, the dual attention mechanism in S32 includes two types: feature correlation attention and topology connection strength attention. Feature correlation attention is used to calculate the feature correlation between nodes, identify key nodes that are significantly affected by extreme weather, and adaptively increase their feature weights. Topology connection strength attention is used to quantify the topology connection strength according to the real-time operating status of the line and adjust the connection weights between nodes accordingly.

[0089] In step S33, when the feature reconstruction error exceeds the preset error threshold, a secondary verification of the distribution network topology data and a node feature outlier removal operation are triggered simultaneously to avoid erroneous data interfering with subsequent control decisions.

[0090] By clearly dividing the dual attention mechanism into feature relevance attention and topological connection strength attention, focusing on node feature association and topological connection state respectively, graph neural networks can accurately identify key nodes and weak topological links under extreme weather conditions, improving the targeting of feature extraction and solving the problems of homogeneous feature extraction and inability to adapt to dynamic topology and the impact of extreme weather in traditional graph neural networks. The secondary verification and outlier removal operations when the feature reconstruction error exceeds the threshold can promptly detect and correct errors in topological data and node features, preventing erroneous information from entering the subsequent control decision-making process, further improving the reliability of feature extraction, providing accurate and clean feature support for reinforcement learning decision-making, and ensuring the scientific and effective nature of control decisions.

[0091] Example 2:

[0092] A large model-driven multi-objective photovoltaic-storage-distribution network control device under extreme weather conditions includes a sensing and scene generation module, a dynamic reward function generation module, a topology adaptive feature extraction module, a multi-objective collaborative control decision module, and a model training and deployment module.

[0093] The perception and scene generation module is used to collect real-time electrical data and extreme weather data of the power distribution network, identify topology change events based on voltage and current mutations, dynamically update the adjacency matrix and calculate the topology change confidence, and generate a structured natural language scene description.

[0094] The dynamic reward function generation module is communicatively connected to the perception and scene generation module. It is used to input the structured natural language scene description into a lightweight large language model fine-tuned by power field corpus. Based on preset prompt templates, output format constraints and parsing rules, it generates a dynamic multi-objective reward function containing four objectives: voltage stability, frequency recovery, sustainable energy storage and photovoltaic consumption, and extracts weight parameters and scoring rules.

[0095] The topology adaptive feature extraction module is communicatively connected to the perception and scene generation module. It is used to construct a node feature matrix that integrates distribution network electrical features, extreme weather features and topology association features. Based on the adjacency matrix, a topology and feature dual attention graph convolutional network is used to extract features, output a global high-dimensional feature vector, and verify the accuracy of the features by reconstructing the error.

[0096] The multi-objective collaborative regulation and control decision module is communicatively connected to the dynamic reward function generation module and the topology adaptive feature extraction module, respectively. It is used to set the policy network input state of the near-end policy optimization algorithm as a concatenated vector of the global high-dimensional feature vector and the weight parameters, set the action space as a combination of energy storage charging and discharging power, photovoltaic flexible curtailment power and transferable load adjustment power, set the reward function as the dynamic multi-objective reward function, and generate photovoltaic, energy storage and controllable load collaborative regulation and control instructions; perform multi-level verification of the regulation and control instructions based on energy balance constraints and equipment rated constraints, and output standardized regulation and control instructions after the verification is passed;

[0097] The model training and deployment module is communicatively connected to the dynamic reward function generation module, the topology adaptive feature extraction module, and the multi-objective collaborative control decision module, respectively. It is used to train the graph neural network and the near-end policy optimization network based on the extreme weather dataset, fine-tune the lightweight large language model, and deploy the trained algorithm to the distribution network control device to achieve real-time closed-loop control.

[0098] The following is an application example of the present invention:

[0099] Step 1, Extreme Weather - Topological Dynamic Sensing and Data Preprocessing:

[0100] 1.1 Data Acquisition:

[0101] Real-time electrical operation data of the power distribution network and raw meteorological observation data related to extreme weather are collected to provide basic input for subsequent modeling and sensing processing. The real-time electrical operation data of the power distribution network includes voltage, current, and power, while the extreme weather data includes temperature, wind speed, irradiance, icing thickness, and storm intensity.

[0102] 1.2 Dynamic Evolution Modeling of Extreme Weather:

[0103] (1) Temperature sequence generation:

[0104] An autoregressive (AR-1) stochastic process combined with a sinusoidal baseline was used to simulate hourly fluctuations in high-temperature heat waves (≥35℃) and cold waves (≤-10℃). Load changes were correlated with a temperature correction factor, which amplified the air conditioning load during high temperatures and the heating load during cold waves.

[0105] Temperature correction factor formula: ;

[0106] In the formula, This is the temperature influence coefficient. For real-time temperature, This is the reference temperature.

[0107] (2) Wind speed and storm modeling:

[0108] Storm conditions are simulated based on the superposition of random perturbations and gusts, with gusts described by an exponential decay function. State transition probabilities are described by a three-state Markov chain (stationary, storm, extreme storm) and mapped to a light attenuation factor. Stable state Extreme storms .

[0109] Gust attenuation function formula: ;

[0110] In the formula, The attenuation coefficient; The maximum instantaneous wind speed at the peak of the gust wave is the initial maximum value of the entire attenuation process; t is the attenuation time variable. Let t be the instantaneous gust wind speed.

[0111] (3) Generation of illumination sequence:

[0112] By employing a clear-sky radiation model, combined with a random cloud cover factor, and superimposed with cloud drift and storm shielding effects, a rapid fluctuation sequence of irradiance under extreme weather conditions is generated.

[0113] Clear-sky radiation model formula: ;

[0114] In the formula, The solar constant, , These are the sunrise and sunset times, respectively.

[0115] Formula for irradiance sequence under extreme weather conditions: ;

[0116] In the formula, For random cloud cover factor, This is the light attenuation factor.

[0117] 1.3 Dynamic sensing of distribution network topology:

[0118] To address the issues of drastic fluctuations in distribution network electrical quantities and frequent changes in topology caused by the extreme weather dynamic evolution model constructed in section 1.2, this invention further constructs a dynamic topology perception and confidence verification mechanism for the distribution network. By designing dual criteria of voltage and current mutations, it identifies topological events such as line tripping and switch switching in real time, dynamically corrects the topology adjacency matrix, and provides accurate topological feature input and data support for subsequent multi-source data fusion and structured scene generation.

[0119] (1) Topological change identification:

[0120] By detecting voltage and current surges at distribution network nodes, topology changes can be identified in real time, such as line disconnections, energy storage switching, and switch actions. The threshold values ​​for surge detection are as follows:

[0121] Voltage sudden change:

[0122] Sudden change in current:

[0123] in, The rated voltage of the distribution network node; This refers to the rated current of the line. , For distribution network node numbering, At the current sampling time, This refers to the previous sampling time.

[0124] (2) Dynamic adjacency matrix update:

[0125] Update the adjacency matrix based on topology change events. The rules are as follows:

[0126] If node and The lines between them are normally connected. If the line is disconnected or the switch is off, .

[0127] (3) Confidence check of topological changes:

[0128] The confidence level of topology change is calculated by judging the consistency between the amplitude of voltage and current surges and the fault recording. To avoid topological discrimination errors caused by false data under extreme weather conditions, the judgment rules are as follows: If the change is determined to be a genuine topological change, the adjacency matrix is ​​updated. If the data is deemed noise, the adjacency matrix will not be updated.

[0129] 1.4 Multi-source data fusion and scene description generation:

[0130] Extreme weather parameters, topology change events, and equipment operating status are transformed into structured natural language scene descriptions, serving as input prompts for a large language model. Extreme weather parameters include type and level, while equipment operating status includes voltage, frequency, energy storage SOC, and photovoltaic absorption rate.

[0131] Example as follows:

[0132] Current scenario: Strong wind storm (level 4), topology change: Line disconnection between nodes 3 and 4 (confidence level 0.92), operating status: Node 5 voltage deviation 2.1%, system frequency deviation 0.12Hz, energy storage SOC 38%, photovoltaic absorption rate 82%, the storm is forecast to intensify in the next hour, and the irradiance will decrease. 。

[0133] Step 2: Generation of dynamic reward function driven by large language model:

[0134] This step sets fixed prompt templates, fixed output formats, and fixed parsing rules. At the same time, a rule verification module is added to automatically verify the weight range and scoring threshold of the Large Language Model (LLM). Results that do not meet industry standards will trigger secondary prompts and force adjustments to ensure that the reward function meets the power system safety specifications. A pre-trained generative language model is introduced to dynamically generate an adapted multi-objective reward function based on real-time scene descriptions, replacing the traditional manual fixed reward function.

[0135] 2.1 LLM Input Prompt Construction:

[0136] Design a standardized prompt template that includes scenario information, output requirements, and output format constraints to ensure that the LLM generated results are parsable and executable. The fixed template is as follows:

[0137] Task: Generate a multi-objective reward function for the regulation and control of photovoltaic, energy storage, and distribution networks under extreme weather conditions.

[0138] I. Current Scene Information:

[0139] Extreme weather type: {weather type}, level: {w} (level 1-5, level 5 is the most severe)

[0140] Topological change: {Topological event description}, Confidence:

[0141] Real-time operating status: Voltage deviation %, frequency deviation }Hz, Energy Storage SOC (SOC%), Photovoltaic Grid Integration Rate %

[0142] Weather forecast: Forecast details for the next {Δt} hours

[0143] II. Reward Function Requirements:

[0144] The reward function needs to cover voltage recovery. Frequency recovery Sustainable energy storage Photovoltaic consumption Four major goals.

[0145] Output format: ,in , , , For dynamic weights, satisfying The reasons for each weight value must be explained; The scoring rules for each segment must be provided, clearly defining the acceptable range, reward value, and penalty value.

[0146] III. Constraints:

[0147] The weights need to be adapted to the current scenario priority, such as prioritizing voltage safety during stormy weather and prioritizing photovoltaic peak consumption.

[0148] The scoring rules must comply with the power distribution network safety standards, with a voltage of 0.95-1.05 pu and a frequency of 49.8-50.2 Hz.

[0149] 2.2 Generation and Analysis of LLM Reward Function:

[0150] LLM generates dynamic reward functions based on prompt templates and knowledge of the power system domain, such as the "Distribution Network Operation Regulations" and "Photovoltaic Grid Connection Technical Requirements." Example output is as follows:

[0151] The reward function adapted to the current strong wind storm (level 4) scenario is:

[0152]

[0153] Reason for weighting: The risk of line disconnection is high during storms; ensuring voltage safety is a priority. ) and frequency stability ( Secondly, maintain energy storage regulation capabilities. =0.2, photovoltaic grid connection priority is the lowest ( =0.1).

[0154] Scoring rules:

[0155] : 10 points for voltage deviation ≤1%, linearly decreasing from 1% to 3% (10 → 0 points), and deduct 200 points for >3%;

[0156] : 8 points for frequency deviation ≤ 0.05Hz, linearly decreasing from 0.05 to 0.1Hz (8 to 0 points), and deduct 150 points for deviation > 0.1Hz;

[0157] : SOC∈[40%,60%] gets 6 points, 30%-40% or 60%-70% gets 3 points, <30% or >70% deducts 50 points;

[0158] : 4 points for a disposal rate of ≥90%, 2 points for 80%-90%, and 10 points deducted for <80%.

[0159] Structured parsing is performed based on the generated reward function, and weight parameters in the LLM output are extracted using natural language processing (NLP) tools. , , , The scoring rules for each objective are transformed into mathematical expressions, as shown in the following example:

[0160] The weight parameters are , , ,

[0161] Voltage reward function:

[0162] Frequency recovery Sustainable energy storage Photovoltaic consumption All adopt the same segmented scoring rule as the voltage reward function mentioned above.

[0163] 2.3 Reasonableness Verification:

[0164] If there is a conflict between the analysis results and the distribution network safety standards, a secondary prompt for correction will be triggered until the generated results meet the engineering specifications.

[0165] Example of a secondary prompt: "Current" The acceptable frequency range does not conform to GB / T15945-2018. Please readjust the scoring rules according to 49.8-50.2Hz.

[0166] Step 3: Dynamic Topology Adaptive GNN Feature Extraction:

[0167] A graph neural network (GNN) module integrating dynamic topology and extreme weather features is constructed to provide accurate feature input for reinforcement learning decision-making. The specific implementation is as follows:

[0168] 3.1 Construction of Node Feature Matrix:

[0169] By integrating the electrical characteristics and extreme weather characteristics of distribution network nodes, a node characteristic matrix is ​​formed. , For the number of nodes, Feature dimensions are categorized into three types:

[0170] Electrical characteristics: node voltage Active power reactive power Current Energy storage Photovoltaic power output ;

[0171] Extreme weather characteristics: temperature in the area where the node is located Wind speed Irradiance Storm level ;

[0172] Topological characteristics: number of lines connected to a node, line impedance Confidence of topological change at the i-th node .

[0173] 3.2 Topology-Feature Dual Attention Graph Convolutional Network (GCN) Convolution:

[0174] For adjacency matrix A dual-attention mechanism is introduced to enhance feature extraction of critical nodes (such as faulty nodes and high-deviation nodes). The dual-attention mechanism is used for feature correlation attention and topology strength attention, respectively. The former identifies critical nodes that are significantly affected by weather; the latter reflects the health status of the line based on real-time impedance changes.

[0175] GCN convolutional layer formula:

[0176]

[0177] In the formula, It is a dual attention weight matrix. for The degree matrix, For the first Hidden features in layers For trainable weights, This is the ReLU activation function.

[0178] Formula for calculating the dual attention weight matrix:

[0179]

[0180] In the formula, This is the attention balance coefficient.

[0181] Feature-related attention expression:

[0182]

[0183] Nodes are measured by calculating cosine similarity. and The higher the relevance of a feature, the greater its weight.

[0184] Topology connection strength attention expression:

[0185]

[0186] Through line impedance The lower the impedance, the greater the weight in measuring the strength of the topology connection.

[0187] 3.3 Feature Output and Consistency Verification:

[0188] After two layers of GCN convolution, the output of the second layer... Global average pooling is performed to obtain a high-dimensional global feature vector that integrates dynamic topology, extreme weather, and node states. , The output feature dimension is calculated; at the same time, the GNN reconstruction error is calculated to detect topology update errors or sensor anomalies.

[0189] GNN reconstruction error formula:

[0190]

[0191] In the formula, The feature matrix reconstructed by GNN decoding.

[0192] Anomaly detection rules: When reconstruction error When the value exceeds a threshold, the system triggers secondary topology verification to prevent incorrect topology from affecting decision-making. The threshold is determined by offline verification, such as 0.5.

[0193] Step 4: Multi-device collaborative decision-making and constraint verification:

[0194] Based on the Proximal Policy Optimization (PPO) algorithm, combined with the dynamic reward function generated by LLM and the features extracted by GNN, collaborative decision-making of photovoltaic, energy storage and transferable loads is realized, and the feasibility of regulation is ensured through constraint verification.

[0195] 4.1 Reinforcement Learning Decision Framework:

[0196] The core design includes the state space, action space, reward function, and PPO policy update rules, as detailed below:

[0197] (1) State space S:

[0198] Global feature vector output by GNN The four weight parameters of the reward function generated by LLM ( Composed of, namely:

[0199]

[0200] (2) Action space A:

[0201] It consists of the actions of three types of control objects: energy storage, photovoltaics, and load, as detailed below:

[0202] Energy storage actions: for each energy storage node The action includes its charging and discharging power commands. Positive values ​​represent discharging to support the load, while negative values ​​represent charging to absorb excess photovoltaic power. The action constraints are as follows:

[0203]

[0204] In the formula, , These represent the upper and lower limits of the energy storage charging and discharging power, respectively.

[0205] Photovoltaic Action For each photovoltaic node , representing the flexible curtailment power of photovoltaic power, satisfies the following constraints:

[0206]

[0207] In the formula, Let j be the available power of the photovoltaic system.

[0208] Load action For each controllable load node , representing transferable load The regulating power is positive when the load is transferred to the current time period, and negative when the load is transferred out of the current time period, satisfying the constraint:

[0209]

[0210] In the formula, , These represent the upper and lower limits of the transferable load, respectively.

[0211] In summary, the complete action vector is:

[0212]

[0213] (3) Reward function :

[0214] The dynamic reward function generated by LLM is directly used to calculate the reward value corresponding to the current action in real time, guiding the agent to learn the control strategy adapted to the scenario.

[0215] (4) PPO strategy update:

[0216] The PPO algorithm includes an Actor network and a Critic network: the Actor network is state-based. Output action probability distribution, Critic network evaluates state value .

[0217] Actor network update: The update parameters are truncated using a clipping strategy to avoid excessively large policy update magnitudes. The loss function is as follows:

[0218]

[0219] In the formula, For the dominant function, This is the clip cutoff factor, with a value of 0.1. , These are the Actor network parameters before and after the update.

[0220] Critic network update: Parameters are updated by minimizing the state value prediction error, using temporal difference (TD) error optimization. The loss function is:

[0221] In the formula, These are the parameters for the Critic network. For the Critic network, the current state Value estimation; The target value is calculated using generalized advantage estimation (GAE), as shown in the following formula:

[0222]

[0223] In the formula, For TD error, As a discount factor, For GAE coefficients, These are the target network parameters.

[0224] 4.2 Constraint Verification:

[0225] To avoid equipment overload or system instability caused by control actions, the decision output needs to be verified through multiple layers of constraints, as follows:

[0226] (1) Energy balance constraint:

[0227] To ensure the balance of active power in the distribution network and avoid frequency fluctuations caused by power deficits or surpluses, the constraint formula is as follows:

[0228]

[0229] In the formula, Let t be the sum of the total output power of all photovoltaic devices at time t; Let t be the sum of the flexible curtailment power of all photovoltaic devices; The sum of the original active power of all loads in the distribution network at time t; It is the sum of the regulating power of all transferable loads at time t; Let t be the sum of the charging and discharging power of all energy storage devices at time t; Let t be the total power loss of the distribution network lines.

[0230] (2) Equipment rated constraints:

[0231] Apparent power constraint of photovoltaic inverters: ;

[0232] In the formula, Contributing reactive power to photovoltaic power generation; This refers to the inverter's rated apparent power.

[0233] Energy storage constraints: In addition to charge / discharge power constraints, SOC dynamic constraints must be met.

[0234]

[0235] In the formula, To improve the charging and discharging efficiency of energy storage; For energy storage, the power is self-lossing. This refers to the rated capacity of the energy storage.

[0236] Simultaneously, the upper and lower limits of SOC must be satisfied: ;

[0237] In the formula, , .

[0238] Line transmission constraints: ;

[0239] In the formula, Let be the active power transmitted by line ij at time t; Let be the reactive power transmitted by line ij at time t; This represents the rated maximum transmission capacity of line ij.

[0240] Step 5, Algorithm Training and Deployment:

[0241] 5.1 Data Preparation:

[0242] Based on the IEEE 33-node distribution network model, which includes three 100kW photovoltaic clusters, four 50kWh energy storage nodes, and a 200kW transferable load cluster, training and testing datasets are constructed using multi-source historical meteorological datasets. These datasets cover high-temperature, cold-wave, storm, and icing scenarios, as well as irradiance... The sampling frequency is 10Hz and the duration is 100s.

[0243] 5.2 Network Training and Optimization:

[0244] (1) LLM fine-tuning:

[0245] The lightweight LLM is supervisedly fine-tuned using a power sector corpus to optimize its understanding of power scenarios. After fine-tuning, the parsing accuracy of the LLM's reward function generation should be ≥95%. Power sector corpora include documents such as the "Distribution Network Operation Regulations" and extreme weather control cases.

[0246] (2) GNN training:

[0247] Train a dynamic topology adaptive GNN based on topology change samples and node feature samples to minimize reconstruction error. The feature extraction accuracy after training should be ≥92%.

[0248] (3) PPO training and updating:

[0249] Initialize the Actor / Critic network with hidden layer dimensions [64, 32], setting the learning rate to 0.0003, GAE discount factor to 0.95, training epochs to 14000, and batch size to 32. Samples are then collected through empirical replay to jointly optimize the GNN reconstruction loss. Critic loss Actor loss The network parameters are updated using the Adam optimizer.

[0250] After training, the model is tested and verified, and finally the trained control strategy is deployed in the field device.

[0251] Based on the IEEE 33-node system and the NREL2023 extreme storm dataset, with irradiance ranging from 50 to 1000 W / m² and a duration of 100 s, the control effects of the method described in this invention are compared with those of GNN-PPO, the multi-agent reinforcement learning algorithm MAPPO, and the large model-assisted reinforcement learning algorithm LLM-PPO. The comparison results are as follows:

[0252] First, the reward function curve of the proposed regulation method is as follows: Figure 2 As shown.

[0253] Voltage deviation rate for example Figure 3As shown.

[0254] As can be seen from the figure, the proposed method consistently remains below the 1% safety threshold, significantly outperforming other methods. It exhibits the best voltage stability under extreme storm conditions, and the larger model-assisted PPO method reduces voltage by approximately 40%.

[0255] Frequency fluctuations, for example Figure 4 As shown.

[0256] As can be seen from the figure, the fluctuation amplitude of the proposed method is far below the 0.05Hz safety threshold, which is better than the three selected comparison algorithms, and the frequency control accuracy is the highest in extreme scenarios.

[0257] Photovoltaic grid integration rate for example Figure 5 As shown.

[0258] As shown in the figure, the absorption rate of the method presented in this paper is consistently above 95%, far exceeding other methods, and the photovoltaic utilization rate is optimal under extreme weather conditions.

[0259] Energy storage SOC maintenance rate, for example Figure 6 As shown.

[0260] As shown in the figure, the method presented in this paper consistently maintains a safety threshold of 85%, demonstrating the best stability and sustainability in energy storage regulation.

[0261] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A large-model-driven method for multi-objective photovoltaic-storage-distribution network regulation under extreme weather conditions, characterized in that: Includes the following steps: S1. Collect real-time electrical data and extreme weather data of the power distribution network, identify topology change events based on voltage and current mutations, dynamically update the adjacency matrix and calculate the topology change confidence, and generate a structured natural language scene description. S2. Input the structured natural language scene description into a lightweight large language model fine-tuned by power field corpus. Based on preset prompt templates, output format constraints and parsing rules, generate a dynamic multi-objective reward function containing four objectives: voltage stability, frequency recovery, sustainable energy storage and photovoltaic consumption. Extract weight parameters and scoring rules. S3. Construct a node feature matrix that integrates distribution network electrical features, extreme weather features, and topological association features. Based on the adjacency matrix, use a topological and feature dual attention graph convolutional network to extract features, output a global high-dimensional feature vector, and verify the accuracy of the features by reconstructing the error. S4. Set the input state of the policy network of the near-end policy optimization algorithm to the concatenation vector of the global high-dimensional feature vector and the weight parameters, set the action space to the combination of energy storage charging and discharging power, photovoltaic flexible curtailment power and transferable load adjustment power, set the reward function to the dynamic multi-objective reward function, and generate photovoltaic, energy storage and controllable load coordinated control instructions; perform multi-level verification of the control instructions on energy balance constraints and equipment rated constraints, and output standardized control instructions after the verification is passed; S5. Based on the extreme weather dataset, train a graph neural network and a near-end policy optimization network, fine-tune the lightweight large language model, and deploy the trained algorithm to the distribution network control device to achieve real-time closed-loop control.

2. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 1, characterized in that, The specific steps in S1 include: S11. Data Acquisition: Collect real-time electrical operation data of the power distribution network, as well as raw meteorological observation data related to extreme weather, to provide basic input for subsequent modeling and sensing processing. S12. Dynamic evolution modeling of extreme weather: For different types of extreme weather such as high temperature, cold wave, and storm, dynamic evolution models of meteorological elements corresponding to temperature, wind speed, and light intensity are constructed respectively to simulate the real-time fluctuation patterns of various meteorological elements under extreme weather conditions, and at the same time, the impact characteristics of extreme weather on power distribution network load and photovoltaic output are quantitatively characterized. S13. Distribution network topology dynamic perception: Identify topology change events corresponding to line on / off, equipment switching, and switch actions by real-time mutations in electrical quantities of distribution network nodes. Update the distribution network dynamic adjacency matrix based on the identified topology change events. At the same time, perform confidence verification on the topology change events and determine whether to update the adjacency matrix based on the confidence verification results to avoid topology discrimination errors caused by false data. S14. Multi-source data fusion and scene description generation: Extreme weather-related parameters, topology change events and corresponding verification results, and real-time operating status data of distribution network equipment are fused from multiple sources and transformed into standardized structured natural language scene descriptions, which serve as input prompts for subsequent large language models.

3. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 1, characterized in that, The specific steps in S2 include: S21. Standardized input prompt construction: Design fixed prompt templates adapted to distribution network control scenarios, integrate current extreme weather types and levels, distribution network topology change events, real-time system operating status, and future weather forecast information, while clarifying the coverage target, output format, and industry standard constraints of the reward function; S22. Dynamic reward function generation and structured parsing: Based on the input prompt template, a pre-trained generative large language model is used to generate a multi-objective dynamic reward function that is adapted to the current scenario. Then, natural language processing tools are used to extract the weight allocation logic and the segmented scoring rules of each control objective in the reward function and transform them into executable quantitative evaluation rules. S23. Reward function rationality verification: The parsed reward function is compared and verified with the national and industry safety standards for distribution network operation. If the content does not meet the requirements of the standard, a secondary prompt for correction is triggered until the generated reward function fully complies with the power engineering safety standard.

4. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 1, characterized in that, The specific steps in S3 include: S31. Construction of distribution network node feature matrix: Integrate the electrical operation characteristics of each node in the distribution network, the extreme weather characteristics of the area where the node is located, and the topological association characteristics of the node to construct a multi-dimensional node feature matrix adapted to graph neural network processing. S32. Topology and Feature Dual Attention Graph Convolution Operation: A dual attention mechanism is introduced for the dynamic adjacency matrix of the distribution network that is updated in real time. The weights of feature correlation and topology connection strength between nodes are adaptively allocated, and the deep extraction and global fusion of node features are completed through multi-layer graph convolution operation. S33. Feature Output and Consistency Verification: Pooling is performed on the hidden layer features output by the graph convolution to obtain a global feature vector that integrates dynamic topology information, extreme weather features, and node operating status. At the same time, the feature reconstruction error of the graph neural network is calculated, and a topology secondary verification is triggered based on a preset error threshold to avoid interference from erroneous topology information on subsequent control decisions.

5. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 1, characterized in that, The specific steps in S4 include: S41. Construction of a reinforcement learning control decision framework: Based on the near-end policy optimization algorithm, a reinforcement learning decision framework adapted to the collaborative control of multiple devices in the photovoltaic-storage distribution network is constructed, clarifying the state space, action space, reward function, and policy update rules of the framework; the state space adopts the global feature vector output in step S3 and the weight parameters of the reward function generated in step S2, the reward function directly adopts the dynamic reward function generated in step S2, and the action space covers the adjustable commands of three types of control objects in the distribution network: energy storage, photovoltaic, and controllable load; S42. Multi-device collaborative control decision generation: Based on the current system state, the Actor network of the reinforcement learning framework outputs multi-device collaborative control action commands for energy storage charging and discharging, photovoltaic flexible consumption, and controllable load time-series transfer within the distribution network. At the same time, the Critic network completes the state value assessment and completes the iterative optimization of the strategy based on preset rules. S43. Safety constraint verification of control actions: Perform multi-level safety constraint verification on the output control action commands. The verification dimensions include at least the active power balance constraint of the distribution network, the rated operation constraint of each control device, and the transmission capacity constraint of the distribution network line, to ensure the engineering feasibility of the control actions and the safety of the distribution network operation.

6. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 1, characterized in that, The specific steps of S5 include: S51. Construction of training and testing datasets: Based on the standard distribution network topology model, combined with historical meteorological datasets of various types of extreme weather and distribution network full-scenario operation datasets, an algorithm training and testing dataset covering different extreme weather, different topology states, and different operating conditions is constructed. S52. Multi-network phased training and joint optimization: The large language model, graph neural network, and reinforcement learning policy network are trained in phases, and then the joint optimization of the multiple networks is completed. Among them, the large language model is supervised and fine-tuned using professional corpus in the power field, the graph neural network is trained with the ability to extract topological features, and the reinforcement learning network is trained iteratively with the control strategy and updated with parameters. S53. Model Validation and Engineering Deployment: After training and optimization, the overall control model is tested and validated under multiple scenarios and operating conditions. Once the validation is successful, the trained control strategy is deployed in the distribution network field control device.

7. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 3, characterized in that, The fixed prompt template in S21 also includes mandatory constraint clauses of the distribution network safety regulations. These mandatory constraint clauses include at least the legally acceptable operating range of distribution network voltage and frequency, the minimum priority guarantee threshold for each control target, and the normalization constraint rules for the reward function weights. The secondary prompt correction in S23 adopts a targeted correction mode, which only makes local adjustments to the content that does not comply with the mandatory constraint clauses, without changing the core logic of the weight allocation of the reward function to adapt to the current extreme weather scenario.

8. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 3, characterized in that, The pre-trained generative large language model described in S22 is a lightweight large model that has undergone supervised fine-tuning using professional corpus data in the power sector. This professional corpus includes national and industry standards for distribution network operation, historical cases of emergency control of distribution networks under extreme weather conditions, and technical specifications for photovoltaic and energy storage grid connection. The natural language processing tool has built-in fixed field extraction rules, which can directly convert the unstructured text output by the large language model into structured weight parameters and segmentation scoring rules that can be directly called by reinforcement learning algorithms without manual intervention.

9. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 4, characterized in that, The dual attention mechanism in S32 includes two types: feature correlation attention and topology connection strength attention. Feature correlation attention is used to calculate the feature correlation between nodes, identify key nodes that are significantly affected by extreme weather and adaptively increase their feature weights. Topology connection strength attention is used to quantify the topology connection strength according to the real-time operating status of the line and adjust the connection weights between nodes accordingly.

10. The large-model-driven multi-objective photovoltaic-storage-distribution network regulation method under extreme weather conditions according to claim 4, characterized in that, In step S33, when the feature reconstruction error exceeds the preset error threshold, a secondary verification of the distribution network topology data and a node feature outlier removal operation are triggered simultaneously to avoid erroneous data interfering with subsequent control decisions.

Images