A Multi-Agent Voltage Control Method for Distribution Networks Guided by a Large Language Model

Abstract

This invention discloses a multi-agent voltage control method for power distribution networks based on a large language model, comprising the following steps: deploying multiple agents in the power distribution network system, with each agent as the control object and a control objective set; before each new training round, triggering a game-theoretic debate mechanism among multiple heterogeneous large language models to generate physical control parameters; converting the generated physical control parameters into action guidance signals understood by the agents; multiple agents in the continuous execution layer interacting, with dynamically decaying rewards shaping and guiding agent exploration; after the current training cycle ends, extracting multi-dimensional time-series measurement data from the continuous execution layer to trigger a human-like debriefing and reflection mechanism of the large language model, forming a closed-loop reflection chain. This invention relies on an experience memory bank and a cross-round knowledge feedback mechanism to permanently solidify the patterns summarized from historical debriefing, overcoming the catastrophic forgetting problem in reinforcement learning during environmental transfer.

Description

Technical Field

[0001] This invention relates to the field of smart grid industry, and in particular to a multi-agent voltage control method for distribution networks based on a large language model. Background Technology

[0002] With the high proportion of distributed power sources (such as photovoltaic and wind power) integrated into distribution networks, the randomness and bidirectional fluctuations of power flow in distribution network systems have significantly increased, leading to increasingly severe problems of node voltage exceeding limits. In recent years, Multi-Agent Deep Reinforcement Learning (MADRL) has been widely used in distribution network voltage control and flexible resource (such as distributed energy storage) scheduling. Agents can autonomously learn and output voltage control strategies based on real-time grid measurements and through continuous trial and error with the environment. However, purely data-driven neural network models have inherent physical blind spots and collaborative silos, making it difficult to guarantee the stability and security of control when facing complex and ever-changing actual distribution network operating conditions.

[0003] To address the aforementioned control bottlenecks, existing reinforcement learning methods suffer from significant shortcomings in practical engineering applications across the following three dimensions:

[0004] First, existing reinforcement learning strategies suffer from severe deficiencies in physical understanding and interpretability barriers. Traditional reinforcement learning is essentially a black-box optimization based on high-dimensional numerical mappings, where the agent cannot understand the topology and physical operating mechanisms of the distribution network (e.g., it cannot recognize that "the voltage sensitivity of the end node of a long feeder is the highest" or "heavily loaded nodes are prone to instability"). This blind trial and error, lacking common sense about physics, results in extremely low exploration efficiency for the agent in the early stages of training; moreover, the final generated control strategies often lack physical logic support, making it difficult for dispatchers to conduct safety assessments of their decision boundaries, which greatly limits their deployment in real power grids.

[0005] Secondly, multi-agent distributed control faces severe coordination conflicts and the curse of dimensionality. In the scenario of multi-node energy storage coordination in distribution networks, existing methods typically rely on static, manually designed joint reward functions. However, due to the highly complex operating conditions of the power grid (such as a sudden increase in net demand or a sharp drop in renewable energy output), static reward functions cannot be dynamically and adaptively adjusted. This leads to multiple agents easily falling into a situation of acting independently when faced with an emergency voltage limit exceedance, frequently resulting in "action conflicts" (such as simultaneous charging and discharging of adjacent energy storage nodes causing offsetting of regulation) or "extreme imbalance of state of charge (SOC)," seriously affecting the overall control margin of the system and the lifespan of the equipment.

[0006] Finally, traditional reinforcement learning training paradigms lack human-like debriefing and reflection capabilities, resulting in severe experiential forgetting and knowledge gaps. Existing reinforcement learning models update neural network weights through temporal differencing or policy gradients, essentially a form of passive, muscle-memory-like learning. When an agent completes a training round or faces control failure, it cannot, like a human expert, deeply review and reflect on its historical control trajectory. For example, it cannot summarize the physical experience that the previous failure was due to premature discharge of energy stored at the end of the feeder, leading to a subsequent voltage collapse. This blank-slate learning mode, lacking post-training review and knowledge extraction, forces the agent to engage in blind, full-domain trial and error again whenever it switches training scenarios or starts a new round of training. Excellent collaborative strategies discovered by chance in the past are locked within a massive number of neuron parameters, easily forgotten due to gradient overlay in subsequent training, failing to form explicit, reusable, and cross-round inheritable physical control rules. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a simple and robust multi-agent voltage control method for distribution networks based on a large language model.

[0008] The technical solution of the present invention to solve the above-mentioned technical problems is: a multi-agent voltage control method for distribution networks based on a large language model is implemented based on a three-layer spatial architecture, which includes a multi-model cognition and game theory layer at the top layer, a physical prior bridging layer in the middle layer, and a continuous execution layer at the bottom layer.

[0009] In the multi-model cognition and game layer, multiple heterogeneous large language models are deployed and experience memory banks and their respective reputation values ​​are maintained. Each language model performs independent virtual scheduling, proposes strategies, debates with each other and performs knowledge distillation for the current power grid situation, and obtains the physical control parameters of the winner of the debate.

[0010] In the physical prior bridging layer, a strategy evaluation and physical guidance module is deployed, which is responsible for converting the physical control parameters of the winner of the large language model debate into continuous action guidance signals;

[0011] In the continuous execution layer, a power distribution network simulation environment and multiple intelligent agents are deployed. With the assistance of action guidance signals, the intelligent agents try and fail, and output continuous charging and discharging power.

[0012] Includes the following steps:

[0013] Step 1: Experimental scenario and environment setup. Deploy multiple intelligent agents in the power distribution network system, using the intelligent agents as control objects and setting control objectives;

[0014] Step 2: After completing the experimental scenario and environment settings, initialize the training task and obtain the current operating situation and historical experience of the power distribution network system. Before each new training round starts, trigger the game debate mechanism of multiple heterogeneous large language models to generate topologically differentiated physical control parameters for the agents of the continuous execution layer.

[0015] Step 3: Convert the generated physical control parameters into action guidance signals that the agent can understand with zero latency during millisecond-level execution;

[0016] Step 4: Multiple agents in the continuous execution layer interact, and the agents are guided to explore through dynamically decaying rewards without changing the basic optimization goal;

[0017] Step 5: After the current training cycle ends, extract the multi-dimensional time series measurement data of the continuous execution layer, trigger the human-like debriefing and reflection mechanism of the large language model, extract the cause and effect of success and failure, and form a closed-loop reflection chain.

[0018] The above-mentioned multi-agent voltage control method for distribution networks based on a large language model, in step 1, the agent... At time step The continuous action output is , After physical gain amplification, the corresponding charging and discharging power is [-4.0, 4.0] megawatts;

[0019] The enhanced state space of an agent is defined as follows:

[0020] ;

[0021] in, Represents intelligent agents At time step An enhanced state space; This represents a vector concatenation operation; For intelligent agents At time step The actual activation prompt vector received; For intelligent agents The node at time step Real-time voltage; For intelligent agents The node at time step The state of charge; These are the normalized eigenvalues ​​for each time step; For time step The system's real-time load power; For time step The photovoltaic power generation capacity; For time step The wind power generation capacity;

[0022] The control objective is to minimize the system-level voltage deviation objective function of the distribution network nodes while ensuring the safety of the energy storage charge state of each agent, and at the same time, to minimize the grid oscillations caused by blind trial and error by multiple agents.

[0023] The specific process of step 2 in the above-mentioned multi-agent voltage control method for distribution networks guided by a large language model is as follows:

[0024] Step 21, Generate independent proposals: Multiple large language models read historical trajectory features from the experience memory bank, and each language model formulates a strategy based on the differences in the sensitivity of the power grid spatial topology, and independently outputs a set of physical control parameters for each agent;

[0025] Step 22, Cross-examination and Revision: Multiple rounds of mutual review are conducted among the large language models;

[0026] Step 23: Strategies are scored: The strategies developed by each language model are scored from four dimensions: physical rationality, spatial difference, debate quality, and model historical reputation. The strategy with the highest score is selected as the action configuration for this round, and the language model corresponding to the strategy with the highest score wins.

[0027] In the aforementioned multi-agent voltage control method for distribution networks guided by a large language model, step 21 includes physical control parameters in the physical control parameter set, which encompass the agents. Voltage droop gain Intelligent agents Target reference voltage Intelligent agents State of charge protection gain and intelligent agents static bias .

[0028] In the above-mentioned distribution network multi-agent voltage control method based on large language model, in step 22, in each round of mutual review, two large language models are selected, denoted as model A and model B respectively. Model A reviews the proposal of model B, points out the risk of cooperative failure that model B may cause, and issues criticism. After receiving the criticism, model B evaluates the physical rationality of the criticism and modifies or physically and logically defends the initial proposal.

[0029] In the aforementioned multi-agent voltage control method for distribution networks guided by a large language model, the formula for calculating the comprehensive score in step 23 is as follows:

[0030] ;

[0031] in, Representing a large language model Overall score; For large language models The score for physical plausibility; For large language models Spatial difference score; For large language models The debate quality score; Large language models recorded in the experience memory bank Historical reputation value; For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weight coefficients, and satisfying ;

[0032] The formula used to evaluate whether the physical control parameters generated by the large language model meet the safe operation boundary of the distribution network is:

[0033] ;

[0034] in, A collection of intelligent agents; For intelligent agents Voltage droop gain; For intelligent agents State-of-charge protection gain; This is the effectiveness index function for voltage droop gain; The margin index function for the state-of-charge protection gain; Configure physical penalty terms for parameters that may trigger energy storage exceeding limits; The physical evaluation weighting coefficients for the corresponding effectiveness indicators; For the physical evaluation weighting coefficients corresponding to the sufficiency index; This refers to the physical evaluation weighting coefficient corresponding to the physical penalty item;

[0035] The calculation formula is:

[0036] ;

[0037] in, The total number of agents; For large language models For intelligent agents The generated physical control parameter vector; For large language models For intelligent agents The generated physical control parameter vector; This is the spatial difference sensitivity adjustment coefficient; The Euclidean norm 2 of a vector;

[0038] The calculation formula is:

[0039] ;

[0040] in, For large language models The validity score indicates the vulnerabilities of other model proposals; For large language models The strategic gain component resulting from modifying the initial proposal after incorporating valid criticism; For the corresponding Debate scoring coefficient; For the corresponding The debate scoring coefficient.

[0041] The specific process of step 3 in the above-mentioned multi-agent voltage control method for distribution networks guided by a large language model is as follows:

[0042] Step 31, Physical Reference Motion Generation: At each time step in the underlying layer Physical prior bridging layer acquires intelligent agents The node at time step Real-time voltage With intelligent agents The node at time step State of charge ;

[0043] Using the physical control parameters output by the winning large language model, continuous recommended actions are calculated in real time through the voltage droop control equation in the physical domain:

[0044] ;

[0045] in, Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function, used to force the calculated action to be limited to the upper and lower limits allowed by the agent's hardware. Within the range;

[0046] Step 32, Enhance observation space stitching: Calculate the... As a semantic cue vector, it is directly concatenated into the agent's augmented state space to construct an augmented observation state space;

[0047] ;

[0048] in, Represents intelligent agents At time step Enhanced observation state space; Represents intelligent agents At time step An enhanced state space; This indicates a vector concatenation operation.

[0049] The specific process of step 4 in the above-mentioned multi-agent voltage control method for distribution networks guided by a large language model is as follows:

[0050] Step 41, Calculate the dynamic guided reward: agent Output Then, calculate and To ensure consistency, a guiding reward that monotonically decays with the training process is applied.

[0051] ;

[0052] in: Represents intelligent agents At time step The dynamic guided reward obtained; For intelligent agents At time step Continuous action output; Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function; This represents the basic weighting coefficient for the guiding reward; This represents a smooth decay factor that monotonically decreases with the overall training progress;

[0053] Step 42, Composite-driven smooth transition: Calculate the total single-step reward that the agent ultimately uses for gradient updates;

[0054] ;

[0055] in, Represents intelligent agents At time step The final comprehensive reward value used for network gradient updates; Indicates at time step Basic physical penalties for feedback in power distribution network simulation environments;

[0056] As the training progresses, As the dynamic rewards gradually decrease and the control is smoothly transferred to the distribution network itself, the intelligent agent can autonomously explore extreme collaborative strategies that surpass large language models on the basis of safety benchmarks.

[0057] The specific process of step 5 in the above-mentioned multi-agent voltage control method for distribution networks guided by a large language model is as follows:

[0058] Step 51, Multi-dimensional temporal profile generation and reputation update: Collect the global temporal trajectory of the underlying control in this round, transform the global temporal trajectory into a natural language structured profile readable by the large language model, and update the historical reputation values ​​of the various language models participating in the debate based on the actual voltage control performance.

[0059] Step 52, In-depth review and cross-distillation based on the perspective of victory and defeat: triggering a five-step closed-loop reflection chain to extract physical laws;

[0060] Step 53, Reflection Memory Consolidation and Cross-Round Feedback: Store the physical laws extracted from the review and reflection into a persistent experience memory bank. When starting Step 2 of the next round, force all major language models to prioritize reading their own exclusive failure lessons, and use the failure lessons as a mandatory constraint boundary for generating new proposals.

[0061] In the aforementioned multi-agent voltage control method for distribution networks guided by a large language model, step 52 includes a five-step closed-loop reflection chain:

[0062] Step 1, in-depth review of the winner: Combine the winning big language model with the timing profile to reflect on the physical context in which the voltage droop gain played a key role.

[0063] Step 2, Loser's Attribution Reflection: The defeated big language model compares its initial proposal with the actual power grid trajectory and reflects on its blind spots in physical cognition;

[0064] Step 3, cross-model teaching: The winning big language model provides customized improvement suggestions based on the reflection results of the losing big language model;

[0065] Step 4, defensive feedback: The defeated big language model points out the potential hidden dangers or overfitting risks in the strategies formulated by the winning big language model;

[0066] Step 5, Consensus Law Synthesis: The conclusions of the multidimensional reflections in Steps 1 to 4 above are synthesized and condensed into bilateral shared principles and specific lessons for improvement.

[0067] The beneficial effects of this invention are as follows:

[0068] 1. This invention innovatively introduces a global review mechanism after training, breaking the underlying logic of passive gradient calculation in traditional reinforcement learning. By introducing a large language model to compare extreme samples and reflect on causality after each round of training, it can autonomously summarize the physical root causes of the success or failure of the strategy. It makes the excellent strategies accidentally discovered in the neural network explicit into an interpretable logical rule structure based on explicit condition triggers, thus meeting the stringent requirements of power dispatch for safety and transparency.

[0069] 2. Constructing a knowledge loop of trial and error, review, accumulation, and feedback to overcome the barrier of forgetting experience: This invention relies on an experience memory bank and a cross-cycle knowledge feedback mechanism to permanently solidify the patterns summarized from historical reviews. When facing a new round of control tasks, the agent starts directly from historical experience, which greatly accelerates the cold start convergence speed and overcomes the catastrophic forgetting pain point of reinforcement learning in environmental transfer.

[0070] 3. This invention transforms high-level, macroscopic, discrete collaborative instructions into continuous, stable low-level guidance signals. This effectively avoids the agent's observation domain jumps and training crashes caused by hard rule injection, ensuring the absolute stability and security of the model during the absorption of high-dimensional semantic knowledge. Attached Figure Description

[0071] Figure 1 This is the overall flowchart of the present invention.

[0072] Figure 2 This diagram shows the results of multiple rounds of training for the present invention and traditional multi-agent reinforcement learning algorithms. Detailed Implementation

[0073] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0074] like Figure 1 As shown, the multi-agent voltage control method for distribution networks guided by a large language model is implemented based on a three-layer spatial architecture, which includes a multi-model cognition and game theory layer at the top, a physical prior bridging layer in the middle, and a continuous execution layer at the bottom.

[0075] In the multi-model cognition and game layer, multiple heterogeneous large language models are deployed and experience memory banks and their respective reputation values ​​are maintained. Each language model performs independent virtual scheduling, proposes strategies, debates with each other and performs knowledge distillation for the current power grid situation, and obtains the physical control parameters of the winner of the debate.

[0076] In the physical prior bridging layer, a strategy evaluation and physical guidance module is deployed, which is responsible for converting the physical control parameters of the winner of the large language model debate into continuous action guidance signals.

[0077] In the continuous execution layer, a power distribution network simulation environment and multiple intelligent agents are deployed. With the assistance of action guidance signals, the intelligent agents try and fail, and output continuous charging and discharging power.

[0078] A multi-agent voltage control method for distribution networks based on a large language model includes the following steps:

[0079] Step 1: Experimental scenario and environment setup. In the IEEE 33-node distribution network system, deploy 3 agents. The set of agents is defined as follows: , These correspond to grid-connected nodes 17, 21, and 32, respectively, with the intelligent agent as the control object and the control target set.

[0080] intelligent agent At time step The continuous action output is , After physical gain amplification, the corresponding charging and discharging power is [-4.0, 4.0] megawatts;

[0081] The enhanced state space of an agent is defined as follows:

[0082] ;

[0083] in, Represents intelligent agents At time step An enhanced state space; This represents a vector concatenation operation; For intelligent agents At time step The actual activation prompt vector received; For intelligent agents The node at time step Real-time voltage; For intelligent agents The node at time step The state of charge; These are the normalized eigenvalues ​​for each time step; For time step The system's real-time load power; For time step The photovoltaic power generation capacity; For time step The wind power generation capacity.

[0084] The control objective is to minimize the system-level voltage deviation objective function of the distribution network nodes while ensuring the safety of the energy storage charge state of each agent, and at the same time, to minimize the grid oscillations caused by blind trial and error by multiple agents.

[0085] Step 2: After completing the experimental scenario and environment settings, initialize the training task and obtain the current operating situation and historical experience of the power distribution network system. Before each new training round starts, trigger the game debate mechanism of multiple heterogeneous large language models to generate topologically differentiated physical control parameters for the agents of the continuous execution layer.

[0086] The specific process of step 2 is as follows:

[0087] Step 21, Generate independent proposals: Multiple large language models read historical trajectory features from the experience memory bank, and each language model formulates a strategy based on the differences in the sensitivity of the power grid spatial topology, and independently outputs a set of physical control parameters for each agent;

[0088] The physical control parameters in the physical control parameter set include those of the intelligent agent. Voltage droop gain Intelligent agents Target reference voltage Intelligent agents State of charge protection gain and intelligent agents static bias .

[0089] Step 22, Cross-examination and Revision: Multiple rounds of mutual review are conducted among the large language models;

[0090] In each round of mutual review, two large language models are selected, denoted as Model A and Model B respectively. Model A reviews Model B's proposal, points out the risk of collaborative failure that Model B may cause, and issues criticisms. After receiving the criticisms, Model B evaluates the physical rationality of the criticisms and modifies or provides physical and logical justifications for the initial proposal.

[0091] Step 23: Strategies are scored: The strategies developed by each language model are scored from four dimensions: physical rationality, spatial difference, debate quality, and model historical reputation. The strategy with the highest score is selected as the action configuration for this round, and the language model corresponding to the strategy with the highest score wins.

[0092] The formula for calculating the overall score is as follows:

[0093] ;

[0094] in, Representing a large language model Overall score; For large language models The score for physical plausibility; For large language models Spatial difference score; For large language models The debate quality score; Large language models recorded in the experience memory bank Historical reputation value; For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weight coefficients, and satisfying ;

[0095] The formula used to evaluate whether the physical control parameters generated by the large language model meet the safe operation boundary of the distribution network is:

[0096] ;

[0097] in, A collection of intelligent agents; For intelligent agents Voltage droop gain; For intelligent agents State-of-charge protection gain; This is the effectiveness index function for voltage droop gain; The margin index function for the state-of-charge protection gain; Configure physical penalty terms for parameters that may trigger energy storage exceeding limits; The physical evaluation weighting coefficients for the corresponding effectiveness indicators; For the physical evaluation weighting coefficients corresponding to the sufficiency index; This refers to the physical evaluation weighting coefficient corresponding to the physical penalty item;

[0098] The calculation formula is:

[0099] ;

[0100] in, The total number of agents; For large language models For intelligent agents The generated physical control parameter vector; For large language models For intelligent agents The generated physical control parameter vector; This is the spatial difference sensitivity adjustment coefficient; The Euclidean norm 2 of a vector;

[0101] The calculation formula is:

[0102] ;

[0103] in, For large language models The validity score indicates the vulnerabilities of other model proposals; For large language models The strategic gain component resulting from modifying the initial proposal after incorporating valid criticism; For the corresponding Debate scoring coefficient; For the corresponding The debate scoring coefficient.

[0104] Step 3: Transform the generated physical control parameters into action guidance signals that the agent can understand with zero latency during millisecond-level execution.

[0105] The specific process of step 3 is as follows:

[0106] Step 31, Physical Reference Motion Generation: At each time step in the underlying layer Physical prior bridging layer acquires intelligent agents The node at time step Real-time voltage With intelligent agents The node at time step State of charge ;

[0107] Using the physical control parameters output by the winning large language model, continuous recommended actions are calculated in real time through the voltage droop control equation in the physical domain:

[0108] ;

[0109] in, Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function, used to force the calculated action to be limited to the upper and lower limits allowed by the agent's hardware. Within the range;

[0110] Step 32, Enhance observation space stitching: Calculate the... As a semantic cue vector, it is directly concatenated into the agent's augmented state space to construct an augmented observation state space, thereby achieving zero-delay injection of prior guidance;

[0111] ;

[0112] in, Represents intelligent agents At time step Enhanced observation state space; Represents intelligent agents At time step An enhanced state space; This indicates a vector concatenation operation.

[0113] Step 4: Multiple agents in the continuous execution layer interact, and the agents are guided to explore through dynamically decaying rewards without changing the basic optimization goal.

[0114] The specific process of step 4 is as follows:

[0115] Step 41, Calculate the dynamic guided reward: To avoid reinforcement learning falling into rigid and inflexible rule-based suboptimal solutions, the agent... Output Then, calculate and To ensure consistency, a guiding reward that monotonically decays with the training process is applied.

[0116] ;

[0117] in: Represents intelligent agents At time step The obtained dynamic guidance reward is used to guide the agent to gravitate towards the physical recommendation actions of the large language model; For intelligent agents At time step Continuous action output; Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function; The basic weighting coefficient representing the guiding reward determines the initial intensity of physical prior intervention on reinforcement learning exploratory behavior; This represents a smooth decay factor that monotonically decreases with the overall training progress;

[0118] Step 42, Composite-driven smooth transition: Calculate the total single-step reward that the agent ultimately uses for gradient updates;

[0119] ;

[0120] in, Represents intelligent agents At time step The final comprehensive reward value used for network gradient updates; Indicates at time step Basic physical penalties for feedback in power distribution network simulation environments;

[0121] As the training progresses... As the dynamic rewards gradually decrease and the control is smoothly transferred to the distribution network itself, the intelligent agent can autonomously explore extreme collaborative strategies that surpass large language models on the basis of safety benchmarks.

[0122] Step 5: After the current training cycle ends, extract the multi-dimensional time series measurement data of the continuous execution layer, trigger the human-like debriefing and reflection mechanism of the large language model, extract the cause and effect of success and failure, and form a closed-loop reflection chain.

[0123] The specific process of step 5 is as follows:

[0124] Step 51, Multi-dimensional temporal profile generation and reputation update: Collect the global temporal trajectory of the underlying control in this round, transform the global temporal trajectory into a natural language structured profile readable by the large language model, and update the historical reputation values ​​of the various language models participating in the debate based on the actual voltage control performance.

[0125] Step 52, In-depth review and cross-distillation based on the perspective of victory and defeat: triggering a five-step closed-loop reflection chain to extract physical laws;

[0126] The five-step closed-loop reflection chain is as follows:

[0127] Step 1, in-depth review of the winner: Combine the winning big language model with the timing profile to reflect on the physical context in which the voltage droop gain played a key role.

[0128] Step 2, Loser's Attribution Reflection: The defeated big language model compares its initial proposal with the actual power grid trajectory and reflects on its blind spots in physical cognition;

[0129] Step 3, cross-model teaching: The winning big language model provides customized improvement suggestions based on the reflection results of the losing big language model;

[0130] Step 4, defensive feedback: The defeated big language model points out the potential hidden dangers or overfitting risks in the strategies formulated by the winning big language model;

[0131] Step 5, Consensus Law Synthesis: The conclusions of the multidimensional reflections in Steps 1 to 4 above are synthesized and condensed into bilateral shared principles and specific lessons for improvement.

[0132] Step 53, Reflection and Memory Consolidation and Cross-Round Feedback: The physical laws extracted from the review and reflection are stored in a persistent experience memory bank. When starting Step 2 of the next round, each language model is required to prioritize reading its own specific lessons learned from failures, using these lessons as mandatory constraints for generating new proposals. This mechanism completely eliminates the repeated trial and error of large language models under similar physical conditions, enabling the perpetual evolution of power distribution network collaborative control strategies across models and scenarios.

[0133] The method of this invention was compared and tested with traditional multi-agent reinforcement learning algorithms, and the results are as follows: Figure 2 As shown, Figure 2 The vertical axis represents the voltage deviation, and the error bars reflect the fluctuation range (i.e., stability) of the control effect. Figure 2 Based on the baseline (i.e., the traditional multi-agent reinforcement learning algorithm without the introduction of a large language model) and the results of the first, second, and third rounds of the method of this invention, the following conclusions can be drawn:

[0134] 1. Control accuracy is significantly improved round by round: The baseline method has the highest average voltage deviation (approximately 0.17) and low exploration efficiency. After introducing the semantic evolution guidance of the large language model of this invention, the voltage deviation in the first round is significantly lower than the baseline level; as training progresses, the voltage deviation in the second and third rounds shows a step-like continuous downward trend, and the average voltage deviation has dropped to the lowest level (approximately 0.145) by the third round.

[0135] 2. Enhanced Strategy Stability and Robustness: Observations Figure 2 As can be seen from the error bars at the top, the baseline method fluctuates greatly when facing complex power grid conditions, which can easily cause power grid oscillations. However, with the gradual intervention of the retrospective reflection mechanism of this invention, the distribution range of the error bars shows a gradual convergence trend, indicating that the cooperative conflict between multiple agents is effectively eliminated and the control action is more stable and safe.

Claims

1. The voltage control method for distribution networks based on large language models is implemented based on a three-layer spatial architecture, which includes a multi-model cognition and game theory layer at the top, a physical prior bridging layer in the middle, and a continuous execution layer at the bottom. In the multi-model cognition and game layer, multiple heterogeneous large language models are deployed and experience memory banks and their respective reputation values ​​are maintained. Each language model performs independent virtual scheduling, proposes strategies, debates with each other and performs knowledge distillation for the current power grid situation, and obtains the physical control parameters of the winner of the debate. In the physical prior bridging layer, a strategy evaluation and physical guidance module is deployed, which is responsible for converting the physical control parameters of the winner of the large language model debate into continuous action guidance signals; In the continuous execution layer, a power distribution network simulation environment and multiple intelligent agents are deployed. With the assistance of action guidance signals, the intelligent agents try and fail, and output continuous charging and discharging power. Its features are, Includes the following steps: Step 1: Experimental scenario and environment setup. Deploy multiple intelligent agents in the power distribution network system, using the intelligent agents as control objects and setting control objectives; Step 2: After completing the experimental scenario and environment settings, initialize the training task and obtain the current operating situation and historical experience of the power distribution network system. Before each new training round starts, trigger the game debate mechanism of multiple heterogeneous large language models to generate topologically differentiated physical control parameters for the agents of the continuous execution layer. Step 3: Convert the generated physical control parameters into action guidance signals that the agent can understand with zero latency during millisecond-level execution; Step 4: Multiple agents in the continuous execution layer interact, and the agents are guided to explore through dynamically decaying rewards without changing the basic optimization goal; Step 5: After the current training cycle ends, extract the multi-dimensional time series measurement data of the continuous execution layer, trigger the human-like debriefing and reflection mechanism of the large language model, extract the cause and effect of success and failure, and form a closed-loop reflection chain.

2. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 1, characterized in that, In step 1, the intelligent agent At time step The continuous action output is , After physical gain amplification, the corresponding charging and discharging power is [-4.0, 4.0] megawatts; The enhanced state space of an agent is defined as follows: ； in, Represents intelligent agents At time step An enhanced state space; This represents a vector concatenation operation; For intelligent agents At time step The actual activation prompt vector received; For intelligent agents The node at time step Real-time voltage; For intelligent agents The node at time step The state of charge; These are the normalized eigenvalues ​​for each time step; For time step The system's real-time load power; For time step The photovoltaic power generation capacity; For time step The wind power generation capacity; The control objective is to minimize the system-level voltage deviation objective function of the distribution network nodes while ensuring the safety of the energy storage charge state of each agent, and at the same time, to minimize the grid oscillations caused by blind trial and error by multiple agents.

3. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 2, characterized in that, The specific process of step 2 is as follows: Step 21, Generate independent proposals: Multiple large language models read historical trajectory features from the experience memory bank, and each language model formulates a strategy based on the differences in the sensitivity of the power grid spatial topology, and independently outputs a set of physical control parameters for each agent; Step 22, Cross-examination and Revision: Multiple rounds of mutual review are conducted among the large language models; Step 23: Strategies are scored: The strategies developed by each language model are scored from four dimensions: physical rationality, spatial difference, debate quality, and model historical reputation. The strategy with the highest score is selected as the action configuration for this round, and the language model corresponding to the strategy with the highest score wins.

4. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 3, characterized in that, In step 21, the physical control parameters in the physical control parameter set include the intelligent agent. Voltage droop gain Intelligent agents Target reference voltage Intelligent agents State of charge protection gain and intelligent agents static bias .

5. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 4, characterized in that, In step 22, in each round of mutual review, two large language models are selected, denoted as Model A and Model B respectively. Model A reviews the proposal of Model B, points out the risk of collaborative failure that Model B may cause, and issues criticism. After receiving the criticism, Model B evaluates the physical rationality of the criticism and modifies or physically and logically defends the initial proposal.

6. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 4, characterized in that, In step 23, the formula for calculating the comprehensive score is as follows: ； in, Representing a large language model Overall score; For large language models The score for physical plausibility; For large language models Spatial difference score; For large language models The debate quality score; Large language models recorded in the experience memory bank Historical reputation value; For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weighting coefficients, For the corresponding The weight coefficients, and satisfying ; The formula used to evaluate whether the physical control parameters generated by the large language model meet the safe operation boundary of the distribution network is: ； in, A collection of intelligent agents; For intelligent agents Voltage droop gain; For intelligent agents State-of-charge protection gain; This is the effectiveness index function for voltage droop gain; The margin index function for the state-of-charge protection gain; Configure physical penalty terms for parameters that may trigger energy storage exceeding limits; The physical evaluation weighting coefficients for the corresponding effectiveness indicators; For the physical evaluation weighting coefficients corresponding to the sufficiency index; This refers to the physical evaluation weighting coefficient corresponding to the physical penalty item; The calculation formula is: ； in, The total number of agents; For large language models For intelligent agents The generated physical control parameter vector; For large language models For intelligent agents The generated physical control parameter vector; This is the spatial difference sensitivity adjustment coefficient; The Euclidean norm 2 of a vector; The calculation formula is: ； in, For large language models The validity score indicates the vulnerabilities of other model proposals; For large language models The strategic gain component resulting from modifying the initial proposal after incorporating valid criticism; For the corresponding Debate scoring coefficient; For the corresponding The debate scoring coefficient.

7. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 6, characterized in that, The specific process of step 3 is as follows: Step 31, Physical Reference Motion Generation: At each time step in the underlying layer Physical prior bridging layer acquires intelligent agents The node at time step Real-time voltage With intelligent agents The node at time step State of charge ; Using the physical control parameters output by the winning large language model, continuous recommended actions are calculated in real time through the voltage droop control equation in the physical domain: ； in, Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function, used to force the calculated action to be limited to the upper and lower limits allowed by the agent's hardware. Within the range; Step 32, Enhance observation space stitching: Calculate the... As a semantic cue vector, it is directly concatenated into the agent's augmented state space to construct an augmented observation state space; ； in, Represents intelligent agents At time step Enhanced observation state space; Represents intelligent agents At time step An enhanced state space; This indicates a vector concatenation operation.

8. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 7, characterized in that, The specific process of step 4 is as follows: Step 41, Calculate the dynamic guided reward: agent Output Then, calculate and To ensure consistency, a guiding reward that monotonically decays with the training process is applied. ； in: Represents intelligent agents At time step The dynamic guided reward obtained; For intelligent agents At time step Continuous action output; Indicates the physical prior bridging layer at time step For intelligent agents Calculate the generated consecutive recommended action values; This represents the amplitude truncation function; This represents the basic weighting coefficient for the guiding reward; This represents a smooth decay factor that monotonically decreases with the overall training progress; Step 42, Composite-driven smooth transition: Calculate the total single-step reward that the agent ultimately uses for gradient updates; ； in, Represents intelligent agents At time step The final comprehensive reward value used for network gradient updates; Indicates at time step Basic physical penalties for feedback in power distribution network simulation environments; As the training progresses, As the dynamic guidance rewards gradually decrease and the control is smoothly transferred to the distribution network itself, the intelligent agent can autonomously explore an ultimate collaborative strategy that surpasses the large language model on the basis of safety benchmarks.

9. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 8, characterized in that, The specific process of step 5 is as follows: Step 51, Multi-dimensional temporal profile generation and reputation update: Collect the global temporal trajectory of the underlying control in this round, transform the global temporal trajectory into a natural language structured profile readable by the large language model, and update the historical reputation values ​​of the various language models participating in the debate based on the actual voltage control performance. Step 52, In-depth review and cross-distillation based on the perspective of victory and defeat: triggering a five-step closed-loop reflection chain to extract physical laws; Step 53, Reflection Memory Consolidation and Cross-Round Feedback: Store the physical laws extracted from the review and reflection into a persistent experience memory bank. When starting Step 2 of the next round, force all major language models to prioritize reading their own exclusive failure lessons, and use the failure lessons as a mandatory constraint boundary for generating new proposals.

10. The method for multi-agent voltage control of distribution networks based on a large language model as described in claim 9, characterized in that, In step 52, the five-step closed-loop reflection chain is as follows: Step 1, in-depth review of the winner: Combine the winning big language model with the timing profile to reflect on the physical context in which the voltage droop gain played a key role. Step 2, Loser's Attribution Reflection: The defeated big language model compares its initial proposal with the actual power grid trajectory and reflects on its blind spots in physical cognition; Step 3, cross-model teaching: The winning big language model provides customized improvement suggestions based on the reflection results of the losing big language model; Step 4, defensive feedback: The defeated big language model points out the potential hidden dangers or overfitting risks in the strategies formulated by the winning big language model; Step 5, Consensus Law Synthesis: The conclusions of the multidimensional reflections in Steps 1 to 4 above are synthesized and condensed into bilateral shared principles and specific lessons for improvement.

Images