Power distribution network intelligent dispatching method based on large model meta-decision and multi-small model cooperation

Abstract

The present application relates to the technical field of power distribution network intelligent scheduling, in particular to a power distribution network intelligent scheduling method based on large model meta-decision and multi-small model cooperation. The method acquires real-time multi-source heterogeneous data of the power distribution network and performs risk perception, converts the structured prompt words into input of a general large model fine-tuned in the field, and queries the historical similar operation scenarios based on the retrieval enhancement generation technology. According to the scene matching degree, the historical scheduling scheme is directly called, or a deep reinforcement learning small model and a deep learning small model are selected to generate a scheduling strategy. The scheduling strategy is subjected to safety verification, and if it fails, the optimal power flow small model is called to solve again. The execution result is recorded to the historical scheme library and the small model training sample, the model decision confidence is dynamically corrected, and the large model incremental fine-tuning is triggered. The present application realizes the cooperation and self-evolution of large and small models, dynamically selects the adaptive solving method according to the scene characteristics, and improves the reliability of the scheduling strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent dispatching technology for distribution networks, specifically to an intelligent dispatching method for distribution networks based on large-scale model meta-decision and multi-small-scale model collaboration. Background Technology

[0002] Intelligent dispatching of distribution networks is a key technology for ensuring the safe operation of distribution networks. Existing distribution network dispatching technologies mainly include knowledge-driven historical scheme database query methods, data-driven deep learning or deep reinforcement learning methods, and model-driven mathematical optimization methods. Knowledge-driven historical scheme database query methods accumulate historical dispatching cases, match similar historical scenarios to the current operating scenario according to expert rules, and then invoke the corresponding dispatching scheme. Data-driven deep learning methods generate dispatching strategies by constructing end-to-end data mapping relationships; deep reinforcement learning methods learn to generate dispatching strategies through continuous interaction between the agent and the distribution network environment. Model-driven mathematical optimization methods establish accurate mathematical models of the distribution network and solve for the optimal dispatching strategy that satisfies various constraints.

[0003] In recent years, general-purpose large model technologies, such as Deepseek and Tongyi Qianwen, have flourished, demonstrating significant advantages in natural language understanding and knowledge retrieval. In the power sector, general-purpose large models have been tested for querying scheduling procedures.

[0004] The aforementioned existing technologies have the following drawbacks: First, knowledge-driven historical solution database query methods typically rely on simple keywords or numerical thresholds for similarity judgment, making it difficult to accurately identify complex similar scenarios at the semantic level, and failing to effectively utilize the query capabilities of general-purpose large models for intelligent retrieval of historical scheduling cases. Second, data-driven deep learning methods generate scheduling strategies through end-to-end data mapping relationships, while deep reinforcement learning methods learn scheduling strategies through continuous interaction between the agent and the distribution network environment. Both are highly dependent on the coverage and quality of training samples, resulting in insufficient generalization ability in scenarios with sparse or out-of-distribution samples. Third, model-driven mathematical optimization methods solve for optimal scheduling strategies that satisfy various constraints by establishing accurate mathematical models of the distribution network. Although this can guarantee the optimality of the solution, the complexity of solving the mixed-integer nonlinear programming problem involving high-dimensional discrete switching actions and continuous source-storage-output coupling in the distribution network increases sharply with the scale of the scenario, making it difficult to directly apply to scheduling scenarios of large-scale complex distribution networks. Fourth, existing technologies have failed to establish an effective collaborative mechanism between large and small models. While general-purpose large models possess powerful natural language understanding and knowledge retrieval capabilities, they have not been organically integrated with smaller models such as deep learning models, deep reinforcement learning models, and mathematical optimization models in current technologies, making dynamic selection based on scenario characteristics impossible. Fifth, existing scheduling methods lack a self-evolution mechanism. The decision confidence of each model and the scheduling strategy cannot be dynamically corrected and continuously optimized based on actual execution results. The updating of historical solution libraries and model training samples relies on manual screening and labeling, making it difficult to achieve autonomous and continuous model evolution.

[0005] In summary, existing intelligent dispatching technologies for distribution networks fail to dynamically select solution methods based on the characteristics of different operating scenarios. When faced with operating scenarios of varying complexity, they cannot adaptively select and query historical solution libraries or invoke deep learning or deep reinforcement learning mini-models to solve dispatching strategies. This results in inconsistent strategy quality generated by a single method in unsuitable scenarios. Furthermore, the models cannot autonomously and continuously evolve with changes in the distribution network operating environment. Therefore, there is an urgent need for a distribution network dispatching method that enables collaboration and self-evolution between large and small models. This method should organically integrate the knowledge query capabilities of the large model, the solution capabilities of the small model, and the experiential knowledge from the historical solution library. It should dynamically select the most suitable solution method based on scenario characteristics and continuously optimize each model based on execution performance, thereby improving the reliability of the dispatching strategy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a smart dispatching method for distribution networks based on large-scale model meta-decision-making and multi-small-scale model collaboration, aiming to solve the problems in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a smart dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration, comprising the following steps: Step S1: Obtain multi-source heterogeneous data of real-time operation of the distribution network. The multi-source heterogeneous data includes operation measurement data and alarm signals. Risk perception is performed based on the operation measurement data. Voltage over-limit, line overload and fault information are recorded based on the alarm signals and converted into structured prompt words. Step S2: Input the structured prompts into the domain-fine-tuned general large model for meta-decision. Based on the retrieval enhancement generation technology, query the historical similar running scenarios in the historical scheme library. Select one of the deep reinforcement learning small model or the deep learning small model according to the scenario matching degree, or directly call the corresponding historical scheduling scheme. The deep reinforcement learning small model adopts the soft actor-critic algorithm, and the deep learning small model adopts the graph attention network. Step S3: If a deep reinforcement learning small model or a deep learning small model is selected in step S2, a scheduling strategy containing a switch opening and closing instruction sequence and source and storage output values ​​is generated by the selected deep reinforcement learning small model or deep learning small model. The generated scheduling strategy is then subjected to a safety verification. If the safety verification passes, the strategy is executed. If a historical scheduling scheme is called in step S2, the historical scheduling scheme is then subjected to a safety verification. If the safety verification passes, the strategy is executed. Step S4: If the scheduling strategy or historical scheduling scheme in step S3 fails the security check, the domain-fine-tuned general large model calls the optimal power flow small model to resolve and execute the scheduling strategy; the scheduling strategy and execution results are recorded in the historical scheme library, and the scheduling strategy and execution results are used for incremental training of the deep reinforcement learning small model and the deep learning small model, and the decision confidence of the deep reinforcement learning small model and the deep learning small model is corrected based on the execution results, triggering the incremental fine-tuning of the domain-fine-tuned general large model.

[0008] Furthermore, in step S1, the operational measurement data includes the voltage amplitude of each node in the distribution network, the current amplitude of each line, the active power output of each distributed power source, the state of charge of each energy storage system, and the active power of each load; after removing bad data from the operational measurement data, the remaining valid operational measurement data is normalized.

[0009] Furthermore, in step S1, the risk perception includes node voltage over-limit risk perception and line current carrying capacity over-limit risk perception; based on the normalized operational measurement data, the voltage over-limit risk index of each node and the current carrying capacity over-limit risk index of each line are calculated respectively.

[0010] Furthermore, in step S2, the scene matching degree is divided into three levels: complete matching, partial matching, and no matching. When the scene semantic similarity reaches the complete matching similarity threshold, and the current scene and the historical most similar scene are semantically consistent in key elements such as fault location, over-limit equipment, topology, source-storage output, and risk level, it is determined to be a complete match, and the corresponding historical scheduling scheme is directly invoked. When the scene semantic similarity is between the partial matching similarity threshold and the complete matching similarity threshold, it is determined to be a partial match. When the scene semantic similarity is lower than the partial matching similarity threshold, it is determined to be a no match. In the case of partial matching or no matching, the general large model, which has been fine-tuned by the domain, performs meta-decision based on the decision confidence of each small model, and selects a deep reinforcement learning small model or a deep learning small model.

[0011] Furthermore, in step S3, when a deep reinforcement learning small model is selected, a soft actor-critic algorithm is used to generate a scheduling strategy. The soft actor-critic algorithm includes an actor network, a critic network, and a target critic network. The actor network takes the current operating state of the distribution network as input and outputs a scheduling action probability distribution containing switch opening and closing commands and source-storage-output setting values. The critic network takes state-action pairs as input and outputs an action value estimate. The parameters of the target critic network are updated by tracking the parameters of the critic network.

[0012] Furthermore, in step S3, when a deep learning small model is selected, a graph attention network is used to generate a scheduling strategy; the distribution network topology is used as a graph structure, the running measurement data is used as node features, the spatial coupling features are extracted through the graph attention layer, and mapped to the switch opening and closing command sequence and the source and storage power output values ​​through the fully connected output layer.

[0013] Furthermore, in step S3, the safety verification is based on power flow calculation of the distribution network, and the verification content includes power flow balance constraints, node voltage constraints, line current carrying capacity constraints, topology constraints, and equipment capacity constraints; if all constraints are satisfied, the safety verification is deemed to pass; otherwise, the safety verification is deemed to fail.

[0014] Furthermore, in step S4, the optimal power flow small model models the scheduling strategy solution problem as a mixed integer nonlinear programming problem; the decision variables include the discrete on / off states of controllable switches and the continuous active power output of distributed power sources and energy storage systems; the objective function is the weighted minimization of distribution network active power loss, switch action cost and load shedding; the constraints include power flow balance equations, node voltage constraints, line current carrying capacity constraints, topology constraints and equipment capacity constraints.

[0015] Furthermore, in step S4, after the scheduling strategy is executed, the execution effect score is calculated based on the actual voltage amplitude of each node in the distribution network, the actual current amplitude of each line, the actual output of each distributed power source and energy storage system, and the actual status of the switches. The current structured prompt word is input into the general large model that has been fine-tuned by the domain, and the query semantic vector of the current scenario is obtained through embedding encoding. The current structured prompt word, scheduling strategy, execution effect score, and query semantic vector of the current scenario are used to construct a new historical scenario record and add it to the historical solution library. The similarity between the query semantic vector of the current scenario and the query semantic vector of the existing historical scenarios in the library is calculated. If there is a historical scenario with a similarity greater than the merging similarity threshold, the execution effect scores of the two are compared. The original historical scenario record is replaced only when the execution effect of the current scenario is better; otherwise, the original record is retained. If there is no existing historical scenario that meets the conditions, a new scenario record is directly added.

[0016] Furthermore, in step S4, the correction of the decision confidence of the deep reinforcement learning small model and the deep learning small model includes: if the small model selected and executed by the current meta-decision mechanism is a deep reinforcement learning small model or a deep learning small model, its decision confidence is updated by weighting according to the execution effect score; for deep reinforcement learning small models or deep learning small models that are not selected by the meta-decision mechanism in adjacent scheduling cycles, their decision confidence is decayed according to the decision confidence decay coefficient; when the number of newly added scene records in the historical solution library reaches the preset incremental fine-tuning sample threshold, or when the decision confidence of the deep reinforcement learning small model is lower than the decision confidence lower limit for multiple consecutive cycles, or when the decision confidence of the deep learning small model is lower than the decision confidence lower limit for multiple consecutive cycles, the incremental fine-tuning of the general large model that has been fine-tuned by the domain is triggered; the incremental fine-tuning adopts a fine-tuning method based on low-rank adaptation, freezes the base parameters of the general large model, and trains only the low-rank adaptation parameter matrix.

[0017] Compared with existing technologies, the present invention has the following advantages: By constructing a meta-decision mechanism, the knowledge query capability of the domain-fine-tuned general large model, the solution capability of the deep reinforcement learning small model and the deep learning small model, and the experience knowledge of the historical solution library are organically integrated. Based on the decision confidence of each model under the current operating scenario and the similarity with the most similar historical scenario, the most suitable solution method is dynamically selected, which overcomes the shortcomings of the existing distribution network scheduling method where a single solution method is difficult to adapt to differentiated operating scenarios. When the safety check fails, the general large model calls the optimal power flow small model to solve the problem again, forming a collaborative mechanism of small model solution and mathematical optimization as a fallback, which ensures the safety and feasibility of the scheduling strategy. Through a closed-loop mechanism involving strategy execution, historical solution library updates, incremental updates of training samples, dynamic correction of model decision confidence, and incremental fine-tuning of the general large model, the collaboration and self-evolution of small and large models are achieved. The solution capabilities of each small model are continuously enhanced by adding new training samples, the meta-decision capability of the general large model is continuously optimized through incremental fine-tuning and expansion of the historical solution library, the historical solution library is kept refined and efficient through a deduplication and merging mechanism, and the model decision confidence dynamically reflects the actual reliability of each model in the current operating environment through execution feedback. Thus, the reliability of dispatch strategy generation is continuously improved under the condition of continuous changes in the distribution network operating environment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the overall framework of the present invention. Detailed Implementation

[0020] like Figure 1 , Figure 2 As shown, the present invention provides a technical solution: a smart dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration, comprising the following steps: Step S1: Acquire and parse multi-source heterogeneous data of real-time operation of the distribution network, including operation measurement data and alarm signals. Perform risk perception based on operation measurement data, record voltage over-limit, line overload and fault information based on alarm signals, and convert them into structured prompt words.

[0021] In this embodiment, step S1 specifically includes: The distribution network dispatching master station collects multi-source heterogeneous data on the real-time operation of the distribution network through the SCADA system at a set period, denoted as: ; In the formula, For a moment A collection of heterogeneous data from multiple sources; For a moment The set of operational measurement data; For a moment The set of alarm signals.

[0022] The time operational measurement data set Including time Distribution network nodes voltage amplitude Each route Current amplitude The active power output of each distributed power source State of charge of each energy storage system and the active power of each load. ; After removing bad data caused by communication anomalies from the operational measurement data, the remaining valid operational measurement data is normalized: ; In the formula, For the first Each operational measurement data at time... The original value; For the first Each operational measurement data at time... The normalized measurement value; For the first The historical maximum value of a single operational measurement data point or the rated upper limit of the equipment; For the first The historical minimum value of the operational measurement data or the lower limit of the equipment's rated value.

[0023] The set of alarm signals The binary signal set sent from relay protection devices and distribution automation terminals (FTU / DTU) to the dispatch master station is divided into five categories according to the physical meaning of the signals: ; In the formula, For a moment The set of protection action signals; For a moment The set of switch position change signals; For a moment The set of overcurrent alarm signals; For a moment The set of overvoltage alarm signals; For a moment A set of low voltage alarm signals.

[0024] In this embodiment, risk perception is performed based on normalized operational measurement data, including node voltage over-limit risk perception and line current carrying capacity over-limit risk perception. To address the risk of voltage exceeding limits at distribution network nodes, define nodes. The voltage over-limit risk indicator is: ; In the formula, For nodes At any moment The normalized voltage amplitude; For nodes Permissible upper voltage limit; For nodes Permissible lower voltage limit; For nodes At any moment The voltage over-limit risk indicator, and The closer it is to 1, the higher the risk level of the voltage limit. To address the risk of line overload, define the line. The overload risk indicator is: ; In the formula, For the line At any moment The normalized current amplitude; For the line Long-term allowable carrying capacity; For the line At any moment The risk indicator of exceeding the carrying capacity limit, and The closer it is to 1, the greater the limitation on the line's capacity and the higher the risk level.

[0025] In this embodiment, alarm signals are used to record information on voltage over-limit, line overload, and faults. time Set of voltage over-limit and line overload information Build as follows: ; In the formula, For the line Overload; For nodes The voltage exceeds the upper limit; For nodes The voltage is below the lower limit; time The following fault information set Build as follows: ; In the formula, For the line A protective action was initiated; For switch When a change occurs, the fault information set records the device number and status of all devices that have performed protection actions or switch changes.

[0026] Risk perceptions, limit violations, or malfunctions are transformed into structured prompts and assembled according to a fixed template. The construction formula is as follows: ; In the formula, For a moment Structured prompts below; For prompt word template mapping function; For a moment A set of voltage over-limit risk indicators for all nodes. , For distribution network nodes; For a moment A set of risk indicators for overload capacity exceeding the limit for all lines. , A collection of distribution network lines; In this embodiment, the template form of the structured prompt words is as follows: ; In the formula, String concatenation operator; This section describes the risk perception results, including the node numbers and risk values ​​of the voltage over-limit risk index being greater than zero, and the line numbers and risk values ​​of the current carrying capacity over-limit risk index being greater than zero. This is the description section for exceeding limits and fault markings, which includes the device number and type of the device that exceeded the limit marked in the alarm signal, as well as the location and type of the fault.

[0027] Step S2: Input the structured prompts into the general large model that has been fine-tuned in the domain for meta-decision. Based on the retrieval enhancement generation technology, query the historical similar running scenarios. Select a deep reinforcement learning small model or a deep learning small model according to the scenario matching degree, or directly call the corresponding historical scheduling scheme.

[0028] Construct a meta-decision mechanism: Input the structured prompts mentioned in step S1 into a general large model that has been fine-tuned by the domain. The general large model queries historical similar running scenarios based on retrieval enhancement generation technology, selects one of the deep reinforcement learning small model and the deep learning small model according to the scenario matching degree, and outputs the selected model and the selection basis; or directly calls the historical scheduling scheme.

[0029] In this embodiment, the domain-fine-tuned general large model is based on a general pre-trained large language model such as Deepseek or Tongyiqianwen, and is incrementally fine-tuned using a distribution network dispatching domain corpus. The domain corpus includes distribution network dispatching procedure text, historical dispatching operation logs, expert experience knowledge text, and vectors labeled with historical scenarios. Corresponding historical scheduling scheme The set of scenarios and decision pairs ,in This represents the total number of historical scenarios in the historical scenario library; the fine-tuning goal is to enable the general-purpose model to grasp the semantic relationships between risk perception indicators, alarm signals, and dispatch strategies of the distribution network, and to be able to respond to structured prompts. The ability to understand medium-risk levels, fault types, and topology status supports subsequent retrieval enhancement generation and meta-decision.

[0030] In this embodiment, the retrieval enhancement generation mechanism is based on a historical solution database. The implementation includes a historical solution library containing records of several historical scenarios. In the middle, the first A historical scene record Including the Historical structured prompts , No. Historical scheduling schemes Passing the exam Historical execution effect markers The construction method of the historical structured prompt words is the same as the structured prompt words in step S1. Consistent, containing a set of voltage over-limit risk indicators for that historical moment. Set of indicators for overload risk Fault marker set and the set of out-of-bounds markers ; The current structured prompt words are input into a domain-fine-tuned general-purpose model, and the time step is obtained through embedding encoding. The semantic vector of the query scenario is used to perform vector similarity retrieval in the historical solution database. The semantic similarity between the current scenario and each historical scenario is calculated, and the historical scenario with the highest similarity is selected as the retrieval result.

[0031] In this embodiment, based on the most similar historical scenes returned by the retrieval enhancement generation mechanism, the scene matching degree is determined through semantic understanding. It represents the degree of matching between the current scene and the most similar historical scene, and is divided into three levels: complete match, partial match, and no match. Define time The semantic similarity between the current scene and the most similar historical scene is: ; The scene matching degree is determined by the large model based on a combination of semantic similarity and domain knowledge: ; In the formula, For a moment The scene matching degree, with a value of ; This is a semantic understanding judgment function for a general large model that has been fine-tuned in the domain. For a moment Structured prompts below; For a moment The semantic similarity between the current scene and the most similar historical scene; Structured prompts corresponding to the most similar historical scenarios; It is a set of semantic understanding rules, including semantic rules for distribution networks such as fault location consistency, over-limit equipment consistency, topological equivalence, source-storage-output similarity, and risk level similarity.

[0032] In this embodiment, when Furthermore, the large model is based on a semantic understanding rule set. When the current scenario and the most similar historical scenario are semantically consistent in key elements such as fault location, out-of-limit equipment, topology, source and storage capacity, and risk level, it is determined to be a complete match. When, it is determined to be a partial match, when When the condition is met, it is determined to be a mismatch; among which To achieve a perfect match, a similarity threshold is required. This is the threshold for partial matching similarity.

[0033] If time Scene matching degree If the value is "exact match", the corresponding historical scheduling scheme is directly invoked. The conditions for invoking the historical scheduling scheme are: the historical scenario and the current scenario are semantically equivalent in key elements such as fault location, out-of-limit equipment, topology, source-storage output, and risk level, and the historical scheme has passed security verification and been successfully executed in the past. The invoked historical scheduling scheme is denoted as […]. It includes the historical switch opening and closing command sequence and the source storage output value.

[0034] If time Scene matching degree If the value is "partial match" or "no match", the domain-adjusted general large model makes a meta-decision based on the decision confidence of each model in the current operating scenario and the similarity between the current operating scenario and the most similar historical scenario, and selects either a deep reinforcement learning small model or a deep learning small model through semantic understanding; the model decision confidence includes the decision confidence of the deep reinforcement learning small model. Decision confidence of deep learning small models The result is obtained by dynamically correcting the historical execution results of each model in the subsequent step S4.

[0035] time Scene matching degree In the case of "partial match" or "no match", the formal expression of the meta-decision mechanism is as follows: ; In the formula, For a moment The selected model identifier in the meta-decision output. , For deep reinforcement learning small models, For deep learning small models; For the meta-decision function of the general large model that has been fine-tuned in the domain; For a moment Structured prompts below; For a moment Scene matching degree; For a moment The semantic similarity between the current scene and the most similar historical scene; For a moment The decision confidence of a small deep reinforcement learning model; For a moment The decision confidence of a small deep learning model.

[0036] In this embodiment, the selection criteria A general-purpose model, fine-tuned for the specific domain, is generated in natural language and used as the basis for human review. Specifically, this includes: The retrieval enhancement generation mechanism returns the historical most similar scene ID and the semantic similarity corresponding to that historical most similar scene. Scene matching degree The decision level (complete match, partial match, or no match) and the reasoning for the decision, and the decision confidence of the deep reinforcement learning small model. Decision confidence of deep learning small models The comparison relationship; the judgment reasons include semantic comparison conclusions between the current scenario and the historical scenario on key elements such as consistency of fault location, consistency of out-of-limit equipment, equivalence of topological structure, similarity of source and storage output, and similarity of risk level.

[0037] Step S3: If a deep reinforcement learning small model or a deep learning small model is selected, a scheduling strategy containing a switch opening and closing instruction sequence and source and storage output values ​​is generated by the selected deep reinforcement learning small model or deep learning small model. The generated scheduling strategy is then subjected to a safety check, and executed after the safety check passes. If a historical scheduling scheme is called, the historical scheduling scheme is then subjected to a safety check, and executed after the safety check passes.

[0038] In this embodiment, the scheduling strategy is denoted as... Includes switch opening and closing command sequence With source and storage output value .

[0039] In this embodiment, regarding policy generation for the deep reinforcement learning small model, when the meta-decision mechanism selects the deep reinforcement learning small model, the SoftActor-Critic (SAC) algorithm is used to generate the scheduling policy. This algorithm is an actor-critic architecture, containing two core networks: an actor network and a critic network, with an additional target critic network used to stabilize the training process. The actor network uses a policy network, which is a neural network structure. It takes the current operating state of the distribution network as input and outputs a probability distribution of scheduling actions including switch opening and closing commands and source / storage output settings, responsible for generating the scheduling policy. The critic network also uses a neural network structure, taking state-action pairs as input and outputting the Q-value estimate (i.e., action value estimate) of the state-action pair, used to evaluate the quality of actions generated by the policy network. The target critic network has the same network structure as the critic network, and its parameters are slowly updated by tracking the critic network parameters. It is used to calculate the target Q-value to avoid frequent fluctuations in the target value during training, thereby stabilizing the training process. The three networks iteratively optimize through continuous interaction with the distribution network operating environment, jointly learning the optimal scheduling policy. To train the three networks mentioned above, the distribution network scheduling problem is modeled as a Markov decision process, and the state space, action space, and reward function are defined as follows: The state space consists of the real-time operating state of the distribution network: ; In the formula, For a moment The state; For at any time The set of voltage amplitudes for all nodes below; For a moment The set of current amplitude values ​​for all lines below; For a moment The active power output of all distributed power sources; For a moment The set of states of charge of all energy storage systems; For a moment The collection of active power of all loads; For a moment The current open / closed state set of all connecting switches and sectionalizing switches; The action space consists of discrete switch actions and continuous source output values: ; In the formula, For a moment The action; For a moment The following is a sequence of switch opening and closing instructions, where a value of 0 indicates that the switch is open and a value of 1 indicates that the switch is closed; For a moment The charging and discharging power settings for the energy storage system are shown below; a positive value indicates discharging, and a negative value indicates charging. This indicates the upper limit of the rated charging and discharging power of the energy storage system; For a moment The active power output setting value of the distributed power source. This indicates the upper limit of the rated active power output of the distributed power source; , , These refer to the number of switches, energy storage systems, and distributed power sources, respectively. reward function Taking into account voltage quality, line current carrying capacity, distribution network losses, operational costs, power generation and storage costs, and load restoration effectiveness, it can be expressed as: ; In the formula, For a moment The agent in the state Next action The instant reward obtained afterward; The voltage over-limit penalty component represents the state of the agent. Next action The penalty amount received after that is due to exceeding the voltage limit; The penalty component for exceeding the line capacity limit represents the state of the agent. Next action The penalty amount received afterward for exceeding the line's capacity limit; The active power loss penalty component of the distribution network represents the state of the agent. Next action The penalty amount incurred due to active power losses in the distribution network; The penalty component represents the number of switching actions, indicating the agent's state. Next action The penalty amount received after the switch operation; The cost component of the source storage represents the state of the intelligent agent. Next action Subsequently, the amount of penalty imposed due to the operating costs incurred by the source and storage capacity; The reward component for load recovery represents the agent's state. Next action The bonus amount obtained after the load is restored to power; The weighting coefficient for the voltage over-limit penalty component; The weighting coefficient for the line's overload penalty component; The weighting coefficient for the active power loss penalty component of the distribution network; The weighting coefficient for the penalty component based on the number of switching actions; The weighting coefficient for the source-storage output cost component; The weighting coefficient for the load recovery reward component; Voltage over-limit penalty component , represented as: ; In the formula, For distribution network nodes; For a moment Distribution network nodes The voltage amplitude; and Distribution network nodes Permissible upper and lower voltage limits; Line overload penalty , represented as: ; In the formula, For a moment line The current amplitude; For the line Long-term allowable carrying capacity; A collection of distribution network lines; Active power loss penalty component of distribution network , represented as: ; In the formula, For the line The resistance; Penalty component for number of switching actions , represented as: ; In the formula, For a moment No. The open / closed state of a controllable switch is 0, which indicates that it is open, and 1 indicates that it is closed. For the previous moment The on / off state of the switch at any given time; A set of controllable switches; Source storage output cost component , represented as: ; In the formula, For a moment No. The active power output of a distributed power source; For the first Unit output cost coefficient of a distributed power source; For a moment No. The charging and discharging power of an energy storage system (discharging is positive, charging is negative). For the first Unit charge / discharge cost coefficient of an energy storage system; , These refer to the number of distributed power sources and energy storage systems, respectively. Load recovery bonus , represented as: ; In the formula, For distribution network nodes; For nodes The active power of the loads that have had their power restored; For nodes Active power demand at the load; The soft actor-critic algorithm trains a policy network by maximizing a weighted sum of expected cumulative reward and policy entropy. Output actions using Gaussian distribution reparameterization: ; In the formula, For action; State; For a moment The action; For policy networks, time-based status The average value of the input and output actions; For policy networks, time-based status The standard deviation of the input and output; This is a noise vector sampled from a standard normal distribution; This is element-wise multiplication; It is a vector with zero mean and covariance matrix as the identity matrix. The multidimensional normal distribution; The commentator network updates by minimizing the Bellman mean square error: ; In the formula, For critic network parameters The loss function; To calculate the expected value of a state-action-reward-next-state quadruple randomly sampled from the experience replay buffer, This is the mathematical expectation operator, used to take the average of the random variables within the parentheses. For experience replay buffer Mid-sampled state-action-reward-next-state quadruple; This serves as a buffer for experience replay. For a moment The instant reward received; For the commentator network on the current state-action pair Q-value estimation, where Q is the action value function; Discount factor; For policy networks The next action of mid-sampling Take the expected value; For the target commentator network on the next state-action pair Q-value estimation; Temperature is a parameter used to balance the exploration and utilization of control strategies. For policy networks in state of time Down Output Action The logarithmic probability, It is the natural logarithm function; In a given state Under the condition of policy network output action The conditional probability; The policy network updates by maximizing a weighted sum of expected cumulative return and policy entropy: ; In the formula, For policy network parameters The loss function; For the experience replay buffer The state of random sampling and from the standard normal distribution Noise vectors randomly sampled from the middle Simultaneously take the mathematical expectation; For the policy network in state Down Output Action The logarithmic probability; For the commentator network on the current state-action pair of Value estimation; The target commentator network employs a soft update mechanism, with the commentator network parameters being slowly tracked. ; In the formula, Parameters for the target critic network; The parameter update assignment sign means that the calculation result on the right side of the equals sign is assigned to the parameter on the left side; This is a soft update coefficient used to control the update speed of the target network parameters; Temperature parameters The policy entropy is automatically adjusted to approximate the target entropy by minimizing the following objectives: ; In the formula, Temperature parameter The optimization objective function; For policy networks Mid-sampling action Take the expected value; Let be the target entropy.

[0040] In this embodiment, regarding the strategy generation of the deep learning small model, when the meta-decision mechanism selects the deep learning small model, a scheduling strategy is generated based on a Graph Attention Network (GAT). Using the distribution network topology as the graph structure and operational measurement data as node features, spatial coupling features are extracted through the graph attention layer and directly mapped to a switch opening / closing command sequence and source / storage output values ​​via a fully connected output layer. Specifically: 1. Constructing the distribution network topology: The distribution network topology is used as a graph structure. The physical nodes in the distribution network include bus nodes, feeder segment nodes, distributed power generation access nodes, energy storage system access nodes, and load access nodes. Each physical node is represented as a vertex in the graph, and the lines between physical nodes are represented as edges in the graph, which together constitute the distribution network topology. 2. Constructing the node feature matrix: The input to the graph attention layer is the node feature matrix. ,in For the number of nodes, For the first Layer feature dimension; The node feature matrix of the first layer, i.e. the input layer, is composed of operational measurement data. Each row of the node feature matrix corresponds to a node feature vector. The node feature vector contains the voltage amplitude measured at the corresponding node, the active power output of the distributed power source connected to the node, the state of charge of the energy storage system connected to the node, and the active power of the load connected to the node. If no distributed power source, energy storage system, or load is connected to a node, the value of the corresponding feature dimension is zero. 3. Calculate the attention coefficient: For each node and its neighboring nodes in the distribution network topology diagram, take the corresponding node feature vector and calculate the attention coefficient between the node and its neighboring nodes. ; In the formula, For the first Layer nodes With nodes Attention coefficients between neighboring nodes; For the first The attention weight vector of the layer; For the first Transpose of the layer attention weight vector; For the first The linear transformation weight matrix of the layer; For the first Layer nodes eigenvectors; For the first Layer nodes eigenvectors; This is a vector concatenation operation; For activation functions; 4. Normalized Attention Coefficient: Take the attention coefficients between a node and its neighbors as exponents with the natural constant as the base, and then divide them by the sum of the exponents of the unnormalized attention coefficients of all the node's neighbors to obtain the normalized attention coefficients. ; In the formula, For the first Layer nodes For nodes The normalized attention coefficient; It is an exponential function with the natural constant as its base; For the first Layer nodes With nodes Attention coefficients between neighboring nodes; For nodes The set of neighboring nodes; 5. Update node feature vectors: For each node, using its corresponding normalized attention coefficient, the feature vectors of all its neighboring nodes are mapped by a linear transformation weight matrix, weighted, and then summed. This summation is then passed through a non-linear activation function to obtain the updated feature vector of the node. ; In the formula, For the first Layer nodes The updated feature vector; It is a non-linear activation function; 6. Global Average Pooling: The last layer obtained in step 5, i.e., the... The summation of the feature vectors of all nodes in the layer, divided by the total number of nodes, yields the global graph embedding vector representing the global characteristics of the entire distribution network topology. ; In the formula, Embed vectors for the global graph; This represents the total number of nodes in the distribution network. For the first The layer refers to the output feature vector of the last layer of nodes; 7. Output scheduling strategy: Embed the global graph into a vector. The first fully connected layer and the sigmoid activation function map the opening and closing probability vectors of each controllable switch. After classification by threshold, discrete 0 / 1 switch states are obtained; the global graph is embedded into a vector. The normalized source-storage output coefficients are mapped through a second fully connected layer and a hyperbolic tangent activation function, and then multiplied element-wise by the upper limit vector of the rated output of the source-storage equipment to obtain the source-storage output value vector. : ; ; In the formula, The vector of the probability distribution of the opening and closing commands of each controllable switch is used to classify the discrete switch states (0 / 1) according to the threshold. The source storage output value; for Activation function; This is the weight matrix for discrete actions, i.e., switching actions, in the output layer; The bias vector for the discrete actions, i.e., the switching actions, of the output layer; It is the hyperbolic tangent activation function; The weight matrix for continuous actions of the output layer, i.e., source and output forces; The bias vector for continuous output layer action, i.e., source and output force. This is element-wise multiplication; This is the upper limit vector of the rated output of the source and storage equipment.

[0041] In this embodiment, the generated scheduling strategy or the invoked historical scheduling scheme is security verified based on the power flow calculation of the distribution network to ensure that the scheduling strategy satisfies: the power flow balance equation of the distribution network, node voltage constraints, line current carrying capacity constraints, topology constraints, and equipment capacity constraints. For distribution networks, lines The power flow equilibrium equation is: ; ; ; ; In the formula, For the line Active power transmitted at the first end, line Defined as a connection node in the distribution network topology and nodes An AC power distribution line; For the line The resistance; For the line The current amplitude; For nodes The net active power demand of the load; For nodes Flow to child nodes The active power; For the line Reactive power transmitted at the first end; For the line The reactance; For nodes The net reactive power demand of the load; For nodes Flow to child nodes reactive power; End node of the line The voltage amplitude; For the first node of the line The voltage amplitude.

[0042] The node voltage constraint is: ; In the formula, and They are nodes Permissible lower and upper voltage limits.

[0043] The line current carrying capacity constraint is: ; In the formula, For the line Long-term allowable carrying capacity; Topological constraints include radial constraints and connectivity constraints; Radial constraints require the distribution network to maintain radial operation after the dispatching strategy is executed, meaning that the subgraph formed by the closed switches is acyclic, represented as: ; In the formula, The number of circuits where the switch is in the closed state; The number of nodes connected to the power supply node; The number of power nodes; this formula guarantees a tree structure and ensures that there are no loops in the powered topology; Connectivity constraints require that all load nodes be connected to the power supply nodes to ensure that there are no islanded nodes, as shown below: ; In the formula, For the set of load nodes; This is the set of nodes connected to the power supply node; this formula ensures that each load node is connected to at least one power supply node. Equipment capacity constraints include output limitations for distributed power sources and energy storage systems, expressed as: ; ; ; ; In the formula, For nodes The active power output of distributed power sources; For nodes The upper limit of the rated active power output of the distributed power source; and They are nodes The charging and discharging power of the energy storage system; and They are nodes The maximum charging power and maximum discharging power of the energy storage system; For nodes The state of charge of the energy storage system; and They are nodes The lower and upper limits of the charge capacity of the energy storage system; If all the above constraints are met, the security check is deemed successful, and the security check flag is set. Allow execution; otherwise, set. This triggers the large model in S4 to call the optimal power flow small model to solve the problem again.

[0044] Step S4: If the scheduling strategy or historical scheduling scheme fails the security check, the domain-fine-tuned general large model calls the optimal power flow small model to resolve and execute the scheduling strategy; the scheduling strategy and execution results are recorded in the historical scheme library, and the scheduling strategy and execution results are used for incremental training of the deep reinforcement learning small model and the deep learning small model, and the decision confidence of the deep reinforcement learning small model and the deep learning small model is corrected based on the execution results, triggering the incremental fine-tuning of the domain-fine-tuned general large model.

[0045] In this embodiment, when the security verification flag At that time, the scheduling master station will implement the scheduling strategy. The remote control command is encapsulated in accordance with the communication rules of the distribution network. The remote control command includes a timestamp, the target status field of each controllable switch in the switch opening and closing instruction sequence, the active power setting field of each distributed power source and energy storage system in the power source and energy storage output setting value, and a verification frame. The dispatch master station sends the remote control command to each distribution network terminal through the dispatch data network. After receiving the remote control command, the distribution network terminal performs the corresponding switch position change operation and power source and energy storage output adjustment, and after the execution is completed, it feeds back the execution result containing actual measurement data or alarm signals to the dispatch master station.

[0046] In this embodiment, if the scheduling policy described in step S3 fails the security check, i.e., the security check flag is set... When this happens, a fallback mechanism is triggered, and the general large model, after domain fine-tuning, calls the optimal power flow small model to resolve the scheduling strategy. The optimal power flow small model models the scheduling strategy solution problem under the current operating scenario as a mixed integer nonlinear programming problem. The objective function of the optimal power flow small model is the weighted minimization of distribution network active power loss, switching action cost, and load shedding. The decision variables include the discrete opening and closing states of controllable switches and the continuous active power output of distributed power sources and energy storage systems. The constraints are the power flow balance equation, node voltage constraints, line current carrying capacity constraints, topology constraints, and equipment capacity constraints.

[0047] In this embodiment, the objective function is: ; In the formula, For a moment The open / closed state vectors of all controllable switches; For a moment Active power output vector of distributed generation; For a moment Active power output vector of energy storage system; For a moment Load shedding vectors for each load node; This is the weighting coefficient for active power losses in the distribution network; For the set of routes; For the line The resistance; For the line The current amplitude; The weighting coefficient for the cost of the switching action; A set of controllable switches; For the first The cost of a single action of a controllable switch; for Time of the first The open / closed state of a controllable switch; for Time of the first The open / closed state of a controllable switch; The weighting coefficient for the load shedding penalty; For the set of load nodes; For a moment node The active power of the load removed at the site; Solving the optimal power flow small model yields a new scheduling strategy. Afterwards, It is encapsulated as a remote control command and executed.

[0048] In this embodiment, after the distribution network terminal executes the scheduling strategy, it feeds back the actual execution results to the scheduling master station. The execution results include the actual voltage amplitude of each node, the actual current amplitude of each line, the actual output of each distributed power source and energy storage system, and the actual status of the switches. The scheduling master station calculates the execution effect score based on the feedback data, which is used for the subsequent correction of the model decision confidence and the updating of the historical scheme database. time Execution performance score Taking into account voltage quality, current carrying capacity margin, distribution network losses, and load recovery rate, it can be expressed as: ; In the formula, , , , These are the weighting coefficients for voltage quality, current carrying capacity margin, distribution network losses, and load recovery rate, respectively, and they satisfy the following conditions: ; for Voltage quality score at any given time; for Real-time capacity margin score; for Real-time distribution network loss scoring; for Time-based load recovery rate score; In this embodiment, the scores for each item are calculated as follows: ; ; ; ; In the formula, For the set of distribution network nodes, , This represents the total number of nodes in the distribution network. For nodes Voltage at the point; For nodes The rated voltage at the location; For the collection of distribution network lines, , Total number of lines; For the line The current amplitude; For the line Long-term allowable carrying capacity; This represents the total active power loss of the distribution network. This represents the maximum permissible active power loss in the distribution network. For nodes The active power of the loads that have had their power restored; For nodes Active power demand of the load.

[0049] The current scheduling scenario and execution result are recorded in the historical solution database, and similarity deduplication and merging operations are performed; specifically, the current structured prompt words are recorded. Scheduling strategy Execution effect score and time Query semantic vector of the scenario Constructing new historical scene records ; Before the historical scheme database was updated, the calculation time was... Query semantic vector of the scenario Query semantic vectors of existing historical scenarios in the database Similarity: ; In the formula, This is a similarity calculation function; For the first The query semantic vector of each historical scenario Let L2 be the norm of the vector; If historical scenarios exist satisfy: ; In the formula, For historical scenes The query semantic vector; The index of the scene in the historical solution library that has the highest semantic similarity to the current scene; The similarity threshold for merging historical scheme databases; Then perform a merge operation: compare the current scene with the historical scene. The execution effect score, if , For historical scenes The performance evaluation score is then replaced with the current scene record to replace the historical scene record. Record, otherwise retain historical scenes. The existing records remain unchanged; if no existing historical scene meets the above conditions, the current scene record is directly appended to the historical solution library; after merging or appending, the vector index of the historical solution library is updated to ensure that the retrieval enhancement generation mechanism can retrieve the latest scene record; The current scenario-decision pair is added to the training samples of the small model to update the training samples of the small model. The scenario-decision pair consists of the current distribution network operation scenario and the corresponding scheduling decision. The operation scenario is the model input and the scheduling decision is the model's expected output. In this embodiment, for a small deep reinforcement learning model, the current experience is used. Added as a sample to the experience replay pool ,in For a moment state, For a moment The action, For a moment Instant rewards received for The state at any given moment; when the number of new samples in the experience replay pool exceeds the threshold, incremental training of the deep reinforcement learning mini-model is triggered, and the parameters of the policy network and the critic network are updated by training on mixed data of new samples and historical samples. For small deep learning models, the current running measurement data Corresponding scheduling strategy Scenario-decision pairs as samples When the number of new samples added to the training sample set reaches a threshold, incremental training of the deep learning mini-model is triggered. The parameters of the deep learning mini-model are updated by training on a mixture of new and historical samples.

[0050] The decision confidence of each model is dynamically adjusted based on the performance score, realizing feedback correction of the model decision confidence; let the decision confidence of the deep reinforcement learning small model be... The decision confidence of the deep learning small model is If at any time After the meta-decision mechanism selects and executes a deep reinforcement learning small model, its decision confidence is updated using the following formula: ; In the formula, The updated decision confidence of the deep reinforcement learning small model; The decision confidence of the deep reinforcement learning small model before the update; The learning rate is the confidence level of the model decision, and its value ranges from [value range missing]. ; For a moment Execution effectiveness score; Similarly, when the meta-decision mechanism selects a small deep learning model, its decision confidence update formula is: ; In the formula, The updated decision confidence of the deep learning small model; The decision confidence of the small deep learning model before the update; If a certain model is adjacent If a decision is not selected by the meta-decision mechanism within a scheduling cycle, its decision confidence level will be corrected. ; In the formula, For the model before the update Decision confidence, Indicates model identifier, Represents a small deep reinforcement learning model. Represents a small deep learning model; For the updated model Decision confidence level; This is the decision confidence decay coefficient, with a value range of [value range missing]. .

[0051] Incremental fine-tuning of the general large model for domain fine-tuning is triggered when any of the following conditions are met: 1. The number of newly added scenario records in the historical solution library reaches the preset incremental fine-tuning sample threshold; 2. The decision confidence of the deep reinforcement learning small model or the deep learning small model is lower than the decision confidence lower limit for several consecutive periods. In this embodiment, the training samples for incremental fine-tuning are constructed by the scheduling master station from the historical scheme database, selecting execution performance scores. Higher than the preset performance quality threshold Historical scene recording; incremental fine-tuning adopts a low-rank adaptation-based fine-tuning method, while keeping the general large model base parameters (i.e., all model parameters learned by the general large model after domain fine-tuning during the large-scale pre-training stage) frozen, training the low-rank adaptation parameter matrix; the optimization objective of incremental fine-tuning is to minimize the following loss function: ; In the formula, The loss function is for incremental fine-tuning; For incremental fine-tuning of the sample set; For the first training samples, For the first Structured cue words for each training sample, For the first Each training sample corresponds to a scheduling strategy; Freeze parameters for the general large model base; The low-rank fitting parameter matrix to be trained; In a given , , The following model outputs a scheduling strategy. The conditional probability; The regularization coefficient is used. The Frobenius norm is used; the low-rank adaptation parameter matrix is ​​defined as a set of trainable incremental parameters injected in a bypass manner next to the attention weight matrix of each layer of the domain-fine-tuned general large model architecture.

[0052] After incremental fine-tuning is completed, the meta-decision function of the general large model is updated. This enables the general large model to continuously evolve its semantic understanding and meta-decision reasoning capabilities for distribution network operation scenarios with actual execution feedback.

[0053] In this embodiment, through the closed-loop mechanism of strategy execution, historical solution library update, incremental update of training samples, dynamic correction of model decision confidence, and incremental fine-tuning of the general large model after domain fine-tuning, the synergy and self-evolution of the small and large models are realized. The solution capability of each small model is continuously enhanced by adding training samples, the meta-decision capability of the general large model after domain fine-tuning is continuously optimized by incremental fine-tuning and expansion of the historical solution library, the historical solution library is kept concise and efficient through deduplication and merging mechanism, and the model decision confidence dynamically reflects the actual reliability of each model in the current operating environment through execution feedback. Thus, under the condition of continuous change in the distribution network operating environment, the reliability of dispatch strategy generation is continuously improved.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent dispatching of distribution networks based on large-model meta-decision and multi-small-model collaboration, characterized in that, Includes the following steps: Step S1: Obtain multi-source heterogeneous data of real-time operation of the distribution network. The multi-source heterogeneous data includes operation measurement data and alarm signals. Risk perception is performed based on the operation measurement data. Voltage over-limit, line overload and fault information are recorded based on the alarm signals and converted into structured prompt words. Step S2: Input the structured prompts into the domain-fine-tuned general large model for meta-decision. Based on the retrieval enhancement generation technology, query the historical similar running scenarios in the historical scheme library. Select one of the deep reinforcement learning small model or the deep learning small model according to the scenario matching degree, or directly call the corresponding historical scheduling scheme. The deep reinforcement learning small model adopts the soft actor-critic algorithm, and the deep learning small model adopts the graph attention network. Step S3: If a deep reinforcement learning small model or a deep learning small model is selected in step S2, a scheduling strategy containing a switch opening and closing instruction sequence and source and storage output values ​​is generated by the selected deep reinforcement learning small model or deep learning small model. The generated scheduling strategy is then subjected to a safety verification. If the safety verification passes, the strategy is executed. If a historical scheduling scheme is called in step S2, the historical scheduling scheme is then subjected to a safety verification. If the safety verification passes, the strategy is executed. Step S4: If the scheduling strategy or historical scheduling scheme in step S3 fails the security check, the domain-fine-tuned general large model calls the optimal power flow small model to resolve and execute the scheduling strategy; the scheduling strategy and execution results are recorded in the historical scheme library, and the scheduling strategy and execution results are used for incremental training of the deep reinforcement learning small model and the deep learning small model, and the decision confidence of the deep reinforcement learning small model and the deep learning small model is corrected based on the execution results, triggering the incremental fine-tuning of the domain-fine-tuned general large model.

2. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 1, characterized in that: In step S1, the operational measurement data includes the voltage amplitude of each node in the distribution network, the current amplitude of each line, the active power output of each distributed power source, the state of charge of each energy storage system, and the active power of each load. After removing bad data from the operational measurement data, the remaining valid operational measurement data is normalized.

3. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 2, characterized in that: In step S1, the risk perception includes node voltage over-limit risk perception and line current carrying capacity over-limit risk perception; Based on the normalized operational measurement data, the voltage over-limit risk index of each node and the current carrying capacity over-limit risk index of each line are calculated respectively.

4. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 3, characterized in that: In step S2, the scene matching degree is divided into three levels: complete matching, partial matching, and no matching. When the scene semantic similarity reaches the complete matching similarity threshold, and the current scene and the historical most similar scene are semantically consistent in key elements such as fault location, over-limit equipment, topology, source and storage output, and risk level, it is determined to be a complete match, and the corresponding historical scheduling scheme is directly invoked. When the semantic similarity of the scene is between the partial matching similarity threshold and the full matching similarity threshold, it is determined to be a partial match; when the semantic similarity of the scene is lower than the partial matching similarity threshold, it is determined to be a mismatch; in the case of partial match or mismatch, the general large model that has been fine-tuned by the domain performs meta-decision based on the decision confidence of each small model, and selects the deep reinforcement learning small model or the deep learning small model.

5. The intelligent dispatching method for distribution networks based on large-model meta-decision-making and multi-small-model collaboration as described in claim 4, characterized in that: In step S3, when a deep reinforcement learning small model is selected, a soft actor-critic algorithm is used to generate a scheduling strategy. The soft actor-critic algorithm includes an actor network, a critic network, and a target critic network. The actor network takes the current operating state of the distribution network as input and outputs a scheduling action probability distribution containing switch opening and closing commands and source-storage-output setpoints. The critic network takes state-action pairs as input and outputs an action value estimate. The parameters of the target critic network are updated by tracking the parameters of the critic network.

6. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 5, characterized in that: In step S3, when selecting a small deep learning model, a graph attention network generation scheduling strategy is adopted; Using the distribution network topology as a graph structure and operational measurement data as node features, spatial coupling features are extracted through the graph attention layer and mapped to switch opening and closing command sequences and source-storage output values ​​through the fully connected output layer.

7. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 6, characterized in that: In step S3, the safety verification is based on power flow calculation of the distribution network. The verification content includes power flow balance constraints, node voltage constraints, line current carrying capacity constraints, topology constraints, and equipment capacity constraints. If all constraints are satisfied, the safety verification is deemed to have passed; otherwise, the safety verification is deemed to have failed.

8. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 7, characterized in that: In step S4, the optimal power flow small model models the scheduling strategy solution problem as a mixed integer nonlinear programming problem; the decision variables include the discrete on / off states of controllable switches and the continuous active power output of distributed power sources and energy storage systems; the objective function is the weighted minimization of distribution network active power loss, switch action cost and load shedding; the constraints include power flow balance equations, node voltage constraints, line current carrying capacity constraints, topology constraints and equipment capacity constraints.

9. The intelligent dispatching method for distribution networks based on large-model meta-decision and multi-small-model collaboration as described in claim 8, characterized in that: In step S4, after the scheduling strategy is executed, the execution effect score is calculated based on the actual voltage amplitude of each node in the distribution network, the actual current amplitude of each line, the actual output of each distributed power source and energy storage system, and the actual status of the switches. The current structured prompt is input into the domain-fine-tuned general model, and the query semantic vector of the current scenario is obtained through embedding encoding. The current structured prompt, scheduling strategy, execution effect score, and query semantic vector of the current scenario are used to construct a new historical scenario record and add it to the historical solution library. The similarity between the query semantic vector of the current scenario and the query semantic vector of the existing historical scenarios in the library is calculated. If there is a historical scenario with a similarity greater than the merging similarity threshold, the execution effect scores of the two are compared. The original historical scenario record is replaced only if the execution effect of the current scenario is better; otherwise, the original record is retained. If there is no existing historical scenario that meets the conditions, the new scenario record is directly appended.

10. The intelligent dispatching method for distribution networks based on large-model meta-decision-making and multi-small-model collaboration as described in claim 9, characterized in that: In step S4, the correction of the decision confidence of the deep reinforcement learning small model and the deep learning small model includes: if the small model selected and executed by the current meta-decision mechanism is a deep reinforcement learning small model or a deep learning small model, its decision confidence is updated by weighting according to the execution effect score; for deep reinforcement learning small models or deep learning small models that are not selected by the meta-decision mechanism in adjacent scheduling cycles, their decision confidence is decayed according to the decision confidence decay coefficient; when the number of newly added scene records in the historical solution library reaches the preset incremental fine-tuning sample threshold, or when the decision confidence of the deep reinforcement learning small model is lower than the decision confidence lower limit for multiple consecutive cycles, or when the decision confidence of the deep learning small model is lower than the decision confidence lower limit for multiple consecutive cycles, the incremental fine-tuning of the general large model that has been fine-tuned by the domain is triggered; the incremental fine-tuning adopts a fine-tuning method based on low-rank adaptation, freezes the base parameters of the general large model, and trains only the low-rank adaptation parameter matrix.

Images