Large model dynamic call heuristic optimization method for power distribution network reconstruction and solver

Abstract

The invention provides a large model dynamic call heuristic optimization method and solver for power distribution network reconstruction, and relates to the technical field of smart power grids, and the method comprises the steps: constructing a heuristic optimization model library comprising a plurality of different optimization preferences; the state of the power distribution network is perceived in real time, a field large language model capable of understanding the causal relationship between the state of the power distribution network and operation indexes serves as an intelligent scheduling center, the current situation is dynamically analyzed, and the most adaptive heuristic optimization model is screened, evaluated and called from a pre-constructed model library for iterative solution. And outputting an optimal power distribution network reconstruction scheme. Evaluating the executed power distribution network reconstruction scheme to obtain a new large language model used for training the large language model in the field, performing reinforcement learning training, and optimizing the future model scheduling and decision-making capability of the new large language model; on the basis, the purpose of considering efficiency, precision and reliability in the power distribution network reconstruction process is achieved.

Description

Technical Field

[0001] This application relates to the field of smart grid technology, and in particular to a dynamic heuristic optimization method and solver for large-scale model reconfiguration of distribution networks. Background Technology

[0002] With the development of smart grid technology and the increasing complexity of distribution network structure, distribution network reconfiguration, as a key means to optimize distribution network operation, can ensure the safe and stable operation of the distribution network, reduce losses, and improve power supply reliability.

[0003] In addressing the computationally complex problem of distribution network reconfiguration, existing methods primarily rely on traditional algorithms such as mathematical programming and heuristics. These traditional algorithms are computationally too time-consuming when dealing with large-scale real-world distribution networks, making them unsuitable for real-time decision-making. While large-scale models are being introduced into the specialized power sector, these generalized or finely tuned models lack a deep and precise understanding of the distribution network's physical mechanisms, operating rules, topology, and specific optimization objectives (such as loss minimization and reliability maximization). Consequently, the reconfiguration schemes they directly output carry significant risks in terms of mathematical feasibility and physical security. Furthermore, although large-scale models possess a certain degree of generalization ability, the "one-step" planning approach struggles to guarantee the correctness and effectiveness of solutions under multiple constraints when faced with distribution network reconfiguration scenarios requiring precise calculation and optimization.

[0004] It is evident that existing technologies, when dealing with dynamic and complex distribution network reconfiguration tasks, cannot intelligently select, invoke, and drive the most suitable underlying optimization model and parameters based on the dynamic task and real-time status. Consequently, efficiency, accuracy, and reliability cannot be balanced during the distribution network reconfiguration process. Summary of the Invention

[0005] In view of this, embodiments of this application provide a large-scale model dynamic calling heuristic optimization method and solver for distribution network reconfiguration, in order to solve the problem that existing methods cannot balance efficiency, accuracy and reliability in the process of distribution network reconfiguration.

[0006] To address the above problems, this application provides the following technical solutions:

[0007] The first aspect of this application discloses a heuristic optimization method for dynamic invocation of a large-scale model of distribution network reconfiguration, the method comprising:

[0008] Generate a family of heuristic optimization models containing various optimization preferences, and construct a heuristic optimization model library based on the family of heuristic optimization models;

[0009] Real-time acquisition of multi-source data from the distribution network, performance of condition perception and status feature extraction, and generation of natural language contextual summaries describing the current operating status of the distribution network;

[0010] Based on the natural language context summarization and scheduling requirements, a suitable candidate heuristic optimization model is selected and called from the heuristic optimization model library using a domain-wide large language model; the domain-wide large language model is trained based on the large language model.

[0011] Evaluate the candidate heuristic optimization models and determine the optimal strategy based on the evaluation results;

[0012] The optimal strategy is transformed into an executable sequence of scheduling instructions and executed on-site. The sequence of scheduling instructions is iteratively adjusted in a rolling optimization manner until the process termination condition is met, and the power distribution network reconfiguration scheme is output.

[0013] The proposed power distribution network reconfiguration scheme is evaluated to obtain a new large language model for training the domain large language model.

[0014] Optionally, the power distribution network reconfiguration scheme is evaluated to obtain a new large language model for training the domain large language model, including:

[0015] The evaluation module is invoked to calculate the quantitative impact of the power distribution network reconfiguration scheme on key indicators of actual operation, and a quantitative effect report is generated.

[0016] The quantitative effect report, the distribution network reconfiguration scheme, and the entire decision trajectory of executing the distribution network reconfiguration scheme are fed back to the large language model. The large language model is then trained using scheduling effect preference reward and distribution network condition perception reward to obtain a new large language model for training the domain large language model.

[0017] Optionally, a heuristic optimization model library is constructed based on the aforementioned family of heuristic optimization models, including:

[0018] For each heuristic optimization model in the heuristic optimization model family, a corresponding multi-dimensional metadata tag is established. The metadata tag includes functional tags, performance characteristic tags, applicable operating condition constraints, and interface specifications.

[0019] By using a database that supports multi-tag combined queries, each heuristic optimization model and its corresponding multi-dimensional metadata tags are associated and stored to obtain a heuristic optimization model library.

[0020] Optionally, the training process of the domain-specific large language model includes:

[0021] Construct a professional corpus containing distribution network professional corpus, and an instruction-output pair dataset containing causal reasoning for complex scheduling scenarios. The distribution network professional corpus includes at least one or more of the following: distribution network operation procedures, technical manuals, operation tickets, work tickets, historical fault handling reports, real-time operation logs, and related academic literature.

[0022] Using the aforementioned professional corpus as training data, a domain-adaptive pre-training of a general large language model is performed using a self-supervised learning task. This enables the pre-trained large language model to recognize entities, operational terms, and key operating indicators specific to the power distribution network domain, as well as the interrelationships among these entities, operational terms, and key operating indicators.

[0023] By using the instruction-output pair dataset and comparative learning samples designed based on the instruction-output pair dataset, the reasoning ability of the pre-trained large language model for the causal relationship between power distribution network operations and results is supervised and fine-tuned, thus obtaining a domain-specific large language model.

[0024] Optionally, multi-source data from the distribution network can be acquired in real time, and operating condition perception and state feature extraction can be performed to generate a natural language context summary describing the current operating status of the distribution network, including:

[0025] Based on a standardized industrial communication interface, multiple information sources are accessed, and the real-time collected multi-source data is synchronized, cleaned, and fused to obtain fused multi-source data.

[0026] State feature vectors are extracted and calculated from the fused multi-source data. These state feature vectors are used to characterize the current operating status of the distribution network and include fault features, network status features, resource features, and environmental features.

[0027] The state feature vector is combined with the distribution network topology information, and a natural language context summary describing the current operating status of the distribution network is generated based on the combined information.

[0028] Optionally, based on the natural language scenario summarization and scheduling requirements, a suitable candidate heuristic optimization model is selected and invoked from the heuristic optimization model library using a domain-wide large language model, including:

[0029] The natural language scenario summary and scheduling instructions are input into the domain big language model, and the current core task, operational constraints and user expected optimization preferences are analyzed and understood in the domain big language model to obtain understanding information;

[0030] The understanding information is converted into a semantic query for the heuristic optimization model library using the domain-wide language model. Based on the semantic query, a preliminary candidate strategy list is retrieved from the heuristic optimization model library. The candidate strategy list includes the preliminary strategy encapsulated as a heuristic optimization model.

[0031] Using the domain-specific large language model, for each candidate strategy in the candidate strategy list, combined with the current distribution network operating parameters, an adjustment strategy for the key operating parameters of each candidate strategy is output. Based on the adjustment strategy, the corresponding heuristic optimization model is adjusted to obtain an adapted candidate heuristic optimization model.

[0032] Optionally, the candidate heuristic optimization model is evaluated, and the optimal strategy is determined based on the evaluation results, including:

[0033] Using the current real distribution network state as the initial condition, the candidate strategies output by the candidate heuristic optimization model are simulated and executed in parallel in a lightweight simulation environment to obtain simulation results. The simulation results include the operational consequences and distribution network state changes produced by executing each candidate strategy within a preset future time.

[0034] Extract multi-dimensional performance metrics for each candidate strategy from the simulation results. These multi-dimensional performance metrics include: expected power recovery time, system bus loss, minimum node voltage, maximum device load rate, and switching operation complexity.

[0035] The multi-dimensional performance indicators are input into the multi-objective trade-off model. Based on the weights preset or learned online in the multi-objective trade-off model, a comprehensive utility score and ranking are performed to obtain the output ranking list. The weights are used to reflect the scheduling priority under different operating scenarios.

[0036] Based on the finite-time rolling optimization objective function, the expected comprehensive cost of the candidate strategies in the sorted list is evaluated, and the candidate strategy with the minimum expected comprehensive cost is determined as the optimal strategy.

[0037] Optionally, the optimal strategy is transformed into a sequence of executable scheduling instructions for on-site execution, and the sequence of scheduling instructions is iteratively adjusted using a rolling optimization approach until the process termination condition is met, outputting a distribution network reconfiguration scheme, including:

[0038] With an optimization time window length of T and the current distribution network fault state as the initial state, the optimal strategy is transformed into a sequence of scheduling instructions containing specific executable operations and sent to the field execution equipment, so that the field execution equipment executes the sequence of scheduling instructions, which includes switching operations.

[0039] After the first scheduling instruction sequence is executed, the step of acquiring multi-source data of the distribution network in real time, performing condition perception and state feature extraction is executed to obtain the new state of the distribution network as the new initial state, and updating the remaining optimization time window length to T-1.

[0040] Based on the new initial state and the remaining optimization window length, the optimal strategy is re-determined by combining the finite-time-domain rolling optimization objective function. The process then returns to the step of converting the optimal strategy into a sequence of scheduling instructions containing specific executable operations and issuing them to the field execution equipment, until the process termination condition is met, and the distribution network reconfiguration scheme is output. The process termination condition includes the new initial state being a non-fault state, or the operating indicators meeting a preset level, or the optimization window length being 0.

[0041] A second aspect of this application discloses a solver, the solver comprising:

[0042] An offline construction module is used to generate a family of heuristic optimization models containing multiple different optimization preferences, and to annotate and store the family of heuristic optimization models with structured metadata in order to build a heuristic optimization model library. Each heuristic optimization model in the heuristic optimization model library is encapsulated by a policy.

[0043] The online dynamic scheduling solution module is used to acquire multi-source data of the distribution network in real time, perform condition perception and state feature extraction, and generate a natural language scenario summary describing the current operating status of the distribution network. Based on the natural language scenario summary and scheduling requirements, it uses a domain-wide language model to select and call suitable candidate heuristic optimization models from the heuristic optimization model library. The domain-wide language model is trained based on a general language model. The candidate heuristic optimization models are evaluated, and the optimal strategy is determined based on the evaluation results. The optimal strategy is transformed into an executable scheduling instruction sequence and executed on-site. The scheduling instruction sequence is iteratively adjusted in a rolling optimization manner until the process termination condition is met, and the distribution network reconfiguration scheme is output.

[0044] An offline closed-loop learning optimization module is used to evaluate the power distribution network reconfiguration scheme and obtain a new general-purpose large language model for training the domain large language model.

[0045] A third aspect of this application discloses a solver, which includes at least one processor and at least one memory and bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the large-scale dynamic calling heuristic optimization method for power distribution network reconfiguration as described in the first aspect of this application.

[0046] Based on the above-described embodiments of this application, a large-scale model-based heuristic optimization method and solver for power distribution network reconfiguration is provided. This method constructs a closed-loop paradigm of "offline preparation - online solving - continuous optimization." In the offline preparation phase, a heuristic optimization model library containing various optimization preferences is built. In the online solving phase, the power distribution network status is perceived in real time. A domain-specific large-scale language model, capable of understanding the causal relationship between the power distribution network status and operational indicators, serves as the intelligent scheduling hub. It dynamically analyzes the current situation and selects, evaluates, and calls the most suitable heuristic optimization model from the pre-built model library for iterative solving, outputting the optimal power distribution network reconfiguration scheme. In the continuous optimization phase, the executed power distribution network reconfiguration scheme undergoes multi-dimensional performance quantification evaluation. Based on the evaluation results, the domain-specific large-scale language model optimizes its ability to select heuristic optimization models and make scheduling decisions in the next power distribution network reconfiguration process, achieving autonomous evolution of system performance and balancing efficiency, accuracy, and reliability during power distribution network reconfiguration. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0048] Figure 1 A flowchart illustrating a large-scale model for power distribution network reconfiguration that dynamically invokes a heuristic optimization method, as provided in an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the structure of a solver provided in an embodiment of this application;

[0050] Figure 3 This is a schematic diagram of another solver provided in an embodiment of this application. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0052] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] The following are descriptions of the relevant technologies involved in the embodiments of this application:

[0054] Large Language Model (LLM) refers to a natural language processing model based on deep learning with a massive number of parameters, capable of understanding, generating, and reasoning about text.

[0055] Domain-specific large language model: This refers to a language model that, based on a large language model, undergoes further pre-training or fine-tuning using professional texts in the power distribution network field (such as dispatching procedures, technical manuals, operation tickets, work tickets, and historical cases), thereby acquiring the ability to understand and reason about power distribution network professional knowledge. In the embodiments of this application, it specifically refers to a domain-specific large language model that has been pre-trained or fine-tuned with professional knowledge in the power distribution network field and is used to analyze power distribution network operating conditions, dispatching decisions, and generate strategies.

[0056] Heuristic optimization models refer to fast solution algorithms or strategies based on experience or intuitive rules designed for distribution network optimization problems (such as load transfer and fault recovery). In the embodiments of this application, such heuristic optimization models typically exist in the form of programmed code, parameterized rule sets, etc., and can provide approximate optimal solutions within an acceptable time.

[0057] Comparative evolution refers to a method for automatically generating and improving heuristic optimization models. This method extracts improvement strategies by comparing and analyzing the differences between solutions generated by seed heuristic models and better solutions, and iteratively applies them to model updates, thereby evolving a family of heuristic models with better performance and greater diversity.

[0058] Strategy Knowledge Base: A structured database that stores and manages heuristic optimization models and their metadata. Each model comes with detailed tags, including functional descriptions, applicable scenarios, optimization objectives, performance characteristics, etc., supporting rapid semantic-based retrieval and matching.

[0059] Real-time operating condition awareness: refers to the real-time acquisition of the operating status information of the distribution network through data acquisition and monitoring systems (such as SCADA), including fault information, load data, equipment status, power flow distribution, etc., and extracting key features from them to form a comprehensive description of the current operating status of the distribution network.

[0060] Multi-objective trade-offs refer to the process of comprehensively evaluating and ranking candidate strategies in distribution network optimization scheduling, while simultaneously considering multiple conflicting objectives (such as shortest recovery time, minimum line loss, and optimal voltage quality), through a certain trade-off model or preference setting.

[0061] Rolling optimization refers to a dynamic optimization method that recalculates the optimization based on the new system state after each decision is executed, thereby continuously adjusting subsequent decisions and forming a closed loop of "perception-decision-execution-re-perception" to cope with system changes and uncertainties.

[0062] Dual reward mechanism: This refers to using two reward signals simultaneously to guide the agent's learning during reinforcement learning training. In the embodiments of this application, it specifically refers to a training mechanism that combines rewards based on scheduling effect preference (used to evaluate the overall effectiveness of scheduling schemes) and rewards based on distribution network condition awareness (used to evaluate the accuracy of identifying key state characteristics of the distribution network).

[0063] Policy evolution: refers to the process of continuously training and optimizing decision-making models (such as policy networks of large language models) by collecting the actual effects of historical scheduling cases and using methods such as reinforcement learning, so that their decision-making ability can be continuously improved over time.

[0064] Online dispatching refers to the process by which the system automatically generates and executes a dispatching scheme based on real-time operating conditions when a distribution network experiences a fault or requires optimization. This process requires low latency and high reliability and typically includes steps such as candidate strategy generation, rapid evaluation, and optimal selection.

[0065] Offline modeling refers to preparatory work such as model training and knowledge base construction using historical data and simulation environment before the system is officially deployed. It does not require real-time performance, but emphasizes the accuracy and completeness of the model.

[0066] Performance closed loop: refers to feeding back the actual execution effect of the scheduling scheme to the decision model for evaluation and optimization, thus forming a complete cycle from decision-making to execution to evaluation and improvement.

[0067] Adaptive solving: refers to the ability of a solution method to automatically adjust its solution strategy or parameters based on the specific characteristics and constraints of the current problem, in order to improve the efficiency and effectiveness of the solution.

[0068] Intelligent solver scheduling unit: This refers to the core component proposed in this patent. It is composed of a specialized large language model and is responsible for analyzing the real-time operating conditions of the distribution network, dynamically selecting, combining and adjusting heuristic optimization models from the strategy knowledge base, and generating the final scheduling instructions.

[0069] As can be seen from the background technology, when dealing with dynamic and complex distribution network reconfiguration tasks, existing technologies cannot intelligently select, call and drive the most suitable underlying optimization model and parameters according to the dynamic task and real-time status, thus making it impossible to balance efficiency, accuracy and reliability in the distribution network reconfiguration process.

[0070] Therefore, this application discloses a large-scale model-based heuristic optimization method and solver for distribution network reconfiguration. This method and solver are applied to systems related to distribution network reconfiguration. By constructing a closed-loop paradigm of "offline preparation - online solving - continuous optimization," in the offline preparation phase, a heuristic optimization model library containing various optimization preferences is built. In the online solving phase, the distribution network status is sensed in real time. A domain-specific large-scale language model, capable of understanding the causal relationship between the distribution network status and operating indicators, serves as the intelligent scheduling hub, dynamically analyzing the current situation and selecting, evaluating, and calling the most suitable heuristic optimization model from the pre-built model library for iterative solving, outputting the optimal distribution network reconfiguration scheme. In the continuous optimization phase, the executed distribution network reconfiguration scheme undergoes multi-dimensional performance quantification evaluation. Based on the evaluation results, the domain-specific large-scale language model optimizes its ability to select heuristic optimization models and make scheduling decisions in the next distribution network reconfiguration process, achieving autonomous evolution of system performance and realizing the goal of balancing efficiency, accuracy, and reliability in the distribution network reconfiguration process.

[0071] like Figure 1 The diagram shown is a flowchart illustrating a large-scale model dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in this application. This large-scale model dynamic invocation heuristic optimization method and solver for distribution network reconfiguration are applied to systems related to distribution network reconfiguration. The method includes the following steps:

[0072] S101: Generate a family of heuristic optimization models containing various optimization preferences, and build a heuristic optimization model library based on this family of heuristic optimization models.

[0073] In S101, each heuristic optimization model in the heuristic optimization model library is encapsulated by a policy.

[0074] In this embodiment, S101 is executed before the relevant system is deployed to complete the construction and training of core computing resources and decision-making agents, laying a solid foundation for subsequent online intelligent scheduling. By introducing the automatic evolution of heuristic optimization models, deep pre-training and fine-tuning of domain-specific large language models, and structured annotation and storage of policy knowledge bases, a heuristic optimization model library that can understand complex power distribution network scenarios, generate diverse policies, and is easy to call is constructed.

[0075] In the specific execution of S101, a family of heuristic optimization models containing various optimization preferences is generated, and structured metadata is labeled and stored to build a heuristic optimization model library.

[0076] This involves creating diverse, high-quality heuristic optimization strategies based on a comparative evolution approach and an automated "generate-compare-refine" cycle, and encapsulating these strategies into callable heuristic optimization models. Based on this, embodiments of this application can start from basic optimization rules and, through simulation and analysis, gradually refine strategies that are more suitable for specific scenarios.

[0077] S102: Acquire multi-source data of the distribution network in real time, perform condition perception and state feature extraction, and generate a natural language context summary describing the current operating status of the distribution network.

[0078] S103: Based on natural language context summarization and scheduling requirements, use a domain-wide large language model to select and call suitable candidate heuristic optimization models from the heuristic optimization model library.

[0079] In S103, the domain-specific large language model is trained based on the general large language model.

[0080] In one embodiment of this application, domain-specific training is specifically performed based on a general-purpose large language model, enabling the general-purpose large language model to possess accurate scheduling and decision-making capabilities. The specific training process is as follows:

[0081] Step 11: Collect and construct specialized corpora.

[0082] During step 11, high-quality professional corpora, including distribution network operation procedures, equipment technical manuals, historical fault handling reports, real-time operation logs, and relevant academic literature, were systematically collected. Based on the logical correspondences in the collected professional corpora, a representative instruction-output pair dataset was constructed. The instructions cover complex distribution network dispatching scenarios, and the corresponding outputs include detailed event chains, key operations performed, and analyses of the specific impacts of these operations on important power grid indicators (such as SAIDI, voltage deviation, and network loss).

[0083] That is, execute step 11 to build a professional corpus containing power distribution network professional corpus, and an instruction-output pair dataset containing causal reasoning for complex scheduling scenarios.

[0084] Step 12: Perform domain-adaptive pre-training based on professional corpora.

[0085] During step 12, the professional corpus collected in step 11 is used as training data. Self-supervised learning tasks, including Masked Language Modeling (MLM) and Next Sentence Prediction (NSP), are employed to further pre-train the general-purpose large language model. Through domain-adaptive training, the general-purpose large language model gains a deep understanding of entities specific to the distribution network domain (such as "Feeder Automation Terminal (FTU)" and "In-phase Operation") and their attributes, the meaning of operational terms (such as "Negative Sequence" and "Islanding Operation Mode"), and the concepts and interrelationships of key operational indicators (such as "Power Supply Reliability Index SAIFI" and "Three-phase Imbalance").

[0086] That is, in step 12, the professional corpus is used as training data, and the general large language model is pre-trained in a domain-adaptive manner using a self-supervised learning task, so that the pre-trained large language model has the ability to identify entities, operational terms and key operating indicators specific to the power distribution network, as well as the interrelationships between entities, operational terms and key operating indicators specific to the power distribution network.

[0087] Step 13: Fine-tune the instructions in the instruction-output dataset based on the professional corpus, and strengthen causal reasoning.

[0088] During step 13, the instructions in the instruction-output dataset are supervised and fine-tuned using a pre-built instruction dataset. This pre-built instruction dataset contains contrastive learning samples, which enhance the causal reasoning and indicator-based decision-making capabilities of the general large language model.

[0089] That is, by executing step 13, the ability of the pre-trained large language model to reason about the causal relationship between power distribution network operations and results is supervised and fine-tuned using the instruction-output pair dataset and the comparative learning samples designed based on the instruction-output pair dataset, thus obtaining the domain large language model.

[0090] Example: Contrastive learning samples specifically involve: given the causes of a "good" and a "bad" reconfiguration scheme under the same fault scenario, the model outputs an effect explaining the reasons for their superiority or inferiority; or given the cause of an operational result (such as "low voltage at a certain node"), the model infers the effect of the operational cause that may have led to this result. Through causal-based training, the model can more accurately understand the logical relationship between operations and results, providing a more reliable basis for strategy selection and parameter adjustment in online scheduling. Through this process, the general-purpose large language model is trained into a domain-specific large language model capable of understanding the complex constraints of the distribution network and evaluating the potential impact of operations on multiple key indicators, providing a cognitive foundation for subsequent contextual awareness and intelligent decision-making.

[0091] In one embodiment of this application, a general large language model is pre-trained and fine-tuned based on a distribution network professional corpus to obtain a domain large language model that can understand the causal relationship of distribution network scheduling.

[0092] In the specific execution of S103, the domain-wide language model is used as the intelligent solver scheduling unit. Based on natural language context summarization and scheduling requirements, candidate heuristic optimization models are dynamically selected, their parameters are adapted, and candidate strategies are invoked from the heuristic optimization model library. The specific execution process is as follows:

[0093] Step 21: Input the natural language scenario summary and scheduling requirements into the domain large language model, and output the understanding information of the current power grid scenario and scheduling requirements.

[0094] During step S21, the aforementioned generated natural language context summary and scheduling requirements, which may be temporary scheduling instructions (e.g., a dispatcher might input "Please prioritize ensuring the continuity of power supply to the hospital" via voice or text), are input into the trained domain-wide language model. Context understanding and requirement parsing are then performed within this domain-wide language model to achieve a deep understanding of the current core tasks (such as "rapid power restoration" and "minimizing economic losses"), key operational constraints (such as "ensuring uninterrupted power supply to critical users"), and users' expected optimization preferences.

[0095] That is, in step 21, the natural language context summary and scheduling instructions are input into the domain big language model, and the current core task, operation constraints and user expectation optimization preferences are analyzed and understood in the domain big language model to obtain understanding information.

[0096] Step 22: Based on the understood information, perform matching in the offline-built heuristic optimization model library to obtain a preliminary list of candidate strategies.

[0097] During step S22, based on the understanding of the current power grid situation and scheduling requirements from the domain-wide language model in step S21, this understanding is transformed into a semantic query of the offline-built policy knowledge base. Specifically, using the tagging system of policy metadata in the heuristic optimization model library, multi-dimensional matching is performed within this library to determine a preliminary candidate policy list. This candidate policy list includes the initially selected policies encapsulated as heuristic optimization models.

[0098] That is, in step S22, the understood information is converted into a semantic query for the heuristic optimization model library using the domain large language model, and a preliminary candidate strategy list is determined from the heuristic optimization model library based on the semantic query.

[0099] Step S23: Using the domain large language model, for each candidate strategy in the candidate strategy list, combined with the current distribution network operating parameters, output the adjustment strategy for the key operating parameters of each candidate strategy, and adjust the corresponding heuristic optimization model based on the adjustment strategy to obtain the adapted candidate heuristic optimization model.

[0100] During the execution of step S23, a heuristic optimization action library is defined as follows: Each action For each heuristic optimization model invoked under specific parameter configurations, a set of specific switching operation sequences or equipment adjustment instructions can be generated. This applies to each real-time state of the distribution network. The domain-specific large language model generates N candidate actions based on the current context. The values ​​of N and k are greater than 1.

[0101] This involves determining the adjustment strategy for key operating parameters of each candidate strategy model in the candidate strategy list, taking into account current power grid operating parameters (e.g., precise load values ​​at different nodes and precise impedance parameters of lines). Based on this adjustment strategy, the candidate heuristic optimization model can be more accurately adapted to the current specific power grid operating conditions. This enables online fine-tuning and personalized adaptation of the preset heuristic optimization model.

[0102] S104: Evaluate candidate heuristic optimization models and determine the optimal strategy based on the evaluation results.

[0103] In the specific implementation of S104, a lightweight Monte Carlo parallel simulation was used to simulate the results of different heuristic optimization models performing M operations under the current problem state, and to evaluate the multi-dimensional performance indicators (power outage cost) of the candidate heuristic optimization models. Line loss cost Voltage over-limit cost Complex operation and high cost Based on preset or online learning scheduling preferences, the target weight coefficients are dynamically adjusted and the candidate strategies are ranked according to their expected comprehensive costs. Furthermore, the strategy with the minimum expected comprehensive cost is determined as the optimal strategy through finite-time rolling optimization, which can be specifically reflected in the following finite-time rolling optimization objective function.

[0104] The objective function for finite-time rolling optimization is: at each decision time t, consider the optimization window for the next T scheduling steps. Define the scheduling quality function. To execute the action sequence starting from state s The expected total cost afterward:

[0105]

[0106] in, For state The corresponding average system outage duration contributes to the cost. For state The corresponding network line loss cost, For state The corresponding penalty cost for exceeding the voltage limit, For action The corresponding operational complexity cost (such as the number of switching actions). These are the weighting coefficients for each objective, which can be dynamically adjusted according to scheduling priority.

[0107] For each candidate action Its expected performance was evaluated using lightweight Monte Carlo simulations:

[0108]

[0109] in, From Begin, execute the action first The subsequent T-1 steps use a randomized strategy to generate the decision sequence, where M is the number of simulations.

[0110] S105: Transform the optimal strategy into an executable sequence of scheduling instructions and execute it on-site. Iterate and adjust the sequence of scheduling instructions in a rolling optimization manner until the process termination condition is met, and output the distribution network reconfiguration scheme.

[0111] In the specific execution of S105, the optimal strategy is transformed into an executable instruction sequence for field execution. After the instruction is executed, the distribution network status is immediately re-sensed, and the aforementioned steps are iteratively executed in a rolling optimization manner until the process termination condition is met, and the final distribution network reconfiguration scheme is output.

[0112] S106: Evaluate the power distribution network reconfiguration scheme to obtain a new large language model for training the domain large language model.

[0113] In one embodiment of this application, the specific process of performing S106 includes:

[0114] S1061: Call the evaluation module to calculate the quantitative impact of the distribution network reconfiguration scheme on key indicators of actual operation and generate a quantitative effect report.

[0115] In the specific implementation of S1061, after the distribution network reconfiguration is completed based on the distribution network reconfiguration scheme, the evaluation module is called to calculate the quantitative impact of the distribution network reconfiguration scheme on the key indicators of actual operation and generate a quantitative effect report.

[0116] In this embodiment of the application, after a distribution network reconfiguration task (such as fault handling) is completely completed, the system will call the evaluation module to quickly calculate the actual impact of the scheme on key indicators such as SAIDI (system average outage duration), SAIFI (system average outage frequency), line loss rate, and voltage qualification rate, and generate a comprehensive quantitative effect report.

[0117] S1062: The quantitative effect report, distribution network reconfiguration scheme, and the entire decision trajectory of the distribution network reconfiguration scheme are fed back to the domain big language model. The GRPO reinforcement learning algorithm is used to train the domain big language model by combining the scheduling effect preference reward and the distribution network condition perception reward, so as to obtain a dual-reward driven scheduling decision model.

[0118] Among them, the dual-reward driven scheduling decision model can improve the quality and adaptability of strategy selection when screening and calling heuristic optimization models, enabling it to understand the distribution network status more accurately and balance multi-objective optimization more intelligently, thereby improving the overall performance of the method and realizing the autonomous evolution of system performance.

[0119] In this embodiment, the effect report, the final adopted solution, and the state-action sequence of the entire solution process are integrated and fed back to the large language model for learning. The learning process employs a dual-reward design, with the strategy selecting the model's parameters. Update by maximizing the expected total reward:

[0120]

[0121] in, , For balance coefficient, For Parameterized strategy selection model To provide a reward function that favors scheduling performance, The reward function for distribution network condition perception is defined. Training employs the GRPO reinforcement learning algorithm, which continuously optimizes the algorithm based on collected historical decision trajectory data. This enables the strategy selection model to simultaneously possess the ability to select high-performing strategies and accurately understand the state of the distribution network.

[0122] The dual rewards include: scheduling effect preference reward. and distribution network operating condition perception rewards .

[0123] The scheduling performance preference reward is calculated based on the difference between the actual scheduling results and the simulation prediction results of candidate strategies. If the actual scheduling performance is better than the prediction result of a certain candidate strategy, the model receives a positive reward for that actual scheduling behavior; otherwise, it receives a negative reward. The magnitude of the reward can be scaled according to the magnitude of the actual performance improvement or decline to incentivize the model to learn towards better results. This reward mechanism does not rely on absolute performance values ​​but trains the model to identify better combinations of scheduling schemes through relative comparisons between schemes. The specific implementation process is as follows:

[0124] If the current state is s, optimize the action library as follows: Each action The multidimensional index vector of this scheme was obtained through simulation. Each component represents the reduction in power outage time. Increase in line loss Voltage exceeding limits Operational complexity .

[0125] Define the alternative preference function Representation scheme Superior Solution Degree:

[0126]

[0127] in, For the Sigmoid function, These are weighting coefficients used to reflect scheduling preferences.

[0128] The scheduling performance preference reward is:

[0129]

[0130] In this embodiment, a scheduling effect preference reward is set to enable the large language model to learn to make reasonable trade-offs when multiple objectives conflict. For example, when scheme A can significantly reduce power outage time but slightly increase line loss compared to scheme B, the model can still correctly identify A as better if the preference weights are set reasonably. This relative comparison mechanism is robust to evaluation noise.

[0131] The reward for distribution network condition awareness is as follows: The system reviews the analysis summary of the distribution network state generated by the domain language model during the decision-making process. It compares the domain language model's analysis of the distribution network state at that time with the key state information reconstructed from real data afterward. High rewards are given to behaviors where the domain language model accurately identifies and effectively utilizes real key operating condition information for decision-making; penalties are given for situations where the domain language model omits, misjudges, or fails to fully consider key operating conditions. This reward directly incentivizes the domain language model to accurately understand the distribution network state, decoupling it from the final decision and improving the interpretability of the decision.

[0132] Specifically, define the key operating condition feature set. Each f indicates a different critical operating condition, such as f1 indicating whether it is during peak load period; f2 indicating whether there is a power outage for critical users; f3 indicating whether the equipment load rate exceeds the safety threshold, etc.; the distribution network state feature vector is... The large language model's perception output vector for the current working condition is: The feature weight vector is Weight Representation of features The importance of scheduling decisions.

[0133] Define the perceptual accuracy of features:

[0134]

[0135] in, For indicator functions, For the Sigmoid function, In the formula, is the scaling factor. Variance is the variance of a continuous feature.

[0136] The reward function is:

[0137]

[0138] in, For the collaborative reward coefficient, control the "all-correct bonus" (i.e. By accurately perceiving all features simultaneously, the reward model avoids focusing solely on high-weight features while ignoring others. This reward model forms an accurate state perception before making a decision.

[0139] After executing S1062, a dual-reward-driven scheduling decision model can be trained and generated based on the domain large language model. Then, S103 is executed. In the next process, the corresponding operation is performed based on the regenerated dual-reward-driven scheduling decision model.

[0140] In the large-scale model dynamic heuristic optimization method for distribution network reconfiguration provided in this application embodiment, a closed-loop paradigm of "offline preparation - online solution - continuous optimization" is constructed. In the offline preparation stage, a heuristic optimization model library containing various optimization preferences is built. In the online solution stage, the distribution network status is perceived in real time. A domain-specific large-scale language model capable of understanding the causal relationship between the distribution network status and operating indicators serves as the intelligent scheduling hub, dynamically analyzing the current situation and selecting, evaluating, and calling the most suitable heuristic optimization model from the pre-built model library for iterative solution, outputting the optimal distribution network reconfiguration scheme. In the continuous optimization stage, the executed distribution network reconfiguration scheme undergoes multi-dimensional performance quantification evaluation. Based on the evaluation results, a dual-reward mechanism integrating scheduling effect preference rewards and distribution network condition perception rewards is adopted to reinforce the large-scale language model, optimizing its future model scheduling and decision-making capabilities. Based on this optimization, the domain-specific large-scale language model enhances its ability to select heuristic optimization models and make scheduling decisions in the next distribution network reconfiguration process, achieving autonomous evolution of system performance and realizing the goal of balancing efficiency, accuracy, and reliability in the distribution network reconfiguration process.

[0141] Based on the large-scale dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the above embodiments of this application, step S101 is executed to generate a family of heuristic optimization models containing various different optimization preferences. The family of heuristic optimization models is then annotated with structured metadata and stored to construct a heuristic optimization model library. The specific process is as follows:

[0142] S1011: Constructing a power distribution network simulation environment and seed strategy.

[0143] In S1011, the distribution network simulation environment is used to simulate a benchmark case library containing different test scenario configurations; the seed strategy is a basic optimization rule represented by code or rule description.

[0144] In this embodiment, a highly simulated distribution network environment is constructed. This environment should be able to simulate a baseline case library containing different fault types (such as short circuits and ground faults), different fault locations, different load levels (from normal operation to heavy load and light load), and different topologies (such as radial and ring networks). A basic optimization rule with clear physical meaning (e.g., "prioritize restoring non-faulty power sources closer to the fault point" or "prioritize transferring feeder loads with the highest load rate") is selected as a seed strategy and represented in the form of code or rule description.

[0145] S1012: Run the seed strategy in the distribution network simulation environment to make decisions on the selected test scenario, record the generated decision trajectory and the achieved performance indicators, and construct the basic solution by combining the decision trajectory and performance indicators.

[0146] In S1012, the basic solution is constructed from the decision trajectory and performance results generated by the execution decisions of the selected test scenario in the distribution network simulation environment using the seed strategy.

[0147] In this embodiment, the driving seed strategy makes decisions and executes them in a simulation environment for a specific scenario, recording its decision trajectory (such as a switching operation sequence) and the final performance indicators (such as power restoration time and line loss). The resulting decision trajectory and performance results constitute the basic solution.

[0148] S1013: Generate multiple variant solutions by performing random perturbations based on the basic solution, and select the variant solutions that meet the performance requirements from the multiple variant solutions as the comparison solution.

[0149] In this embodiment, key decision points in the basic solution (e.g., the order of switching operations, the priority of load transfer) are systematically randomly perturbed to generate multiple candidate variant schemes. After these variant schemes are executed in the simulation environment, their performance indicators are compared with those of the basic solution, and the variant schemes that perform better in the key indicators are selected and defined as the comparison solutions.

[0150] S1014: Input the basic solution and the comparative solution into the pre-trained domain big language model. The domain big language model analyzes the key differences between the basic solution and the comparative solution on the decision trajectory, identifies the operation mode and logic that determines the performance improvement, determines the policy improvement principle, and outputs the policy improvement principle in the form of natural language description.

[0151] In this embodiment, the basic solution and the comparative solution (for the same initial state of the distribution network) are taken as input, and their decision-making processes are analyzed and presented to a domain-wide language model. The domain-wide language model is trained to analyze the key differences between the two solutions on their decision trajectories, particularly the operational choices that lead to significant performance improvements. By understanding the physical constraints and optimization objectives of the distribution network operation, the domain-wide language model identifies the operational patterns that lead to performance improvements and their underlying logic, and summarizes abstract policy improvement principles in the form of natural language descriptions.

[0152] S1015: Reconstruct the seed policy based on the policy improvement principle, generate a heuristic optimization model based on the reconstructed seed policy, and return to execute S1012 until a family of heuristic optimization models containing multiple different optimization preferences is generated based on the iterative loop.

[0153] In this embodiment, the extracted strategy improvement principles are used as guidance and fed back to the domain-wide language model. The domain-wide language model modifies, expands, or reorganizes the original code or rule description of the seed strategy to generate a new generation of strategies. These new generation strategies are also considered independent heuristic optimization models. To further verify their effectiveness, the new generation strategies are tested on independent validation sets. If their performance significantly improves compared to the previous generation strategy in key metrics, the new strategy is retained and used as the starting point for the next round of evolution. This iterative process is repeated until a family of strategies containing multiple different optimization preferences is generated. That is, a group of heuristic optimization models containing multiple different optimization preferences is generated.

[0154] S1016: Perform structured metadata annotation and storage on the heuristic optimization model family to build a heuristic optimization model library.

[0155] In the specific implementation of S1016, a corresponding multi-dimensional metadata tag is established for each heuristic optimization model in the heuristic optimization model family. Using a database that supports multi-tag combination queries, each heuristic optimization model and its corresponding multi-dimensional metadata tag are associated and stored to obtain a heuristic optimization model library. The metadata tags include functional tags, performance characteristic tags, applicable operating condition constraints, and interface specifications.

[0156] In this embodiment, a detailed and multi-dimensional metadata tagging system is created for each generated heuristic optimization model. This heuristic optimization model can exist in the form of code files, model checkpoints, or preset parameter configuration sets. The tagging system includes functional tags, performance characteristic tags, applicable operating condition constraints, and interface specifications. Relational databases or graph databases are used to associate and store the heuristic optimization model (strategy model file) itself with its carefully designed metadata tags. This database should support multi-tag-based combined queries (e.g., retrieving strategies with "fault isolation" and "recovery speed priority") to achieve fast and accurate candidate strategy retrieval. Furthermore, a semantic relationship graph between strategies can be constructed to support more complex strategy recommendation, combination, or transfer learning applications.

[0157] Among them, the functional labels are used to clearly define the core distribution network problems that the strategy model can solve, such as "fault isolation and isolation point selection" and "distributed power dispatch and optimization".

[0158] Performance feature labels are used to describe the key behavioral tendencies or optimization objectives that the strategy model exhibits when it is executed, such as "recovery speed priority" or "line loss minimization orientation".

[0159] Applicable operating condition constraints are used to explicitly define the effective application boundaries and preconditions of the strategy model, such as: "applicable to a specific topology type", "requires at least one available backup power source", and "only works best when the load factor is below a certain threshold".

[0160] The interface specification is used to precisely define the input data format required by the model and the output data format.

[0161] Based on the large-scale dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the above embodiments of this application, S102 is executed to acquire multi-source data of the distribution network in real time, perform condition perception and state feature extraction, and generate a natural language context summary describing the current operating status of the distribution network, including:

[0162] S1021: Based on a standardized industrial communication interface, multiple information sources are accessed, and the real-time collected multi-source data is synchronized, cleaned, and fused to obtain fused multi-source data.

[0163] In this embodiment, under online power distribution network dispatching, the system achieves real-time data access with multiple information sources, including SCADA, fault indicators (FI), PMU (Phasor Measurement Unit), distributed power monitoring systems, energy storage monitoring systems, and meteorological information systems, through standardized industrial communication interfaces (such as IEC 61850 and DL / T 860 protocols). The received multi-source data undergoes unified time synchronization, bad data detection, missing value imputation, and outlier cleaning to ensure data accuracy and consistency. Finally, fused multi-source data is obtained.

[0164] The state space is the set of all possible operating states of a distribution network at a given moment. Each state (each specific operating state) within the state space is a set of key variables describing the current state of the distribution network. The action space is the set of all legal and valid operations that can be taken in the current state.

[0165] The state space and action space of a distribution network under online dispatching are defined as follows: Let This represents the distribution network status at time t, including comprehensive information such as topology, load, fault information, and equipment status.

[0166]

[0167] in, The active and reactive load vectors of the nodes. The node voltage magnitude vector. The branch current vector, Indicates the current topology (switch state). Indicates fault information (type, location, and scope of impact).

[0168] S1022: Extract and calculate state feature vectors from the fused multi-source data.

[0169] In S1022, this state feature vector is used to characterize the current operating status of the distribution network. This state feature vector includes fault characteristics, network state characteristics, resource characteristics, and environmental characteristics.

[0170] The fault characteristics include the type of fault, the estimated fault location, and the amount of power loss and the number of affected users assessed based on the fault range; the types of faults include single-phase grounding, phase-to-phase short circuit, etc.

[0171] Network status characteristics include the voltage amplitude and phase angle of each node, the real-time load rate of critical lines and transformers, and the connectivity of the network topology (such as whether islands exist).

[0172] Resource characteristics include the real-time output limit of adjustable distributed power sources, the state of charge (SOC) of energy storage systems, and the available capacity of controllable loads.

[0173] Environmental characteristics include whether there are severe weather warnings and the current natural time of day (such as morning rush hour, evening rush hour, and nighttime off-peak).

[0174] In this embodiment, a series of high-level, abstract state features are extracted and calculated from cleaned and fused multi-source data, and then integrated into a state feature vector describing the current operating status of the power distribution network. .

[0175] S1023: Combine state feature vectors with distribution network topology information, and generate a natural language scenario summary based on the combined information to describe the current operating status of the distribution network.

[0176] In the specific execution of S1023, the extracted state feature vector is combined with the topology single-line diagram information of the distribution network to automatically generate a structured, concise natural language context summary containing key information.

[0177] Based on the large-scale dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the above embodiments of this application, S104 is executed to evaluate candidate heuristic optimization models, and the optimal strategy is determined based on the evaluation results, including:

[0178] S1041: Using the current real distribution network state as the initial condition, simulate and execute the candidate strategies output by the candidate heuristic optimization model in parallel in a lightweight simulation environment to obtain simulation results.

[0179] In S1041, the simulation results include the operational consequences and changes in the distribution network state resulting from the execution of each candidate strategy within a preset future timeframe.

[0180] In this embodiment, a lightweight simulation environment is initiated to evaluate the feasibility and performance of multiple candidate strategies in a short time. This simulation environment uses the current state of the real distribution network as initial conditions and simulates the execution of each parameter-adapted candidate strategy in parallel. The simulation process employs efficient fast power flow calculation and rule-based logical judgment, enabling it to quickly predict the operational consequences and distribution network state changes that may result from the execution of each candidate strategy over a future period.

[0181] S1042: Extract multi-dimensional performance metrics for each candidate strategy from the simulation results.

[0182] In S1042, the multi-dimensional performance metrics include: expected power recovery time, system bus loss, minimum node voltage, maximum device load rate, and switching operation complexity.

[0183] In this embodiment of the application, after the simulation is completed, a series of multi-dimensional performance indicators are extracted from the simulation results of each candidate strategy. These multi-dimensional performance indicators can comprehensively reflect the reconfiguration effect, including but not limited to: the expected power recovery time (SAIDI contribution), the bus loss caused, the minimum node voltage, the maximum device (line / transformer) load rate, and the complexity of operation (e.g., the total number of switching actions required).

[0184] S1043: Input multi-dimensional performance indicators into the multi-objective trade-off model, perform comprehensive utility scoring and ranking based on the preset or online learned weights in the multi-objective trade-off model, and obtain the output ranking list.

[0185] In S1043, the weights are used to reflect the scheduling priorities under different operating scenarios.

[0186] In this embodiment, the extracted multi-dimensional performance indicators are input into a pre-defined or online-learned multi-objective trade-off model. This model incorporates the actual preferences and priorities of scheduling operations (e.g., in major failure scenarios, the weight of "recovery time" is much higher than "line loss"; while in routine optimization scenarios, "economic efficiency" may be the dominant factor). Based on these weights, the model performs a comprehensive utility score on each candidate strategy and generates a ranking list accordingly, clearly defining the optimal strategy and its backup options.

[0187] S1044: Based on the finite-time-domain rolling optimization objective function, evaluate the expected comprehensive cost of candidate strategies in the ranking list, and determine the candidate strategy with the minimum expected comprehensive cost as the optimal strategy.

[0188] In this embodiment, candidate strategies are evaluated and weighed against current operational priorities before any scheduling instruction is executed to ensure that the final selected strategy is not only feasible but also optimal. Based on this, simulation evaluation and preference-based ranking provide a solid scientific basis for online scheduling decisions.

[0189] Based on the large-scale dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the above embodiments of this application, step S105 converts the optimal strategy into an executable sequence of scheduling instructions for on-site execution, and iteratively adjusts the scheduling instruction sequence in a rolling optimization manner until the process termination condition is met, outputting a distribution network reconfiguration scheme, including:

[0190] S1051: With an optimization window length of T and the current distribution network fault state as the initial state, the optimal strategy is transformed into a sequence of scheduling instructions containing specific executable operations, and then sent to the field execution equipment to execute the sequence of scheduling instructions.

[0191] In S1051, the scheduling instruction sequence includes switching operations.

[0192] In this embodiment of the application, with an optimization time window length of T and the current distribution network fault state as the initial state, the optimal strategy is... This is converted into an executable sequence of scheduling instructions (e.g., disconnecting switch S1, closing switch S2, adjusting the tap of transformer T1 to the X position).

[0193] These instructions are issued to the field execution equipment via a standard dispatching instruction ticket system or directly through the automatic control interface. After the instructions are issued, the system will continuously monitor the execution status of the instructions, the feedback information from the equipment, and the real-time response of the distribution network.

[0194] S1052: After the first scheduling instruction sequence is completed, execute S102 to obtain the new state of the distribution network as the new initial state, and update the remaining optimization window length to T-1.

[0195] S1053: Based on the new initial state and the remaining optimization window length, the optimal strategy is re-determined by combining the finite time domain rolling optimization objective function. Then, the process returns to step S1051 until the process termination condition is met, and the distribution network reconfiguration scheme is output.

[0196] In S1053, the process termination conditions include the new initial state being a non-faulty state, or the operating indicators meeting a preset level, or the optimization window length being 0.

[0197] In this embodiment of the application, after the first batch of dispatch instructions is executed, the operating status of the distribution network changes from... Transfer to Where T is the state transition function, describing the action to be performed. The physical changes in the state of the distribution network are then detected. The system then immediately returns to S102 to re-sensor the latest state of the current distribution network. Then, based on the latest status... Given the new initial state, the remaining optimization window is updated to T=T-1, and the optimization problem is reconstructed, i.e., the new expected total cost:

[0198]

[0199] If the optimal strategy is selected, and the new state matches the expectation and no adjustment is needed, the subsequent steps of the original strategy are continued; if deviations are detected (such as equipment failure to operate or output fluctuations) or new optimization opportunities are discovered, a new round of decision-making is initiated based on this.

[0200] Repeat the above steps until the termination condition is met, such as the fault being eliminated, the operating indicators reaching a satisfactory level, or the remaining decision steps being exhausted.

[0201] In this embodiment of the application, through this closed-loop mechanism of "perception-decision-execution-re-perception", the system can realize continuous monitoring and dynamic adjustment of the distribution network status, forming a rolling optimization process until the fault is completely eliminated, the distribution network returns to normal operation, or reaches the preset satisfactory operating state.

[0202] Based on the large-scale dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the embodiments of this application, the embodiments of this application also disclose a solver, such as... Figure 2 As shown, the solver includes: an offline construction module 201, an online dynamic scheduling solver module 202, and an offline closed-loop learning optimization module 203.

[0203] Offline build module 201 is used to generate a family of heuristic optimization models containing various optimization preferences, and to build a heuristic optimization model library based on this family of heuristic optimization models.

[0204] The online dynamic scheduling solution module 202 is used to acquire multi-source data of the distribution network in real time, perform condition perception and state feature extraction, and generate a natural language scenario summary describing the current operating status of the distribution network. Based on the natural language scenario summary and scheduling requirements, it uses a domain-wide language model to select and call suitable candidate heuristic optimization models from the heuristic optimization model library. The domain-wide language model is trained based on a general-purpose language model. The candidate heuristic optimization models are evaluated, and the optimal strategy is determined based on the evaluation results. The optimal strategy is transformed into an executable scheduling instruction sequence and executed on-site. The scheduling instruction sequence is iteratively adjusted in a rolling optimization manner until the process termination condition is met, and the distribution network reconfiguration scheme is output.

[0205] The offline closed-loop learning optimization module 203 is used to evaluate the power distribution network reconfiguration scheme and obtain a new general-purpose large language model for training the domain large language model.

[0206] In the embodiments of this application, the specific implementation methods and principles of the offline construction module 201, the online dynamic scheduling solution module 202, and the offline closed-loop learning optimization module 203 are described in the preceding description of the corresponding execution steps.

[0207] like Figure 3 As shown, based on the large-scale model dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the above embodiments of this application, this application also discloses a solver. The electronic device includes at least one processor 300, and at least one memory 301 and a bus 302 connected to the processor 300. The processor 300 and the memory 301 communicate with each other through the bus 302. The processor 300 is used to call the program instructions in the memory 301 to execute the large-scale model dynamic invocation heuristic optimization method for distribution network reconfiguration disclosed in the aforementioned embodiments of this application.

[0208] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and apparatus embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0209] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0210] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic heuristic optimization method for large-scale power distribution network reconfiguration, characterized in that, The method includes: Generate a family of heuristic optimization models containing various optimization preferences, and construct a heuristic optimization model library based on the family of heuristic optimization models; Real-time acquisition of multi-source data from the distribution network, performance of condition perception and status feature extraction, and generation of natural language contextual summaries describing the current operating status of the distribution network; Based on the natural language context summarization and scheduling requirements, a suitable candidate heuristic optimization model is selected and called from the heuristic optimization model library using a domain-wide large language model. Evaluate the candidate heuristic optimization models and determine the optimal strategy based on the evaluation results; The optimal strategy is transformed into an executable sequence of scheduling instructions and executed on-site. The sequence of scheduling instructions is iteratively adjusted in a rolling optimization manner until the process termination condition is met, and the power distribution network reconfiguration scheme is output. The proposed power distribution network reconfiguration scheme is evaluated to obtain a new large language model for training the domain large language model.

2. The method according to claim 1, characterized in that, The power distribution network reconfiguration scheme is evaluated to obtain a new large language model for training the domain large language model, including: The evaluation module is invoked to calculate the quantitative impact of the power distribution network reconfiguration scheme on key indicators of actual operation, and a quantitative effect report is generated. The quantitative effect report, the distribution network reconfiguration scheme, and the entire decision trajectory of executing the distribution network reconfiguration scheme are fed back to the large language model. The large language model is then trained using scheduling effect preference reward and distribution network condition perception reward to obtain a new large language model for training the domain large language model.

3. The method according to claim 1, characterized in that, A heuristic optimization model library is constructed based on the aforementioned family of heuristic optimization models, including: For each heuristic optimization model in the heuristic optimization model family, a corresponding multi-dimensional metadata tag is established. The metadata tag includes functional tags, performance characteristic tags, applicable operating condition constraints, and interface specifications. By using a database that supports multi-tag combined queries, each heuristic optimization model and its corresponding multi-dimensional metadata tags are associated and stored to obtain a heuristic optimization model library.

4. The method according to claim 1, characterized in that, The training process of the domain-specific large language model includes: Construct a professional corpus containing distribution network professional corpus, and an instruction-output pair dataset containing causal reasoning for complex scheduling scenarios. The distribution network professional corpus includes at least one or more of the following: distribution network operation procedures, technical manuals, operation tickets, work tickets, historical fault handling reports, real-time operation logs, and related academic literature. Using the aforementioned professional corpus as training data, a domain-adaptive pre-training of a general large language model is performed using a self-supervised learning task. This enables the pre-trained large language model to recognize entities, operational terms, and key operating indicators specific to the power distribution network domain, as well as the interrelationships among these entities, operational terms, and key operating indicators. By using the instruction-output pair dataset and comparative learning samples designed based on the instruction-output pair dataset, the reasoning ability of the pre-trained large language model for the causal relationship between power distribution network operations and results is supervised and fine-tuned, thus obtaining a domain-specific large language model.

5. The method according to claim 1, characterized in that, Real-time acquisition of multi-source data from the distribution network, performance condition sensing and status feature extraction, and generation of natural language contextual summaries describing the current operating status of the distribution network, including: Based on a standardized industrial communication interface, multiple information sources are accessed, and the real-time collected multi-source data is synchronized, cleaned, and fused to obtain fused multi-source data. State feature vectors are extracted and calculated from the fused multi-source data. These state feature vectors are used to characterize the current operating status of the distribution network and include fault features, network status features, resource features, and environmental features. The state feature vector is combined with the distribution network topology information, and a natural language context summary describing the current operating status of the distribution network is generated based on the combined information.

6. The method according to claim 1, characterized in that, Based on the aforementioned natural language context summarization and scheduling requirements, a domain-wide large language model is used to select and call suitable candidate heuristic optimization models from the heuristic optimization model library, including: The natural language scenario summary and scheduling instructions are input into the domain big language model, and the current core task, operational constraints and user expected optimization preferences are analyzed and understood in the domain big language model to obtain understanding information; The understanding information is converted into a semantic query for the heuristic optimization model library using the domain-specific large language model. Based on the semantic query, a preliminary candidate strategy list is determined from the heuristic optimization model library. The candidate strategy list includes the preliminary strategy encapsulated as a heuristic optimization model. Using the domain-specific large language model, for each candidate strategy in the candidate strategy list, combined with the current distribution network operating parameters, an adjustment strategy for the key operating parameters of each candidate strategy is output. Based on the adjustment strategy, the corresponding heuristic optimization model is adjusted to obtain an adapted candidate heuristic optimization model.

7. The method according to claim 1, characterized in that, Evaluate the candidate heuristic optimization models and determine the optimal strategy based on the evaluation results, including: Using the current real distribution network state as the initial condition, the candidate strategies output by the candidate heuristic optimization model are simulated and executed in parallel in a lightweight simulation environment to obtain simulation results. The simulation results include the operational consequences and distribution network state changes produced by executing each candidate strategy within a preset future time. Extract multi-dimensional performance metrics for each candidate strategy from the simulation results. These multi-dimensional performance metrics include: expected power recovery time, system bus loss, minimum node voltage, maximum device load rate, and switching operation complexity. The multi-dimensional performance indicators are input into the multi-objective trade-off model. Based on the weights preset or learned online in the multi-objective trade-off model, a comprehensive utility score and ranking are performed to obtain the output ranking list. The weights are used to reflect the scheduling priority under different operating scenarios. Based on the finite-time rolling optimization objective function, the expected comprehensive cost of the candidate strategies in the sorted list is evaluated, and the candidate strategy with the minimum expected comprehensive cost is determined as the optimal strategy.

8. The method according to claim 1, characterized in that, The optimal strategy is transformed into an executable sequence of scheduling instructions for on-site execution. The sequence of scheduling instructions is iteratively adjusted using a rolling optimization approach until the process termination condition is met, outputting a distribution network reconfiguration scheme, including: With an optimization time window length of T and the current distribution network fault state as the initial state, the optimal strategy is transformed into a sequence of scheduling instructions containing specific executable operations and sent to the field execution equipment, so that the field execution equipment executes the sequence of scheduling instructions, which includes switching operations. After the first scheduling instruction sequence is executed, the step of acquiring multi-source data of the distribution network in real time, performing condition perception and state feature extraction is executed to obtain the new state of the distribution network as the new initial state, and updating the remaining optimization time window length to T-1. Based on the new initial state and the remaining optimization window length, the optimal strategy is re-determined by combining the finite-time-domain rolling optimization objective function. The process then returns to the step of converting the optimal strategy into a sequence of scheduling instructions containing specific executable operations and issuing them to the field execution equipment, until the process termination condition is met, and the distribution network reconfiguration scheme is output. The process termination condition includes the new initial state being a non-fault state, or the operating indicators meeting a preset level, or the optimization window length being 0.

9. A solver, characterized in that, The solver includes: An offline building module is used to generate a family of heuristic optimization models containing various optimization preferences, and to build a heuristic optimization model library based on the family of heuristic optimization models. The online dynamic scheduling solution module is used to acquire multi-source data of the distribution network in real time, perform condition perception and state feature extraction, and generate a natural language scenario summary describing the current operating status of the distribution network. Based on the natural language scenario summary and scheduling requirements, it uses a domain-wide language model to select and call suitable candidate heuristic optimization models from the heuristic optimization model library. The candidate heuristic optimization models are evaluated, and the optimal strategy is determined based on the evaluation results. The optimal strategy is transformed into an executable scheduling instruction sequence and executed on-site. The scheduling instruction sequence is iteratively adjusted in a rolling optimization manner until the process termination condition is met, and the distribution network reconfiguration scheme is output. An offline closed-loop learning optimization module is used to evaluate the power distribution network reconfiguration scheme and obtain a new general-purpose large language model for training the domain large language model.

10. A solver, characterized in that, The solver includes at least one processor, and at least one memory and bus connected to the processor; wherein the processor and the memory communicate with each other through the bus; the processor is used to call program instructions in the memory to execute the large-scale dynamic calling heuristic optimization method for power distribution network reconfiguration as described in any one of claims 1 to 8.

Images