Dynamic power distribution network reconstruction method and system under large model enhancement, and medium

Abstract

The application discloses a dynamic power distribution network reconstruction method and system under a large model enhancement, relates to the technical field of power distribution network reconstruction, and comprises the following steps: acquiring operation state quantities of a dynamic power distribution network at a current time and historical time periods, and load and distributed energy prediction data of a future preset time period; predicting a short-term electrical state evolution trajectory of the dynamic power distribution network; generating a natural language reconstruction strategy summary in a continuous time dimension; providing gradient information provided by a power grid dynamic representation unit as a physical dynamic constraint of a differentiable optimization problem; and completing the reconstruction of the dynamic power distribution network by executing the switch action sequence. The application solves the technical problems that the existing technology has rough power distribution network reconstruction timing planning, lacks semantic guidance for decision-making, and is difficult to adapt to the continuous dynamic evolution law of the power grid, and achieves the technical effects of realizing continuous time domain dynamic power distribution network timing reconstruction under large model guidance, and improving the decision-making rationality and operation stability of power distribution network reconstruction.

Description

Technical Field

[0001] This invention relates to the field of distribution network reconfiguration technology, specifically to a dynamic distribution network reconfiguration method, system, and medium under large model enhancement. Background Technology

[0002] Current distribution network reconfiguration mostly adopts static optimization methods with discrete time sections, relying on traditional mathematical programming algorithms to solve the timing of switch actions. This makes it difficult to adapt to the continuous timing evolution characteristics of the power grid caused by the high penetration of distributed generation and dynamic load fluctuations. At the same time, existing reconfiguration strategies lack high-level semantic decision guidance and do not make sufficient use of implicit intentions such as scheduling experience, action dependencies, and timing preferences. This can easily lead to problems such as unreasonable topology adjustment timing, single optimization objectives, and insufficient consideration of dynamic constraints. As a result, the reconfiguration schemes have poor practicality and slow convergence speed, making it difficult to meet the actual operational needs of refined and continuous timing reconfiguration of dynamic distribution networks.

[0003] Existing technologies suffer from technical problems such as coarse timing planning for distribution network reconfiguration, lack of semantic guidance in decision-making, and difficulty in conforming to the continuous dynamic evolution of the power grid. Summary of the Invention

[0004] This application provides a method, system, and medium for dynamic distribution network reconfiguration under large model enhancement, which is used to address the technical problems of existing technologies, such as coarse timing planning for distribution network reconfiguration, lack of semantic guidance in decision-making, and difficulty in conforming to the continuous dynamic evolution of the power grid.

[0005] In view of the above problems, this application provides a method, system and medium for dynamic distribution network reconfiguration under large model enhancement.

[0006] A first aspect of this application provides a dynamic distribution network reconfiguration method under large model enhancement, the method comprising: The system acquires the operational status variables of the dynamic distribution network at the current moment and in historical time periods, as well as load and distributed energy prediction data for a preset future time period. It inputs these operational status variables and the load and distributed energy prediction data into a power grid dynamic characterization unit. Based on the neural differential equation network built into the power grid dynamic characterization unit, it predicts the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain. Finally, it inputs the current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into a large language model decision guidance unit to generate a natural language reconstruction strategy overview in the continuous time dimension. The natural language reconstruction strategy summary includes the reconstruction objective, the expected action sequence, and its correlation with changes in the power grid state. The natural language reconstruction strategy summary is compiled into a differentiable optimization problem containing weight coefficients and conditional constraints, and the gradient information provided by the power grid dynamic characterization unit is used as the physical dynamic constraint of the differentiable optimization problem. In the differentiable optimization solution unit, the differentiable optimization problem is solved, and a switching action sequence defined on a continuous time axis is output. By executing the switching action sequence, the reconstruction of the dynamic distribution network is completed. The switching action sequence includes the execution time and action content of each switching action.

[0007] A second aspect of this application provides a dynamic distribution network reconfiguration system under large model enhancement, the system comprising: The system includes: an operational state acquisition module for acquiring operational state data of the dynamic distribution network at the current time and in historical periods, as well as load and distributed energy prediction data for a preset future period; an electrical state evolution trajectory prediction module for inputting the operational state data and the load and distributed energy prediction data into a power grid dynamic characterization unit, and predicting the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain based on the neural differential equation network built into the power grid dynamic characterization unit; and a natural language reconstruction strategy summary generation module for inputting the current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into a large language model decision guidance unit to generate a continuous time dimension... The natural language reconstruction strategy summary includes the reconstruction objective, the expected action sequence, and its correlation with changes in the power grid state. A physical dynamic constraint generation module is used to compile the natural language reconstruction strategy summary into a differentiable optimization problem containing weight coefficients and conditional constraints, and uses the gradient information provided by the power grid dynamic representation unit as the physical dynamic constraints of the differentiable optimization problem. A dynamic distribution network reconstruction module is used to solve the differentiable optimization problem in the differentiable optimization solution unit, output a switching action sequence defined on a continuous time axis, and complete the reconstruction of the dynamic distribution network by executing the switching action sequence, wherein the switching action sequence includes the execution time and action content of each switching action.

[0008] In a third aspect of this application, a computer-readable storage medium is provided storing a computer program for executing the dynamic distribution network reconfiguration method under large model enhancement provided in this application.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: The system acquires the operational status variables of the dynamic distribution network at the current moment and in historical time periods, as well as load and distributed energy prediction data for a preset future time period. These operational status variables and the predicted load and distributed energy data are input into a power grid dynamic characterization unit. Based on the neural differential equation network built into the power grid dynamic characterization unit, the system predicts the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain. The current topology, topology violation diagnosis results, the predicted load and distributed energy data, and the short-term electrical state evolution trajectory are then input into a large language model decision guidance unit to generate a natural language reconstruction strategy summary in the continuous time dimension. The gradient information provided by the power grid dynamic characterization unit is used as the physical dynamic constraint of the differentiable optimization problem. A switching action sequence defined on the continuous time axis is output, and the reconstruction of the dynamic distribution network is completed by executing the switching action sequence. This achieves the technical effect of realizing continuous-time domain dynamic distribution network time-series reconstruction under the guidance of a large model, improving the decision rationality and operational stability of distribution network reconstruction. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating the dynamic distribution network reconfiguration method under large model enhancement provided in this application embodiment.

[0012] Figure 2 A schematic diagram of the structure of the dynamic distribution network reconfiguration system under large model enhancement provided in the embodiments of this application.

[0013] Figure labeling: Module 10 for acquiring operating state quantities, Module 20 for predicting electrical state evolution trajectory, Module 30 for generating a summary of natural language reconstruction strategy, Module 40 for generating physical dynamic constraints, and Module 50 for dynamic distribution network reconstruction. Detailed Implementation

[0014] This application provides a dynamic distribution network reconfiguration method, system, and medium under large model enhancement, which addresses the technical problems of existing technologies, such as coarse timing planning for distribution network reconfiguration, lack of semantic guidance in decision-making, and difficulty in conforming to the continuous dynamic evolution of the power grid.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] Example 1, as Figure 1 As shown, this application provides a dynamic distribution network reconfiguration method under large model enhancement, the method comprising: Step S100: Obtain the operating status of the dynamic distribution network at the current time and in historical periods, as well as the load and distributed energy forecast data for future preset periods.

[0017] Specifically, the system collects real-time operating status data such as voltage, current, and power of the dynamic distribution network through the distribution network monitoring system interface, retrieves archived operating status data from historical periods, and connects to the load forecasting and distributed energy forecasting systems to obtain load forecasting data and distributed energy output forecasting data for future preset periods, providing complete multi-dimensional basic data for power grid dynamic modeling and reconfiguration decisions.

[0018] Step S200: Input the operating state variables and the load and distributed energy prediction data into the power grid dynamic characterization unit, and predict the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain based on the neural differential equation network built into the power grid dynamic characterization unit.

[0019] Specifically, the power grid dynamic characterization unit receives operating state variables and load and distributed energy prediction data. It extracts grid node voltage, branch current and power from the operating state variables as initial state variables of the neural differential equation network. Based on the future power distribution given by the prediction data and combined with the physical laws of the power grid, it infers the future switching state change trend and quantifies it into a hypothetical control input sequence. The neural differential equation network takes this initial state as the starting point and the hypothetical control input as the driving force, and carries out forward simulation based on the embedded continuous dynamics model. Finally, it calculates the future change path of the short-term electrical state in the continuous time domain and obtains the short-term electrical state evolution trajectory of the dynamic distribution network.

[0020] Step S300: Input the current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into the large language model decision guidance unit to generate a natural language reconstruction strategy summary in the continuous time dimension. The natural language reconstruction strategy summary includes the reconstruction target, the expected action sequence, and its correlation with the changes in the power grid state.

[0021] Specifically, the real-time status of switches is obtained from the distribution network monitoring system to determine the current topology. Connectivity and circulation analysis are performed on the topology to obtain topology violation diagnosis results. The current topology, topology violation diagnosis results, load and distributed energy forecast data, and short-term electrical state evolution trajectory are then integrated into structured input prompt information. Subsequently, the prompt information is enhanced to identify key switches and network areas with violations, locate sensitive periods of electrical state changes, and complete annotation and guidance instruction addition. At the same time, operator natural language scheduling instructions and operating preferences can be incorporated as hard constraints and input into the large language model decision guidance unit for reasoning. Finally, a continuous time-dimensional natural language reconstruction strategy summary is generated, which includes the reconstruction target, expected action sequence, and its correlation with power grid state changes.

[0022] Step S400: Compile the natural language reconstruction strategy summary into a differentiable optimization problem containing weight coefficients and conditional constraints, and use the gradient information provided by the power grid dynamic characterization unit as the physical dynamic constraint of the differentiable optimization problem.

[0023] Specifically, the natural language reconstruction strategy summary is semantically parsed by a pre-defined strategy compilation module to identify the reconstruction goal, action description, state association, and constraint information. The reconstruction goal is quantified into the objective function of a differentiable optimization problem according to predefined mapping rules. The action description and state association are transformed into equality and inequality constraints, and corresponding weight coefficients are set according to the semantic strength of the constraints. At the same time, the gradient information of the state variables output by the neural differential equation network with respect to the switching action is obtained from the power grid dynamic representation unit and added as a physical dynamic constraint to the differentiable optimization problem to complete the construction of a complete optimization model.

[0024] Step S500: In the differentiable optimization solution unit, solve the differentiable optimization problem, output the switching action sequence defined on the continuous time axis, and complete the reconstruction of the dynamic distribution network by executing the switching action sequence. The switching action sequence includes the execution time and action content of each switching action.

[0025] Specifically, in the differentiable optimization solution unit, the execution time and action content of the candidate switch action sequence are iteratively adjusted using the differentiable characteristics of the differentiable optimization problem. After each adjustment, the differentiable forward simulation function of the power grid dynamic characterization unit is called to evaluate the system performance and constraint satisfaction, and the gradient of the objective function and constraints with respect to the switch action is calculated to guide the iterative optimization. At the same time, a secondary analysis is performed on the natural language reconstruction strategy summary to extract implicit guidance information such as action timing, time window preference and dependency relationship, and to construct a switch action timing reference trajectory. Minimizing the deviation between the candidate sequence and the reference trajectory is used as an auxiliary optimization objective or regularization term to guide the solution to converge quickly. After the preset convergence condition is met, a continuous time axis switch action sequence containing accurate execution time and action content is output. By executing this switch action sequence, the safe and stable reconstruction of the dynamic distribution network is finally completed.

[0026] In one possible implementation, step S200 further includes: Step S210: The power grid dynamic characterization unit receives the operating status data and the load and distributed energy prediction data.

[0027] Step S220: Extract the current voltage, branch current and power of the power grid node from the operating state variables, and use them as the initial state variables of the neural differential equation network.

[0028] Step S230: Based on the future power distribution indicated by the load and distributed energy prediction data, and in conjunction with the laws of power grid physics, deduce the switching state change trend of the distribution network topology in the future period, and quantify the switching state change trend into the hypothetical control input sequence of the neural differential equation network.

[0029] Step S240: In the power grid dynamic characterization unit, the neural differential equation network takes the hypothetical control input sequence as its control input and the initial state variable as the starting point of its state variables, and performs forward simulation based on the embedded continuous dynamic model.

[0030] Step S250: Through the forward simulation, the future change path of the electrical state of the power grid in continuous time under the action of the hypothetical control input sequence is calculated, and the short-term electrical state evolution trajectory of the dynamic distribution network is generated.

[0031] Specifically, the power grid dynamic characterization unit receives dynamic distribution network operating status data obtained from previous steps, as well as load and distributed energy prediction data for a future preset period, through a data interface, providing complete input data for the neural differential equation network to carry out continuous time domain power grid electrical state evolution modeling.

[0032] Feature extraction is performed on the acquired operating state variables to select the current voltage of each node in the power grid, the current of each branch, and the power flow data. These key electrical quantities, which can intuitively reflect the real-time operating conditions of the power grid, are then organized and used as the initial state variables for state deduction in the neural differential equation network.

[0033] A machine learning algorithm combining a physically embedded graph neural network and a time-series prediction model is employed. The model takes the power grid topology as input, with node features including predicted load and distributed energy power, historical voltage and current, and edge features including line impedance and connectivity. Spatial topological correlation features are extracted through graph convolutional layers, and then the temporal evolution of future power distribution is learned by combining a Transformer temporal coding layer. During training, historical switch state sequences and power flow physics equations (Kirchhoff's laws) are used as supervision signals and constraints to complete end-to-end training. When in use, real-time load and distributed energy prediction data are input into the trained model, and forward inference is performed using power grid physics laws to directly output the probability of state changes and action trends of each switch in the future time period. Finally, through temporal normalization and vector coding layers, the trend results are quantified into numerical vectors in the continuous time dimension, forming a hypothetical control input sequence adapted to the neural differential equation network.

[0034] In the power grid dynamic characterization unit, the neural differential equation network uses a hypothetical control input sequence as the control driver and the initial state variables as the time starting point. Through the embedded continuous dynamic model, it performs continuous time domain state recursive evolution of the power grid electrical quantities. During the simulation process, it updates the state variables such as node voltage, branch current and power at each time step. It advances the calculation according to the numerical integration method of differential equations, fully characterizes the dynamic process of the power grid operating state changing smoothly over time, and achieves high-precision forward simulation.

[0035] Based on the continuous-time forward simulation results of the neural differential equation network, the continuous change process of key electrical quantities such as grid node voltage, branch current, and power flow under the hypothetical control input sequence is calculated point by point along the time axis. The smooth evolution trend of the distribution network electrical state in the future period is fully restored, and the short-term electrical state evolution trajectory containing timestamps and corresponding electrical quantity values ​​is finally output.

[0036] In one possible implementation, step S300 further includes: Step S310: Obtain the real-time status of the switches from the power distribution network monitoring system to determine the current topology.

[0037] Step S320: Perform connectivity and circulation analysis on the current topology to generate topology violation diagnosis results.

[0038] Step S330: Integrate the current topology, the topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into structured input prompt information.

[0039] Step S340: Input the input prompt information into the large language model decision guidance unit, reason and output a summary of the natural language reconstruction strategy in the continuous time dimension.

[0040] Specifically, through the real-time data interface of the distribution network monitoring system SCADA / DMS, the real-time open and closed status of all switches, tie switches, and sectional switches in the substation and feeders is collected. Based on the switch connection relationship and line connection relationship, the topological connection mode of nodes, branches, and power supply points is automatically identified to form the complete current topology of the distribution network at the current moment.

[0041] Based on the established topology, a breadth-first search algorithm is first used to traverse the entire network connectivity and identify isolated areas that are not powered by the main power source. Then, for closed loops formed in the topology, forward and backward power flow calculations are used to iteratively calculate the voltage, branch power, and current distribution of each node in the loop. This verifies whether abnormal circulating currents, power offsets, and voltage over-limit issues occur in the closed loops. At the same time, it verifies whether the number of switches on and off matches the topology connection logic, and investigates issues such as inconsistent switch states and contradictory topology logic. Finally, violation information such as isolated areas, abnormal loops, abnormal circulating currents, and mismatched number of switches is compiled and labeled to generate a complete topology violation diagnosis result.

[0042] The current topology, topology violation diagnosis results, load and distributed energy forecast data, and short-term electrical state evolution trajectory are standardized and formatted. The topology is converted into a node-branch connection matrix. The topology violation diagnosis results are extracted into structured labels for violation location, type, and severity. The load and distributed energy forecast data are organized into time-series power data. The short-term electrical state evolution trajectory is sorted into a dynamic change sequence of node voltage and branch power along a continuous time axis. Then, it is hierarchically spliced ​​according to the prompt words that can be recognized by the large language model. Finally, it is integrated to generate structured input prompt information containing spatial topology information, violation constraint information, future power information, and dynamic electrical evolution information.

[0043] The structured input prompts are enhanced by first parsing the topology violation diagnosis results within the information to accurately identify the key switches and corresponding network regions that cause the violations. Then, by combining the short-term electrical state evolution trajectory, the sensitive periods of drastic fluctuations in electrical quantities around the key switches and network regions are located. Subsequently, the key switches, network regions, and sensitive periods are explicitly marked in the input prompts, and corresponding strategy planning guidance instructions are added, requiring priority to design corrective actions for key locations during sensitive periods. The enhanced input prompts are then input into the large language model decision guidance unit. This unit performs temporal reasoning based on the marked content and guidance instructions, and finally outputs a natural language reconstruction strategy summary with continuous time dimensions that specifically addresses topology violations.

[0044] In one possible implementation, step S340 further includes: Enhancement processing of the input prompt information includes: a. Analyze the topology violation diagnosis results in the input prompt information to identify the key switches or network areas that caused the violation.

[0045] b. Based on the short-term electrical state evolution trajectory in the input prompt information, locate the sensitive period of drastic changes in electrical state near the critical switch or network area.

[0046] c. In the input prompt information, explicitly mark the key switch or network area and the sensitive time period, and add guiding instructions to indicate that when generating the strategy summary, priority should be given to planning correction actions for the key switch or network area during the sensitive time period.

[0047] The enhanced input prompts are input into the large language model decision guidance unit, which then performs inference based on the annotations and guidance instructions in the enhanced input prompts to generate a summary of natural language reconstruction strategies in the continuous time dimension.

[0048] Specifically, the integrated input prompt information is first enhanced by using regular expressions and keyword extraction algorithms to parse the text information such as violation type, fault location, and abnormal branch in the topology violation diagnosis results of the input prompt information. Then, the distribution network node-switch-branch topology association matrix is ​​retrieved, and the parsed abnormal location is spatially mapped and matched with the topology matrix. The key switching equipment that directly causes islanding, ring network, and switch status contradictions is located by tracing back along the electrical connection path. At the same time, the power supply area and network area affected by the switch are delineated to achieve accurate identification of the violation cause.

[0049] Continuous time-series data of voltage, branch current, and active power of key switches and corresponding network area surrounding nodes are extracted from the short-term electrical state evolution trajectory. The rate of change of electrical quantities is calculated at each time step. The variance of electrical quantity fluctuations in each time period is statistically analyzed by sliding window. Time periods with a rate of change exceeding a preset threshold and a significantly large variance are selected to accurately locate sensitive periods in the region where the electrical state fluctuates drastically and the operation risk is high.

[0050] By using text tagging tools, key switches, network regions, and sensitive time periods identified in the original input prompts are explicitly marked with unique highlight labels. Then, according to the instruction paradigm that the large language model can recognize, customized constraint guidance instructions are embedded. This explicitly requires the model to prioritize the marked key switches and network regions when generating a reconstruction strategy summary in subsequent inferences, and to design topology correction actions and timing control schemes within the corresponding sensitive time periods, thereby completing the targeted enhancement of the input prompts.

[0051] The large language model decision guidance unit is built around a fine-tuned prompt-enhanced temporal reasoning large language model. It adopts an encoder-decoder structure and relies on a pre-trained large language model base. It undergoes supervised fine-tuning through distribution network topology reconstruction samples, violation handling cases, and temporal scheduling rules. An internal attention enhancement layer strengthens the weighting of labels for key switches, network areas, and sensitive time periods. During use, the enhanced input prompt information, with completed annotations and instruction embedding, is input into the model. The model prioritizes the capture of annotation information and guidance instructions through a multi-head attention mechanism, combining power grid safety constraints and power flow temporal change patterns to perform multi-step temporal reasoning. It plans switch control actions and topology correction logic for each time period according to the continuous time dimension, ultimately outputting a temporally relevant and targeted natural language reconstruction strategy summary.

[0052] In one possible implementation, step S340 further includes: Obtain scheduling instructions or running preferences entered by operators in natural language.

[0053] The scheduling instructions or running preferences are parsed to extract time window requirements, operation prohibition switches, or expected optimization goals.

[0054] The time window requirements, prohibited operation switches, or desired optimization objectives are incorporated as additional hard constraints into the input prompts of the large language model decision guidance unit, so as to guide the generated natural language reconstruction strategy summary to conform to the on-site scheduling intent.

[0055] Specifically, the system receives natural language dispatch instructions and operation preference text input by operators in real time through a human-computer interaction interface. Combined with a speech-to-text module that is compatible with voice input, it uses text streaming acquisition and format standardization processing to convert various input formats into parsable structured text data, thus completing the acquisition of dispatch instructions and operation preference information.

[0056] A BERT-based fine-tuned power grid dispatch semantic parsing model is employed for extraction. This model, built upon a pre-trained BERT model, utilizes an encoder-decoder structure. The encoder, composed of multiple bidirectional Transformers, semantically encodes dispatch instructions and operational preferences, capturing textual context and domain-specific semantics. The decoder incorporates Named Entity Recognition (NER) and Intent Classification (ICC) headers. The NER header identifies and extracts entity information such as time windows and prohibited operation switches, while the ICC identifies desired optimization objectives such as minimum network loss, voltage stability, and load balancing. A domain-specific dictionary embedding layer is added to enhance the accuracy of terminology recognition. During model training, historical power grid dispatch instructions and operator operational preference annotations are used as the training set. The text data undergoes preprocessing including word segmentation, part-of-speech tagging, and entity annotation. The cross-entropy loss function is used as the optimization objective, and iterative training with the Adam optimizer is employed. Gradient descent is used to adjust model parameters, while dropout regularization prevents overfitting. After multiple iterations, the model accurately identifies key semantics in dispatch text, ultimately achieving efficient extraction of time window requirements, prohibited operation switches, and desired optimization objectives, which are then converted into standardized structured constraint parameters.

[0057] The extracted time window requirements, list of prohibited operation switches, and expected optimization goals are taken as inviolable hard constraints for scheduling. They are added to the input prompt information after the previous enhancement process through structured tag embedding and instruction supplementation. The constraint priority and execution rules are clearly marked. Then, they are input into the decision guidance unit of the large language model, so that the model strictly follows the operator's scheduling intention during the temporal reasoning and strategy generation process. Finally, the natural language reconstruction strategy summary is output, which fits the actual on-site operation and meets the artificial constraints.

[0058] In one possible implementation, step S400 further includes: Step S410: Using a preset strategy compilation module, perform semantic parsing on the natural language reconstruction strategy summary to identify the reconstruction target, action description, state association, and constraint information.

[0059] Step S420: According to the predefined mapping rules, the reconstruction objective is quantized into the objective function of the differentiable optimization problem, the action description and the corresponding state association are transformed into the equality and inequality constraints of the differentiable optimization problem, and the corresponding weight coefficients are set according to the semantic strength in the constraint information.

[0060] Step S430: Obtain the gradient information of the state variables output by the neural differential equation network with respect to the switching action from the power grid dynamic characterization unit, and add it as a physical dynamic constraint to the differentiable optimization problem.

[0061] Specifically, the pre-defined strategy compilation module employs semantic parsing based on dependency parsing, keyword rule matching, and domain ontology knowledge base verification. First, it segments, words, and filters stop words in the natural language strategy summary text. Then, it uses dependency parsing to break down subject-verb-object and modification logic, extracting core semantic units. Next, it combines the distribution network dispatch ontology knowledge base to match professional entities such as reconstruction targets, switch actions, electrical states, and dispatch constraints. Through pre-defined regular rules and semantic tag classification, it distinguishes and identifies reconstruction targets such as optimal network loss and voltage stability, action descriptions such as switch opening and closing sequences, state relationships between actions and node voltage and branch power changes, and constraint information such as time, equipment, and safety. This achieves accurate extraction of natural language strategies into structured semantic elements.

[0062] By using a predefined distribution network scheduling semantic-mathematical model mapping rule library, reconfiguration objectives such as voltage stability and minimum network loss are mapped into continuously differentiable objective functions containing node voltage and branch power. The timing correlation between switch opening and closing actions and actions and electrical states is converted into power flow balance equation constraints, switch action logic equation constraints, and voltage amplitude and current power upper and lower limit inequality constraints. Then, through a constraint semantic level classification mechanism, safety hard constraints, operation preference constraints, and general optimization constraints are classified, and weighted least squares are used to assign corresponding weight coefficients to different constraint terms, thus completing the quantitative construction of the mathematical model of the differentiable optimization problem.

[0063] The forward propagation output of the neural differential equation network in the power grid dynamic characterization unit is read through a preset data interaction interface. The partial derivative gradient information of the state variables such as node voltage and branch power with respect to each switching action variable is solved using the built-in automatic differentiation module of the model. This is used to characterize the influence relationship of switching action on the physical dynamics of the power grid. The gradient information is constructed as a dynamic constraint term and added to the established differentiable optimization problem in a differentiable form to improve the physical dynamic constraint boundary of the optimization model.

[0064] In one possible implementation, step S500 further includes: Step S510: In the differentiable optimization solution unit, the differentiable properties of the differentiable optimization problem are used to solve it. During the solution process, the execution time and content of the candidate switch action sequence are repeatedly adjusted.

[0065] Step S520: After each adjustment, call the differentiable forward simulation function of the power grid dynamic characterization unit to perform forward simulation on the current candidate switching action sequence, evaluate the corresponding system performance indicators and constraint satisfaction, and calculate the gradient of the objective function and constraints of the differentiable optimization problem with respect to the switching action to guide the next round of adjustment.

[0066] Step S530: When the solution process meets the preset convergence condition, the differentiable optimization solution unit outputs a switching action sequence defined on the continuous time axis.

[0067] Specifically, a gradient projection optimization algorithm with temporal constraints is used to solve the problem. The algorithm consists of a gradient descent main framework and a hierarchical optimization structure for temporal variables. The switching action content and the execution time of the action are treated as continuously differentiable optimization variables. Iterative optimization is carried out based on the gradient achievable property of differentiable optimization problems. First, the candidate switching action sequence is initialized randomly or according to rules. Then, the gradient of the objective function and constraint conditions with respect to the switching action and temporal variables is solved through the automatic differentiation module to determine the descent update direction. Combined with the projection operator, the switching action logic and time window boundaries are ensured to be compliant. During the iteration process, the switching combination, action order and the execution time of each action are adjusted simultaneously to continuously update the candidate switching action sequence.

[0068] After each adjustment of the candidate switch action sequence, the program interface calls the differentiable forward simulation module built on the neural differential equation network of the power grid dynamic characterization unit to perform power flow dynamic simulation on the current switch action sequence in the continuous time domain. The evaluation stage calculates the performance indicators such as node voltage, branch power, and system network loss in real time, and checks whether the constraints such as voltage upper and lower limits, power limits, and power flow balance are met, and outputs the constraint overrun and the quantitative values ​​of the performance indicators. The gradient calculation stage adopts an automatic differential chain differentiation mechanism, which propagates backward along the calculation path of the neural differential equation network to solve the objective function and the partial derivatives of each constraint condition with respect to the switch action state and execution time, obtains accurate gradient information, and sends it back to the optimization solution unit. This gradient direction guides the iterative update of the content and timing of the next round of switch actions.

[0069] A multi-index joint convergence judgment mechanism is adopted to collect four types of data in real time during the iteration process: the rate of change of the objective function, the gradient L2 norm, the total number of constraint violations, and the number of iterations. These data are compared and verified with the corresponding preset thresholds. When any convergence condition is triggered, such as the objective function convergence fluctuation, the gradient tending to zero, all constraints being satisfied, or the number of iterations reaching the upper limit, the gradient iteration optimization process is terminated. The differentiable optimization solution unit performs time-series normalization on the optimization variables and outputs the optimal switching action sequence arranged along the continuous time axis, which includes each switching action and its corresponding precise execution time.

[0070] In one possible implementation, step S510 further includes: Step S511: Perform a secondary analysis on the natural language reconstruction strategy summary to extract implicit guidance information regarding the execution order of switching actions, time window preferences, and dependencies between actions.

[0071] Step S512: Based on the implicit guidance information, construct a reference trajectory regarding the timing of the switching actions.

[0072] Step S513: During the optimization process, minimizing the deviation between the candidate switch action sequence and the reference trajectory is used as an auxiliary optimization objective or regularization term to guide the solution process to quickly converge to a switch action sequence that satisfies both the mathematical model constraints of the differentiable optimization problem and the intent of the implicit guidance information.

[0073] Specifically, a BERT-BiLSTM-CRF joint temporal dependency reasoning algorithm is used to complete the secondary deep semantic analysis. BERT is a bidirectional encoder representation, BiLSTM is a bidirectional long short-term memory network, and CRF is a conditional random field. The algorithm consists of four layers: a semantic encoding layer, a bidirectional feature extraction layer, a sequence labeling layer, and a temporal relation reasoning layer. First, a BERT bidirectional encoder representation model, fine-tuned with data from the power distribution network dispatching domain, is used to globally semantically encode the natural language reconstruction strategy summary, fully capturing the context and professional semantics. Next, the BiLSTM bidirectional long short-term memory network layer is used to bidirectionally mine deep features such as temporal logic and action associations in the text. Then, the CRF conditional random field layer performs precise sequence labeling on switching actions, time information, and logical constraint words. Finally, relying on the temporal relation reasoning layer, the execution order of switching actions, time window preferences, and dependencies between actions (pre- and post-actions, and linkage triggers) are inferred based on the labeling results, achieving refined extraction of implicit guidance information.

[0074] Based on the implicit guidance information extracted from the execution order of switching actions, time window preferences, and pre- and post-action dependencies, and using a continuous time axis as a foundation, the logical order of each switching action is first determined according to the action dependencies. Then, the feasible time interval for each action is defined by combining the time window preferences. Subsequently, an initial expected execution time is assigned to each switching action within the interval, while constraining the time interval between actions with dependencies. Finally, a timing reference trajectory of switching actions containing the corresponding continuous time nodes and timing arrangement rules of each switching action is generated, providing a timing benchmark for subsequent optimization solutions.

[0075] During the gradient iterative optimization process, the timing deviation between the actual execution time and sequence of each action in the candidate switching action sequence and the timing reference trajectory is quantified. Minimizing this deviation is set as an auxiliary optimization objective or a regularization penalty term and integrated into the mathematical model of the differentiable optimization problem. By constraining the degree of deviation of the optimization variables from the reference trajectory, the direction of iterative solution is guided, so that the final switching action sequence meets the mathematical constraints such as power flow balance and power limit, and conforms to the implicit scheduling intention extracted in the early stage, while effectively accelerating the model convergence speed.

[0076] This invention aims to address the technical problems of existing distribution network reconfiguration, such as coarse time-series planning, lack of semantic guidance in decision-making, and inability to conform to the continuous dynamic evolution of the power grid. To verify the actual performance of this method, a 24-hour full-time-series simulation test was conducted using the standard power field example, the IEEE 33-node radial distribution network. Simulation parameters: system base voltage 12.66kV, total active load 3715kW, total reactive load 2300kvar, including 32 branches and 5 sets of interconnection switches; distributed photovoltaic and wind power integration, with a comprehensive penetration rate of 35%; simulation duration 24 hours, time step 15 minutes, totaling 96 continuous time-series sections, simulating dynamic operating scenarios of load peak-valley fluctuations and random output of distributed energy resources.

[0077] Two sets of mainstream existing technologies are set up for comparison: Comparison with Scheme 1: Traditional discrete section static distribution network reconfiguration method, a common industry solution, with coarse time-series planning, lack of continuous dynamic modeling, and lack of semantic decision guidance.

[0078] Comparison with Scheme 2: Conventional dynamic distribution network reconfiguration method, which uses ordinary time series model, without neural differential equation continuous dynamic modeling and without large language model semantic guidance.

[0079] Experimental group: Dynamic distribution network reconfiguration method under large model enhancement of the present invention.

[0080] In response to the technical challenges addressed by this invention, quantitative tests were conducted across four dimensions: continuous dynamic adaptation capability, temporal planning and solution performance, topology operation reliability, and semantic decision adaptation capability. The experimental results are shown in Table 1. Table 1. Comparison of performance indicators of different reconstruction methods

[0081] In summary, this invention completely solves the three core problems of existing distribution network reconfiguration technologies: coarse timing planning, lack of semantic guidance in decision-making, and difficulty in conforming to the continuous dynamic evolution of the power grid. The solution has significant advantages in dynamic adaptation, timing planning, semantic decision-making, and operational reliability, meeting the engineering requirements for continuous and refined reconfiguration of dynamic distribution networks.

[0082] Example 2, based on the same inventive concept as the dynamic distribution network reconfiguration method under large model enhancement in the aforementioned examples, such as... Figure 2 As shown, this application provides a dynamic distribution network reconfiguration system under large model enhancement. The system and method embodiments in this application are based on the same inventive concept. The system includes: The operating status acquisition module 10 is used to acquire the operating status of the dynamic distribution network at the current time and in historical periods, as well as the load and distributed energy forecast data for future preset periods.

[0083] The electrical state evolution trajectory prediction module 20 is used to input the operating state quantities and the load and distributed energy prediction data into the power grid dynamic characterization unit, and predict the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain based on the neural differential equation network built into the power grid dynamic characterization unit.

[0084] The Natural Language Reconstruction Strategy Summary Generation Module 30 is used to input the current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into the large language model decision guidance unit to generate a natural language reconstruction strategy summary in the continuous time dimension. The natural language reconstruction strategy summary includes the reconstruction goal, the expected action sequence, and its correlation with the changes in the power grid state.

[0085] The physical dynamic constraint generation module 40 is used to compile the natural language reconstruction strategy summary into a differentiable optimization problem containing weight coefficients and condition constraints, and to use the gradient information provided by the power grid dynamic characterization unit as the physical dynamic constraint of the differentiable optimization problem.

[0086] The dynamic distribution network reconfiguration module 50 is used to solve the differentiable optimization problem in the differentiable optimization solution unit, output the switching action sequence defined on the continuous time axis, and complete the reconfiguration of the dynamic distribution network by executing the switching action sequence. The switching action sequence includes the execution time and action content of each switching action.

[0087] Furthermore, the system is also used to implement the following functions: The power grid dynamic characterization unit receives the operating state variables and the load and distributed energy prediction data; from the operating state variables, it extracts the current voltage, branch current, and power of the power grid nodes as the initial state variables of the neural differential equation network; based on the future power distribution indicated by the load and distributed energy prediction data, and combined with the physical laws of the power grid, it deduces the switching state change trend of the distribution network topology in the future time period, and quantifies the switching state change trend as a hypothetical control input sequence of the neural differential equation network; in the power grid dynamic characterization unit, the neural differential equation network uses the hypothetical control input sequence as its control input and the initial state variables as the starting point of its state variables, and performs forward simulation based on the embedded continuous dynamics model; through the forward simulation, it calculates the future change path of the power grid electrical state in continuous time under the action of the hypothetical control input sequence, and generates the short-term electrical state evolution trajectory of the dynamic distribution network.

[0088] Furthermore, the system is also used to implement the following functions: The real-time status of switches is obtained from the power distribution network monitoring system to determine the current topology; connectivity and circulation analysis are performed on the current topology to generate topology violation diagnosis results; the current topology, the topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory are integrated into structured input prompt information; the input prompt information is input into the large language model decision guidance unit to infer and output a natural language reconstruction strategy summary in the continuous time dimension.

[0089] Furthermore, the system is also used to implement the following functions: The input prompt information is enhanced by: a. parsing the topology violation diagnosis results in the input prompt information to identify the key switches or network areas causing the violation; b. based on the short-term electrical state evolution trajectory in the input prompt information, locating the sensitive period of drastic electrical state changes near the key switches or network areas; c. explicitly marking the key switches or network areas and the sensitive period in the input prompt information, and adding guidance instructions to indicate that when generating the strategy summary, priority should be given to planning corrective actions for the key switches or network areas during the sensitive period; the enhanced input prompt information is input into the large language model decision guidance unit, which infers based on the annotations and guidance instructions in the enhanced input prompt information to generate a natural language reconstruction strategy summary in the continuous time dimension.

[0090] Furthermore, the system is also used to implement the following functions: The system acquires scheduling instructions or operational preferences input by operators in natural language; it parses the scheduling instructions or operational preferences to extract time window requirements, prohibited operation switches, or expected optimization goals; and it incorporates the time window requirements, prohibited operation switches, or expected optimization goals as additional hard constraints into the input prompts of the large language model decision guidance unit to guide the generated natural language reconstruction strategy summary to conform to the on-site scheduling intent.

[0091] Furthermore, the system is also used to implement the following functions: The natural language reconstruction strategy summary is semantically parsed using a pre-defined strategy compilation module to identify the reconstruction target, action description, state association, and constraint information. Based on predefined mapping rules, the reconstruction target is quantified into the objective function of the differentiable optimization problem. The action description and corresponding state association are transformed into equality and inequality constraints of the differentiable optimization problem, and corresponding weight coefficients are set according to the semantic strength of the constraint information. The gradient information of the state variables output by the neural differential equation network with respect to switching actions is obtained from the power grid dynamic representation unit and added to the differentiable optimization problem as a physical dynamic constraint.

[0092] Furthermore, the system is also used to implement the following functions: In the differentiable optimization solution unit, the differentiability of the differentiable optimization problem is used for solution. During the solution process, the execution time and content of the candidate switching action sequence are repeatedly adjusted. After each adjustment, the differentiable forward simulation function of the power grid dynamic characterization unit is called to perform forward simulation on the current candidate switching action sequence, evaluate its corresponding system performance indicators and constraint satisfaction, and calculate the gradient of the objective function and constraints of the differentiable optimization problem with respect to the switching action to guide the next round of adjustment. When the solution process meets the preset convergence condition, the differentiable optimization solution unit outputs the switching action sequence defined on the continuous time axis.

[0093] Furthermore, the system is also used to implement the following functions: A secondary analysis is performed on the natural language reconstruction strategy summary to extract implicit guidance information regarding the execution order of switching actions, time window preferences, and dependencies between actions. Based on this implicit guidance information, a reference trajectory for the timing of switching actions is constructed. During the optimization process, minimizing the deviation between the candidate switching action sequence and the reference trajectory is used as an auxiliary optimization objective or regularization term to guide the solution process to quickly converge to a switching action sequence that satisfies both the mathematical model constraints of the differentiable optimization problem and the intent of the implicit guidance information.

[0094] Example 3, based on the same inventive concept as the dynamic distribution network reconfiguration method under large model enhancement in the previous examples, provides a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the dynamic distribution network reconfiguration method under large model enhancement in this application. The processor executes various functional applications and data processing of the computer device by running the software programs, instructions, and modules stored in the memory, thereby implementing the aforementioned dynamic distribution network reconfiguration method under large model enhancement.

[0095] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0096] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0097] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A dynamic distribution network reconfiguration method under large model enhancement, characterized in that, The method includes: Acquire the operating status data of the dynamic distribution network at the current moment and in historical periods, as well as the load and distributed energy forecast data for future preset periods; The operating state quantities and the load and distributed energy prediction data are input into the power grid dynamic characterization unit. Based on the neural differential equation network built into the power grid dynamic characterization unit, the short-term electrical state evolution trajectory of the dynamic distribution network is predicted in the continuous time domain. The current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory are input into the large language model decision guidance unit to generate a natural language reconstruction strategy summary in the continuous time dimension. The natural language reconstruction strategy summary includes the reconstruction goal, the expected action sequence, and its correlation with the changes in the power grid state. The natural language reconstruction strategy is summarized and compiled into a differentiable optimization problem containing weight coefficients and conditional constraints, and the gradient information provided by the power grid dynamic characterization unit is used as the physical dynamic constraint of the differentiable optimization problem. In the differentiable optimization solution unit, the differentiable optimization problem is solved, and a sequence of switching actions defined on a continuous time axis is output. By executing the sequence of switching actions, the reconfiguration of the dynamic distribution network is completed. The sequence of switching actions includes the execution time and action content of each switching action. Acquire scheduling instructions or operational preferences entered by operators in natural language; The scheduling instructions or running preferences are parsed to extract time window requirements, operation prohibition switches, or expected optimization targets; The time window requirements, prohibited operation switches, or expected optimization objectives are incorporated as additional hard constraints into the input prompts of the large language model decision guidance unit to guide the generated natural language reconstruction strategy summary to conform to the on-site scheduling intent. Specifically, the natural language reconstruction strategy is summarized and compiled into a differentiable optimization problem containing weight coefficients and conditional constraints, and the gradient information provided by the power grid dynamic representation unit is used as the physical dynamic constraint of the differentiable optimization problem, including: The natural language reconstruction strategy summary is semantically parsed through a preset strategy compilation module to identify the reconstruction target, action description, state association and constraint information. According to the predefined mapping rules, the reconstruction objective is quantized into the objective function of the differentiable optimization problem, the action description and the corresponding state association are transformed into the equality and inequality constraints of the differentiable optimization problem, and the corresponding weight coefficients are set according to the semantic strength of the constraint information. The gradient information of the state variables output by the neural differential equation network with respect to the switching action is obtained from the power grid dynamic characterization unit and added to the differentiable optimization problem as a physical dynamic constraint.

2. The dynamic distribution network reconfiguration method under large model enhancement as described in claim 1, characterized in that, Based on the neural differential equation network built into the power grid dynamic characterization unit, the short-term electrical state evolution trajectory of the dynamic distribution network is predicted in the continuous time domain, including: The power grid dynamic characterization unit receives the operating status data and the load and distributed energy prediction data. The current voltage, branch current, and power of the power grid nodes are extracted from the operating state variables and used as the initial state variables of the neural differential equation network. Based on the future power distribution indicated by the load and distributed energy forecast data, and combined with the laws of power grid physics, the switching state change trend of the distribution network topology in the future period is deduced, and the switching state change trend is quantified into the hypothetical control input sequence of the neural differential equation network. In the power grid dynamic characterization unit, the neural differential equation network takes the hypothetical control input sequence as its control input and the initial state variable as the starting point of its state variable, and performs forward simulation based on the embedded continuous dynamics model; Through the forward simulation, the future change path of the electrical state of the power grid in continuous time under the action of the hypothetical control input sequence is calculated, and the short-term electrical state evolution trajectory of the dynamic distribution network is generated.

3. The dynamic distribution network reconfiguration method under large model enhancement as described in claim 1, characterized in that, The current topology, topology violation diagnosis results, load and distributed energy prediction data, and short-term electrical state evolution trajectory are input into the large language model decision guidance unit to generate a summary of the natural language reconstruction strategy in the continuous time dimension, including: Obtain the real-time status of switches from the power distribution network monitoring system to determine the current topology; Perform connectivity and circulation analysis on the current topology to generate topology violation diagnosis results; The current topology, the topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory are integrated into structured input prompt information. The input prompt information is input into the large language model decision guidance unit, which infers and outputs a summary of natural language reconstruction strategies in the continuous time dimension.

4. The dynamic distribution network reconfiguration method under large model enhancement as described in claim 3, characterized in that, The input prompts are fed into the large language model decision guidance unit, which infers and outputs a summary of the natural language reconstruction strategy in the continuous time dimension, including: Enhancement processing of the input prompt information includes: a. Analyze the topology violation diagnosis results in the input prompt information to identify the key switches or network areas that caused the violation; b. Based on the short-term electrical state evolution trajectory in the input prompt information, locate the sensitive period of drastic changes in electrical state near the key switch or network area; c. In the input prompt information, explicitly mark the key switch or network area and the sensitive time period, and add guiding instructions to indicate that when generating the strategy summary, priority should be given to planning correction actions for the key switch or network area during the sensitive time period; The enhanced input prompts are input into the large language model decision guidance unit, which then performs inference based on the annotations and guidance instructions in the enhanced input prompts to generate a summary of natural language reconstruction strategies in the continuous time dimension.

5. The dynamic distribution network reconfiguration method under large model enhancement as described in claim 1, characterized in that, In the differentiable optimization solution unit, the differentiable optimization problem is solved, and the output is a sequence of switching actions defined on a continuous time axis, including: In the differentiable optimization solution unit, the differentiable properties of the differentiable optimization problem are used to solve it, and the execution time and content of the candidate switch action sequence are repeatedly adjusted during the solution process. After each adjustment, the differentiable forward simulation function of the power grid dynamic characterization unit is invoked to perform forward simulation on the current candidate switching action sequence, evaluate the corresponding system performance indicators and constraint satisfaction, and calculate the gradient of the objective function and constraints of the differentiable optimization problem with respect to the switching action to guide the next round of adjustment. When the solution process meets the preset convergence condition, the differentiable optimization solution unit outputs a switching action sequence defined on a continuous time axis.

6. The dynamic distribution network reconfiguration method under large model enhancement as described in claim 5, characterized in that, Solving the differentiable optimization problem using its differentiable property includes: A secondary analysis is performed on the natural language reconstruction strategy summary to extract implicit guidance information regarding the execution order of switching actions, time window preferences, and dependencies between actions; Based on the implicit guidance information, a reference trajectory regarding the timing of the switching actions is constructed; During the optimization process, minimizing the deviation between the candidate switch action sequence and the reference trajectory is used as an auxiliary optimization objective or regularization term to guide the solution process to quickly converge to a switch action sequence that satisfies both the mathematical model constraints of the differentiable optimization problem and the intent of the implicit guidance information.

7. A dynamic distribution network reconfiguration system enhanced by a large model, characterized in that, The system is used to implement the dynamic distribution network reconfiguration method under large model enhancement as described in any one of claims 1-6, and the system comprises: The operation status acquisition module is used to acquire the operation status of the dynamic distribution network at the current time and in historical periods, as well as the load and distributed energy forecast data for future preset periods. The electrical state evolution trajectory prediction module is used to input the operating state quantities and the load and distributed energy prediction data into the power grid dynamic characterization unit, and predict the short-term electrical state evolution trajectory of the dynamic distribution network in the continuous time domain based on the neural differential equation network built into the power grid dynamic characterization unit. The Natural Language Reconstruction Strategy Summary Generation Module is used to input the current topology, topology violation diagnosis results, the load and distributed energy prediction data, and the short-term electrical state evolution trajectory into the Large Language Model Decision Guidance Unit to generate a continuous time-dimensional Natural Language Reconstruction Strategy Summary. The Natural Language Reconstruction Strategy Summary includes the reconstruction objective, the expected action sequence, and its correlation with changes in the power grid state. The physical dynamic constraint generation module is used to compile the natural language reconstruction strategy summary into a differentiable optimization problem containing weight coefficients and condition constraints, and to use the gradient information provided by the power grid dynamic characterization unit as the physical dynamic constraint of the differentiable optimization problem. The dynamic distribution network reconfiguration module is used to solve the differentiable optimization problem in the differentiable optimization solution unit, output the switching action sequence defined on the continuous time axis, and complete the reconfiguration of the dynamic distribution network by executing the switching action sequence. The switching action sequence includes the execution time and action content of each switching action.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the dynamic distribution network reconfiguration method under large model enhancement as described in any one of claims 1-6.

Images