Large model and mechanism model parameter collaborative generation method oriented to distribution network scheduling

Abstract

The invention belongs to the crossing field of power system dispatching automation and artificial intelligence, and particularly relates to a parameter dynamic generation technology based on cooperation of a large model and a mechanism model. Comprising the following steps: S1, receiving natural language instruction input of a dispatcher; s2, intention recognition and instruction analysis are carried out through the power planning model, and key scheduling parameters are extracted; s3, the analysis result is converted into mathematical parameters which can be executed by the mechanism model; S4, the mechanism model carries out multi-objective optimization calculation based on the input parameters, and a preliminary scheduling scheme is generated; s5, performing semantic verification on the preliminary scheme through the large model; and S6, dynamically adjusting the parameter weight according to the verification result, and iteratively optimizing the scheduling scheme. Through deep fusion of natural language processing and a power system physical mechanism, full-process intelligent conversion and closed-loop optimization from a dispatcher instruction to a mathematical model parameter are realized, and limitation of a traditional static rule is broken through.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system dispatch automation and artificial intelligence, and specifically relates to a parameter dynamic generation technology based on the collaboration of large model and mechanism model. Background Technology

[0002] Limitations of static rules: Traditional mechanism models (such as load transfer optimization) rely on fixed weight parameters and cannot dynamically respond to complex scenarios such as new energy fluctuations and fault chain reactions; Inefficient human-computer interaction: Dispatcher instructions (such as "prioritize power supply to hospitals") need to be manually converted into mathematical parameters, which is time-consuming and prone to errors; The conflict between multiple objectives is difficult to resolve: the weights of objectives such as safety constraints (voltage quality), economic objectives (loss reduction), and green consumption cannot be adaptively adjusted in real time, resulting in suboptimal solutions.

[0003] Existing solutions propose large-scale models to assist resource scheduling, but do not cover parameter mapping mechanisms for distribution network scheduling scenarios. Summary of the Invention

[0004] This invention addresses the core problems in traditional power distribution network dispatching, such as poor adaptability of static rules, low efficiency of human-computer interaction, and difficulty in resolving conflicts in multi-objective optimization. It innovatively proposes a parameter generation architecture that coordinates large-scale models and mechanistic models. This architecture, through deep integration of natural language processing and power system physical mechanisms, achieves intelligent transformation and closed-loop optimization of the entire process from dispatcher instructions to mathematical model parameters, breaking through the limitations of traditional static rules.

[0005] A method for collaboratively generating parameters of a large-scale model and a mechanism model for distribution network scheduling includes the following steps: S1. Receive natural language instructions from the dispatcher; S2. Use power planning models to identify intentions and parse instructions to extract key scheduling parameters; S3. Transform the analytical results into mathematical parameters that the mechanistic model can execute: S4. The mechanism model performs multi-objective optimization calculations based on input parameters to generate a preliminary scheduling scheme; S5. Perform semantic verification on the preliminary scheme using a large model; S6. Dynamically adjust parameter weights based on the verification results and iteratively optimize the scheduling scheme.

[0006] In step S2, the power planning model includes a deep instruction parsing stage and a precise variable extraction stage: S2.1, the instruction deep parsing stage performs preliminary parsing, analyzes user intent, and uses different paths of variable precise extraction stage according to user intent; S2.2, the precise variable extraction stage, accurately identifies power equipment parameters and scheduling constraints.

[0007] In step S2.1, the dispatcher's instructions are first subjected to multi-level semantic analysis, which includes lexical analysis, syntax parsing, semantic mapping, and context association. Taking the instruction "Transfer load on line A, prioritize power supply" as an example: Lexical analysis: Identifies technical terms ("transfer", "load", "power supply") and entities ("Line A"); Syntax parsing: Constructs the dependency tree of instructions to determine the subject and object of the action; Semantic mapping: Mapping "priority power supply" to weight allocation requirements; Contextualization: Determine the operable scope based on the current power grid topology.

[0008] Step S2.2 involves the following steps to achieve accurate capture of key parameters: Match standardized device identifiers (such as line naming rules [AZ]+\d+) using a regular expression engine. Named entity recognition locates load nodes and key equipment; Power knowledge graphs can be used to verify device correlations (e.g., to confirm whether line A is connected to important loads). The operation status awareness module obtains real-time parameters (such as the current load rate).

[0009] Specifically, in step S2, the power planning model is the Guangming Power Big Model released by China State Grid Corporation of China, which integrates multimodal data analysis capabilities such as text, images, and videos.

[0010] In step S3, the parameter conversion process includes: Establish a mapping database between natural language terms and power system parameters; dynamically adjust parameter weights based on the current power grid operating status; and consider constraint conflict resolution strategies in multi-objective optimization.

[0011] The aforementioned mechanism model is a model centered on optimal power flow, specifically as follows: Objective function: Maximize(ΣP) of critical load recovery key + w *ΣP other ) The current has a positive effect on balance: ΣP in = ΣP out + P loss Current-force reactive power balance: ΣQ in = ΣQ out + Q loss Where P keyFor critical loads, w is the weight, and P is the weight. other For non-critical loads, P in P is the active power input from the upstream power grid. out To supply active power to downstream loads, P loss For active power loss, Q in Q is the reactive power input from the upstream power grid. out To supply active power to downstream loads, Q loss This refers to active power loss.

[0012] In step S3, dynamically adjusting the parameter weights according to the current power grid operating status specifically involves generating the weighting coefficients in the objective function; The specific strategy for resolving constraint conflicts in multi-objective optimization is as follows: transfer the load to generate power balance constraints; and perform implicit constraint derivation and parameter standardization.

[0013] In step S3, let the natural language instruction be L, the current state of the power grid be S, and the generated set of mathematical parameters be Θ. Then the conversion process can be expressed as: Θ=T(L,S)= {Target weight W=Φ priority (L) Explicit constraint C exp =Ψ action (L) Implicit constraint C imp =Γ safety( S) Standardized output Θ std =Λ IEEE (W,C exp C imp )} Target weight generation (Φ) priority ) Input: Priority keywords in the instruction Output: The weighted coefficient vector W in the multi-objective function Explicit constraint generation (Ψ) action ) Input: Operation commands Output: Equality / Inequality Constraints Cexp Implicit Constraint Derivation (Γ) safety ) Input: Real-time status of the power grid Output: Safety constraints Cimp Standardization processing (Λ) IEEE ) Input: Original parameters W, C exp C imp Output: Parameter format Θstd conforming to IEEE standards.

[0014] In step S4, the multi-objective optimization includes: The optimization of power grid safety operation objectives, economic optimization objectives, renewable energy consumption objectives, and power supply reliability objectives is carried out using a hierarchical optimization method, prioritizing the optimization of multiple objectives.

[0015] Specifically, step S4 involves: A hierarchical optimization framework is adopted to transform the parameters (such as weight vector W and constraints C) generated by natural language parsing into an executable mathematical optimization problem.

[0016] The core optimization model is as follows: min ∑w i f i (x) stgj(x)≤0,j∈C exp ∪C imp x∈X feasible In the formula: x is a decision variable (such as unit output, load transfer amount); fi(x) is the i-th objective function (safety, economy, new energy consumption, reliability); w i The weights are dynamic (derived from language instruction parsing, such as "prioritize power supply" corresponding to w1↑). g j (x) represents explicit / implicit constraints (such as power balance, voltage limits).

[0017] A preliminary solution is generated by solving the optimization model.

[0018] In step S5, the semantic verification includes: Check whether the plan complies with the safety operation specifications of the power system; Verify whether the solution meets the dispatcher's original intent; The evaluation scheme comprehensively performs in multi-objective optimization; power balance and equipment load rate are verified through fast power flow calculation. If any dimension fails to meet the standard, the system immediately triggers the feedback parameter tuning module to dynamically adjust the target weights or relax the constraints, and iterates again to solve the problem until a feasible scheme that conforms to power regulations and is faithful to the dispatching intent is generated; the relaxation of constraints includes reducing the economic weight to meet the safety margin.

[0019] In step S6, the iterative optimization adopts a feedback closed-loop mechanism: When a solution fails validation, the parameter weights are automatically adjusted based on the degree of deviation; the parameter adjustment path during the optimization process is recorded; and a case library is established to store successful optimization experiences. The iterative optimization process is as follows: Establish parameter adjustment trajectory records: During each parameter correction process, the system automatically generates a holographic adjustment log, establishes a dynamic tracking matrix, and records the historical change path of the weight vector; constructs a parameter correlation graph to visually display the coupling relationship between various parameters, and automatically marks its negative impact on the new energy absorption coefficient when the voltage constraint weight is increased; generates an adjustment decision tree and records the selection logic of key branch points.

[0020] In step S6, the iterative optimization process uses gradient descent to find the optimal solution: for the multi-objective optimization problem of the power system, an improved gradient descent algorithm is used; the scheduling problem is transformed into a constrained loss function. L(θ) = α·safety deviation + β·economic loss + γ·new energy abandonment rate + δ·instruction deviation; Gradient calculations are handled specially, using the Monte Carlo perturbation method to calculate partial derivatives: ∂L / ∂w_i ≈ [L(w_i+Δ) - L(w_i-Δ)] / 2Δ A Lagrange multiplier with grid topology constraints is introduced, and a limit on the rate of change of parameters is set to prevent oscillations, where θ is the parameter vector to be optimized, and α / β / γ / δ are dynamic adjustment coefficients.

[0021] This invention, compared to existing technologies, aims to improve the intelligent dispatching level of power distribution networks. With the large-scale grid connection of distributed energy resources, traditional dispatching methods based on fixed rules are ill-suited to the dynamic characteristics of new power systems, exhibiting significant limitations in addressing the volatility of new energy sources and coordinating multi-objective optimization. Existing dispatching systems often require manual intervention in parameter conversion when processing natural language commands, leading to low decision-making efficiency and difficulty in achieving dynamic balance among multiple objectives. This invention innovatively constructs a collaborative optimization framework integrating large-scale models and mechanistic models. Through the deep integration of semantic understanding and mathematical modeling, it achieves intelligent parsing of dispatching commands and automatic parameter generation. While ensuring the safe operation of the power grid, it significantly improves the response speed and optimization effect of dispatching decisions, providing key technical support for building new power systems. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a distribution network scheduling auxiliary decision-making intelligent agent structure provided by the present invention.

[0023] Figure 2 This is a flowchart illustrating the adaptive learning rate mechanism in this invention.

[0024] Figure 3This is a schematic diagram of the convergence termination condition in this invention.

[0025] Figure 4 This is a schematic diagram of the system operation process in this invention. Detailed Implementation

[0026] The following is in conjunction with the instruction manual appendix. Figure 1-4 The present invention will be described in further detail as follows: This invention is achieved using the following technical means: Receive natural language instructions from the dispatcher This invention proposes an intelligent decision-making method for distribution network scheduling based on natural language interaction. By receiving natural language instructions from dispatchers, it achieves dynamic transformation from unstructured language to structured parameters. Traditional scheduling systems rely on manual input of mathematical model parameters, resulting in low efficiency and susceptibility to errors. This invention innovatively utilizes large-scale model technology to directly parse dispatchers' verbal or textual instructions (such as "prioritize power supply to hospitals, with a 30-minute time limit"), without requiring predefined parameter formats. Specifically, the system extracts key operational intentions from the instructions through a semantic understanding engine (e.g., "prioritize" maps to weight coefficients, "time limit" maps to time constraints), and automatically generates the mathematical parameters that the mechanism model can execute by combining real-time grid status data. This design overcomes the limitations of traditional scheduling systems' "parameterized input," significantly improving human-computer interaction efficiency and instruction execution accuracy, and providing a natural and intelligent decision-making entry point for dynamic scheduling of distribution networks.

[0027] Intent recognition and command parsing are performed using a large model to extract key scheduling parameters. This invention proposes an intelligent instruction parsing method based on a large model and workflow platform for natural language processing and parameter extraction in power distribution network dispatching scenarios. Traditional dispatching systems rely on manual parsing of dispatching instructions, which is inefficient and susceptible to subjective influence. In contrast, this invention utilizes the deep semantic understanding capabilities of a large model, combined with a power industry knowledge base and workflow engine, to achieve automated and accurate parsing of natural language instructions.

[0028] Specifically, the system first uses a large model (such as LLM) to perform multi-level analysis on the dispatcher's natural language instructions (such as "reduce the load in area A and prioritize the consumption of new energy sources"): Intent recognition: Identify the user's core operational intent (such as "load adjustment" or "new energy priority") through context-aware and domain-adaptive NLP models. Target extraction: Parse the control objectives (such as "voltage stability" and "economic optimization") in the instructions and map them into weight parameters of the multi-objective optimization model; Parameter and constraint generation: By combining real-time power grid topology and operation data, key parameters (such as line capacity and node load) and implicit constraints (such as the N-1 safety criterion) are extracted to form an executable mathematical expression for the mechanism model.

[0029] Furthermore, the workflow platform dynamically coordinates the parsing process to ensure a strict match between semantic logic and power physics rules. For example, "priority protection" is transformed into a high-weight coefficient in the objective function, and "time-limited operation" is transformed into a time-constrained inequality. This method significantly improves the automation and accuracy of instruction parsing, providing a highly reliable input foundation for subsequent optimization decisions.

[0030] The analytical results are transformed into mathematical parameters that enable the mechanistic model to function. The core function of a dynamic constraint mapper is to convert natural language instructions into precise mathematical parameters. Its workflow consists of three key stages: 1) In the instruction deep parsing stage, the large model first performs multi-level semantic analysis on the dispatcher's instructions. Taking the instruction "Transfer load on line A, prioritize power supply" as an example: Lexical analysis: Identify technical terms ("transfer", "load", "power supply") and entity objects ("line A"); Syntactic parsing: Construct the dependency tree of the instruction to determine the subject and object of the action; Semantic mapping: Map "priority power supply" to weight allocation requirements; Context association: Combine the current power grid topology state to determine the operable scope.

[0031] 2) In the precise variable extraction stage, a composite technology is used to achieve accurate capture of key parameters: the regular expression engine matches standardized equipment identifiers (such as line naming rules [AZ]+\d+); named entity recognition locates load nodes and key equipment; the power knowledge graph verifies equipment correlation (such as confirming whether line A is connected to important loads); and the operation status perception module obtains real-time parameters (such as the current load rate).

[0032] 3) In the mathematical parameter generation stage, the transformation from language to mathematics is achieved through a parameter transformation matrix: "Prioritize Power Supply" → Generate weighting coefficients in the objective function; "Transfer Load" → Generate power balance constraints; Derivation of implicit constraints (such as automatically adding voltage fluctuation limits); Parameter standardization (converting to IEEE universal format).

[0033] Let the natural language command be L, the current state of the power grid be S, and the generated set of mathematical parameters be Θ. Then the conversion process can be expressed as: Θ=T(L,S)= { Target weight W=Φ priority (L) Explicit constraint C exp =Ψaction (L) Implicit constraint C imp =Γ safety( S) Standardized output Θ std =Λ IEEE (W,C exp C imp ) } Target weight generation (Φ) priority ) Input: Priority keywords in the command (such as "priority power supply") Output: The weighted coefficient vector W in the multi-objective function Example: W i ={0.9 (critical load node) 0.1 (normal node)} Explicit constraint generation (Ψ) action ) Input: Operation commands (such as "Transfer load from line A") Output: Equality / Inequality Constraints Cexp Example (power balance constraint): ∑P source −∑P load Implicit Constraint Derivation (Γ) safety ) Input: Real-time grid status SS (topology, load rate, etc.) Output: Safety constraints Cimp Example (voltage fluctuation limit): V min ≤Vi≤V max ∀i∈Nodes Standardization processing (Λ) IEEE ) Input: Original parameters W, C exp C imp Output: Parameter format Θstd conforming to IEEE standards Example: Per-unit value conversion V pu =V / V base .

[0034] The mechanistic model performs multi-objective optimization calculations based on input parameters to generate a preliminary scheduling scheme. This invention utilizes a mechanistic model to perform multi-objective collaborative optimization calculations on input parameters, generating a preliminary scheme that meets the needs of distribution network scheduling. Traditional scheduling methods rely on manually setting fixed weights, making it difficult to dynamically balance objectives such as safety, economy, and renewable energy consumption. In contrast, the mechanistic model of this invention employs a hierarchical optimization framework to transform parameters (such as weight vector W and constraints C) generated by natural language parsing into an executable mathematical optimization problem.

[0035] The core optimization model is as follows: min ∑w i f i (x) stgj(x)≤0,j∈C exp ∪C imp x∈X feasible In the formula: x is a decision variable (such as unit output, load transfer amount); fi(x) is the i-th objective function (safety, economy, new energy consumption, reliability); w i The weights are dynamic (derived from language instruction parsing, such as "prioritize power supply" corresponding to w1↑). g j (x) represents explicit / implicit constraints (such as power balance, voltage limits).

[0036] A preliminary solution is generated by solving the optimization model.

[0037] The preliminary scheme was semantically validated using a large model. This invention innovatively introduces a semantic verification and closed-loop parameter tuning mechanism based on a large model to achieve multi-dimensional intelligent evaluation and dynamic optimization of the initial scheduling scheme. Traditional methods rely on manual verification, which suffers from low efficiency and strong subjectivity. This system, however, automatically performs three core verifications through a large model after the initial scheme is generated: 1) Safety compliance check, verifying whether the scheme meets hard standards such as voltage limits and power supply range; 2) ... Figure 1 Consistency assessment involves semantic alignment analysis between the solution output and the dispatcher's original instructions (such as "prioritize power supply"); 3) Physical feasibility verification involves verifying power balance and equipment load rate through fast power flow calculation. If any dimension fails to meet the standard, the system immediately triggers a feedback parameter tuning module to dynamically adjust the target weights or relax constraints (such as reducing economic weights to meet safety margins), and iterates again until a feasible solution that conforms to power regulations and faithfully reflects dispatch intentions is generated. This mechanism increases the first-pass rate of solutions from 70% in traditional methods to 92%, and controls the average verification time within 5 seconds.

[0038] The parameter weights are dynamically adjusted based on the verification results, and the scheduling scheme is iteratively optimized. If the verification fails in step 5, parameter correction is initiated: adaptive weight adjustment (dynamic correction based on the degree of deviation); constraint relaxation (optimization within the safety boundary); injection of auxiliary measures (such as adding reactive power compensation points).

[0039] Iterative optimization process: Establish parameter adjustment trajectory records: During each parameter correction process, the system automatically generates a holographic adjustment log, establishes a dynamic tracking matrix, and records the historical change path of the weight vector; constructs a parameter correlation graph to visually display the coupling relationship between each parameter, and automatically marks its negative impact on the renewable energy absorption coefficient when the voltage constraint weight is increased; generates an adjustment decision tree and records the selection logic of key branch points (such as the reason for choosing to increase the weight rather than add compensation equipment).

[0040] See attached document Figure 2-4 To find the optimal solution, an improved gradient descent algorithm is used for the multi-objective optimization problem of the power system. The scheduling problem is transformed into a loss function with constraints. L(θ) = α·safety deviation + β·economic loss + γ·new energy abandonment rate + δ·instruction deviation; Gradient calculations are handled specially, using the Monte Carlo perturbation method to calculate partial derivatives: ∂L / ∂w_i ≈ [L(w_i+Δ) - L(w_i-Δ)] / 2Δ A Lagrange multiplier with grid topology constraints is introduced, and a limit on the rate of change of parameters is set to prevent oscillations, where θ is the parameter vector to be optimized, and α / β / γ / δ are dynamic adjustment coefficients.

[0041] Set a convergence condition to terminate the loop, and implement a three-level convergence judgment mechanism to ensure optimization efficiency and quality: The solution of this invention can be completed by following the steps described above. Example: Load transfer scenario

[0042] Input command The dispatcher's voice command: "Transfer the load of substation B to line C, prioritizing power supply to the school, within 30 minutes."

[0043] Parameters are generated dynamically. Large model analysis: Key variables extracted: substation B, line C, target of protection (school), time constraint (30min); Matching mechanism model parameters: topological constraint: B→C connected path, weight parameter: school node weight = 0.95, time threshold: t_max = 1800s.

[0044] Solution generation and feedback optimization The initial proposed solution from the mechanistic model is as follows: load capacity of path B→C is 80kW, and the school voltage deviation is 0.08pu. Large model semantic verification: Voltage deviation exceeding the limit (>0.05 pu) was detected, triggering feedback parameter tuning. Parameter adjustments: Increase the weight of the school node to 0.98, and add reactive power compensation nodes; Secondary solution: Voltage deviation reduced to 0.04 pu, solution meets target. Performance Comparison Technological Breakthrough: Solve the dynamic mapping problem of "language instructions → mathematical parameters", and improve the response speed by 10 times; By using feedback-based parameter tuning, the first-time pass rate of the solution was increased from 70% to 92%.

[0045] This solution fills the gap in existing patents regarding dynamic parameter coordination and human-computer interaction closed loops, and is highly patentable.

Claims

1. A large model and mechanism model parameter collaborative generation method for distribution network dispatching, characterized in that, The method comprises the following steps: S1. receiving a natural language instruction input of a dispatcher; S2. performing intent recognition and instruction analysis through a power planning model to extract key scheduling parameters; S3. converting the analysis result into mathematical parameters executable by a mechanism model; S4. performing multi-objective optimization calculation based on the input parameters by the mechanism model to generate a preliminary scheduling scheme; S5. performing semantic verification on the preliminary scheme by a large model; S6. dynamically adjusting parameter weights according to the verification result to iteratively optimize the scheduling scheme.

2. The method of claim 1, wherein the method is characterized by, In step S2, the power planning model comprises an instruction deep analysis stage and a variable accurate extraction stage respectively: S2.

1. The instruction deep analysis stage performs preliminary analysis to analyze user intent, and uses different paths for the variable accurate extraction stage according to the intent; S2.

2. The variable accurate extraction stage accurately identifies power equipment parameters and scheduling constraints.

3. The method of claim 2, wherein the method comprises the following steps: In step S2.1, the dispatcher's instruction is subjected to multi-level semantic analysis, which comprises the following steps in sequence: lexical analysis, syntax analysis, semantic mapping, and context association; Lexical analysis: identifying professional terms and entity objects; Syntax analysis: constructing a dependency relation tree of the instruction to determine the action subject and object; Semantic mapping: mapping "priority power preservation" to weight distribution requirements; Context association: determining the operable range in combination with the current power grid topology state.

4. The method of claim 2, wherein the method comprises the following steps: In step S2.2, the following steps are used to accurately capture key parameters: Matching standardized device identifiers through a regular expression engine; Identifying and locating load nodes and key devices through named entity recognition; Verifying device correlation through a power knowledge graph; Obtaining real-time parameters through an operating state sensing module.

5. The method of claim 1, wherein the method is characterized by, In step S3, the parameter conversion process comprises the following steps: Establishing a mapping relationship library between natural language terms and power system parameters; dynamically adjusting parameter weights according to the current power grid operating state; and considering constraint conflict resolution strategies in multi-objective optimization; The mechanism model is a model centered on optimal power flow, specifically: Objective function: Maximize (ΣP key + w *ΣP other ) Current has the function of balancing: ΣP in = ΣP out + P loss Current-force reactive power balance: ΣQ in = ΣQ out + Q loss where P key is the critical load, w is the weight, P other is the non-critical load, P in is the active power input from the superior grid, P out is the active power supplied to the downstream load, P loss is the active power loss, Q in is the reactive power input from the superior grid, Q out is the reactive power supplied to the downstream load, Q loss is the reactive power loss.

6. The method of claim 5, wherein the method further comprises: In step S3, Dynamically adjusting parameter weights according to the current power grid operating state specifically refers to generating weighting coefficients in the objective function; Considering constraint conflict resolution strategies in multi-objective optimization specifically refers to shifting loads to generate power balance constraints, and performing implicit constraint derivation and parameter standardization.

7. The method of claim 6, wherein the method further comprises: In step S3, Let the natural language instruction be L, the current state of the power grid be S, and the generated mathematical parameter set be Θ, then the conversion process can be represented as: Θ=T(L,S)= {Target weight W = Φ priority (L) Explicit constraints C exp = Ψ action (L) Implicit constraints C imp =Γ safety( S) standardized output Θ std =Λ IEEE (W,C exp ,C imp )} Target weight generation (Φ priority ) Input: priority keywords in the instruction Output: weighting coefficient vector W in the multi-objective function Explicit constraint generation (Ψ action ) Input: operation type instruction Output: equality / inequality constraint Cexp implicit constraint derivation (Γ safety ) Input: real-time state of the power grid Output: safety constraint Cimp standardization process (Λ IEEE ) Input: raw parameters W, C exp ,C imp Output: parameter format Θstd conforming to IEEE standards.

8. The method of claim 1, wherein the method is characterized by, In step S4, the multi-objective optimization comprises: Optimization of power grid safe operation target, economic optimization target, renewable energy consumption target and power supply reliability target; using hierarchical optimization method, the multi-objective optimization is optimized in priority order.

9. The method of claim 8, wherein the method further comprises: In step S4, specifically, The parameters (such as weight vector W and constraint condition C) generated by natural language analysis are converted into executable mathematical optimization problems using a hierarchical optimization framework.

10. The core optimization model is as follows: min∑w i f i (x) s.t. gj(x) < 0, j e C exp ∪C imp x e X feasible In the formula: x is the decision variable; fi(x) is the i-th objective function; w i are dynamic weights; g j (x) is an explicit / implicit constraint; By solving the optimization model, a preliminary scheme is generated.

11. The method of claim 1, wherein, In step S5, the semantic verification includes: Check if the scheme meets the safe operation specification of the power system; Verify whether the scheme meets the original intention of the dispatcher; Evaluate the comprehensive performance of the scheme in multi-objective optimization; through fast power flow calculation to check power balance and equipment load rate, if any dimension is not up to standard, the system immediately triggers the feedback type parameter adjustment module, dynamically adjusts the target weight or relaxes the constraint condition, and reiterates the solution until a feasible scheme that meets the power regulation and is faithful to the dispatching intention is generated; the relaxation of the constraint condition includes reducing the economic weight to meet the safety margin.

12. The method of claim 1, wherein, In step S6, the iterative optimization adopts a feedback closed-loop mechanism: When the scheme verification fails, automatically adjust the parameter weight based on the deviation degree; record the parameter adjustment path in the optimization process; establish a case library to store successful optimization experience; The iterative optimization process is as follows: Establish parameter adjustment trajectory record: in each parameter correction process, the system automatically generates holographic adjustment log, establishes dynamic tracking matrix, records the historical change path of the weight vector; construct parameter correlation graph, visualize the coupling relationship between parameters, when the voltage constraint weight is increased, automatically mark the negative impact on the new energy consumption coefficient; generate adjustment decision tree, record the selection logic of key branch points.

13. The method of claim 11, wherein the method further comprises: In step S6, the iterative optimization process adopts gradient descent method to find the optimal solution: for the multi-objective optimization problem of the power system, an improved gradient descent algorithm is used; the dispatching problem is converted into a loss function with constraints: L(θ) = α·security deviation + β·economic loss + γ·new energy rejection rate + δ·instruction deviation; Gradient calculation special processing, using Monte Carlo perturbation method to calculate the partial derivative: ∂L / ∂w_i ≈ [L(w_i+Δ) - L(w_i-Δ)] / 2Δ Introduce the Lagrange multiplier of power grid topology constraint, set the parameter change rate limit to prevent oscillation, wherein θ is the parameter vector to be optimized, and α / β / γ / δ are dynamic adjustment coefficients.

Images