LLM-Agent-based device main circuit parameter sensitivity analysis method

Abstract

The application relates to an LLM-Agent-based device main circuit parameter sensitivity analysis method, and belongs to the field of power electronic device parameter design and optimization, and comprises the following steps: obtaining an operation condition, generating an initial parameter set through a parameter generation LLM; inputting a steady-state analysis tool for solving to obtain a steady-state data set; performing decoupling analysis on the steady-state data set through a sensitivity analysis LLM to generate a preliminary sensitivity analysis matrix; performing confidence diagnosis on the sensitivity analysis matrix and the steady-state data set through a data self-checking LLM; if the diagnosis is passed, output is generated; if the diagnosis is not passed, a supplementary measurement suggestion is generated and fed back to the parameter generation LLM to form a closed-loop iteration, so that automatic and intelligent evaluation of parameter sensitivity is realized. The application relies on three special large language models and an external steady-state calculation tool, realizes automatic and intelligent evaluation of parameter sensitivity, significantly improves the design efficiency and precision of a complex multi-parameter system, and provides a reliable basis for device optimization decision.

Description

Technical Field

[0001] This invention relates to a sensitivity analysis method for main circuit parameters of equipment based on LLM-Agent, belonging to the field of power electronic equipment parameter design and optimization technology. Background Technology

[0002] With the high penetration rate of new energy sources in new power systems, the operational requirements of these systems are becoming increasingly complex. A key challenge in main circuit parameter design is how to effectively improve the steady-state performance during normal operation and the transient performance after faults while reducing system investment costs and increasing economic efficiency, achieving multi-parameter collaborative optimization under physical constraints. This is especially true with the rapid development of power electronics technology, which has significantly increased the sensitivity of systems to main circuit parameters. Due to the strong coupling and nonlinear relationships among electrical equipment parameters, adjusting a single parameter can easily disrupt the original balance, causing complex chain reactions on multiple key performance indicators of the converter. This interconnected nature means that even small deviations in parameter design can easily lead to system reliability issues. Therefore, it is crucial to accurately quantify the degree and direction of the impact of different parameters on various key performance indicators. This not only helps designers focus on core parameters to improve optimization efficiency but also clearly reveals the trade-offs between multiple indicators to support multi-objective decision-making. Furthermore, by quantifying the impact of parameters on safety margins, the risk of exceeding limits during parameter adjustment can be effectively avoided, ensuring the safe and stable operation of the system.

[0003] However, traditional parameter design relies on engineers' empirical formulas and a limited number of simulations for verification. In scenarios with huge parameter spaces and strong coupling between parameters, traditional sensitivity analysis methods are inefficient and prone to misjudgment. The design process depends on human experience, making it difficult to form a standardized and reusable optimization loop. At the same time, it lacks the ability to actively identify and supplement tests for the "blind spots" and "nonlinear mutation regions" of data samples. These problems have become bottlenecks that limit the ability to maintain good equipment operation and improve economic efficiency in practical applications.

[0004] In contrast, Large Language Models (LLMs) possess powerful semantic understanding and logical decision-making capabilities, providing intelligent support for automated optimization and experimental planning of massive parameters. However, directly applying LLMs has two major limitations: first, they cannot directly handle the precise solution of nonlinear physical constraints, easily leading to judgments detached from actual working conditions; second, they lack proactive awareness of the "nonlinear mutation zone" in the parameter space, easily causing the optimization process to get stuck in local optima. Therefore, a novel sensitivity analysis framework is urgently needed in this field. This framework aims to deeply integrate the autonomous logical reasoning capabilities of LLMs with external professional computing tools, constructing a collaborative mechanism based on a Large Language Model Agent, namely LLM-Agent. Through this mechanism, not only can the shortcomings of a single LLM in precise numerical calculations be effectively compensated, but it can also autonomously achieve data iterative planning and feature extraction; at the same time, relying on the proactive exploration closed loop formed by multiple rounds of self-checking feedback, the system can efficiently and accurately complete the quantitative evaluation of sensitivity in a vast parameter space. This new architecture is of great significance for breaking through the bottlenecks of traditional manual design, improving parameter optimization efficiency, and thus ensuring the overall operating performance and economic benefits of equipment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a device main circuit parameter sensitivity analysis method based on LLM-Agent. It is implemented by relying on the LLM-Agent framework, which is constructed in collaboration with three dedicated large language models and external steady-state calculation tools. This method achieves automated and intelligent evaluation of parameter sensitivity, significantly improves the design efficiency and accuracy of complex multi-parameter systems, and provides a reliable basis for device optimization decisions.

[0006] The present invention adopts the following technical solution:

[0007] A method for sensitivity analysis of device main circuit parameters based on LLM-Agent includes the following steps:

[0008] S1: Obtain the power grid operating conditions input by the user, and generate an initial parameter set through parameter generation LLM;

[0009] S2, input the initial parameter set into the steady-state analysis tool for solution, and obtain a steady-state dataset containing the safety margins of each set of parameters;

[0010] S3. Based on sensitivity analysis LLM, decoupling analysis is performed on the steady-state dataset to generate a preliminary sensitivity analysis matrix;

[0011] S4. The sensitivity analysis matrix and steady-state dataset are tested for confidence using the data self-checking LLM. If the test is successful, the final sensitivity analysis matrix is ​​output. If the test fails, a supplementary test suggestion is generated and fed back to the parameter generation LLM. The parameter generation LLM updates the parameter set according to the supplementary test suggestion and returns to step S2, thus forming a closed-loop iteration to achieve automated and intelligent evaluation of parameter sensitivity.

[0012] In this invention, the parameter generation LLM is configured to switch between two modes: initial generation and closed-loop optimization, which is used to generate a structured parameter set based on the initial operating conditions of the system or self-test and supplementary test suggestions.

[0013] The sensitivity analysis LLM configuration is used to perform deep decoupling analysis on steady-state datasets. By matching control variable sample pairs, it calculates the quantitative impact of changes in each parameter on the safety margin of key performance constraints of the equipment.

[0014] The data self-checking LLM configuration performs comprehensive verification on steady-state datasets, automatically diagnoses sample defects, and generates structured retesting suggestions.

[0015] Preferably, the parameter generation LLM corresponds to the initial generation mode and the closed-loop optimization mode when generating the initial parameter set and updating the parameter set, respectively, as follows:

[0016] In the initial generation mode, multiple sets of baseline parameters with different design preferences are estimated based on physical experience, and a single-factor comprehensive scan is performed on each set of baseline parameters, i.e., single-variable step adjustment, to generate the initial parameter set.

[0017] In closed-loop optimization mode, the preferred parameter combination specified in the supplementary test suggestion is extracted as the new benchmark center to generate an updated parameter set.

[0018] Preferably, in step S2, relying on the steady-state analysis tool, the m sets of initial parameters generated in step S1 are calculated in batches to obtain the steady-state constraint electrical quantities corresponding to each set of parameters; then, the obtained steady-state constraint electrical quantities are compared with the safety physical threshold of the current operating condition of the system, and the difference between the safety physical threshold and the actual calculated steady-state constraint electrical quantities is calculated to obtain the safety margin Mar of each key performance indicator, thus forming a steady-state dataset.

[0019] Preferably, in step S3, the sensitivity analysis LLM filters the steady-state dataset to find two sets of data. and It can satisfy the condition that only one parameter is different, and the rest are different. -1 parameters are exactly the same; Indicates the total number of parameters;

[0020] To eliminate the influence of different parameter dimensions and variation ranges, the sensitivity analysis LLM normalizes the safety margin based on a preset minimum step unit, as shown in the following formula:

[0021]

[0022] in, This represents the normalized sensitivity index. , These represent the safety margins of the key performance indicators calculated under data A and data B, respectively. , These represent the specific values ​​of the target parameter that undergoes univariate step adjustment in data A and data B, respectively. This indicates the minimum step unit preset for the target parameters.

[0023] Preferably, based on the normalized sensitivity index, the sensitivity intensity and influence direction of each parameter are comprehensively judged to obtain the sensitivity analysis matrix. The process is as follows:

[0024] Regarding sensitivity intensity, multiple threshold values ​​for change are preset, including a lower threshold value. T 1. Intermediate threshold T n / 2 and upper limit threshold T n When the sensitivity index is less than 0.01, it is judged as having no sensitivity; when the sensitivity index is greater than or equal to 0.01 and less than... T A value of 1 indicates weak sensitivity; when the sensitivity index is greater than or equal to... T 1 and less than T n / 2 When the sensitivity is determined to be moderate; when the sensitivity index is greater than or equal to T n / 2 and less than T n When the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity; when the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity. T n The time limit is determined to be extremely sensitive; lower threshold T 1. Determined based on inherent system noise, used to mask ineffective micro-fluctuations; upper limit threshold. T n Determined based on the extreme value of the maximum allowable safety margin variation of the system; intermediate threshold. T n / 2 Based on T 1 to T n The median division.

[0025] Regarding the direction of influence, if the increase in parameters leads to an increase in the safety margin of the performance index, i.e., the sensitivity index is greater than zero, it is judged as a positive correlation, i.e., the system safety is improved; if the increase in parameters leads to a decrease in the safety margin of the performance index, i.e., the sensitivity index is less than zero, it is judged as a negative correlation, i.e., the system safety is reduced. Finally, a preliminary quantitative sensitivity analysis matrix is ​​output.

[0026] Preferably, in step S4, the data self-checking LLM first performs a defect self-check on the steady-state dataset from three aspects, including missing control variables, conflict, and sparse sampling. When any one of the defects is triggered, it is determined that the self-check has failed and is marked as a low confidence region.

[0027] For low-confidence areas, the data self-checking LLM will generate a structured supplementary test suggestion containing specific reasons and resampling instructions, and feed it back to the parameter generation LLM. After receiving the supplementary test suggestion, the parameter generation LLM triggers a closed-loop optimization mode, extracts the preferred parameter combination specified in the supplementary test suggestion as a new baseline center, reduces the step size or densifies the sampling to regenerate the supplementary parameter set.

[0028] If none of the three defects are triggered, the self-test is considered passed, meaning the sensitivity analysis matrix is ​​converged and stable with high confidence. The iteration then terminates, and the final sensitivity analysis matrix is ​​output, providing a reliable decision-making basis for equipment parameter optimization and trade-off design.

[0029] Preferably, when there are parameters in the steady-state dataset that cannot be independently decoupled for analysis, the control variables are missing.

[0030] A conflict exists when the same parameter exhibits significant differences in sensitivity indices or nonlinear abrupt changes in the direction of influence across different parameter sets.

[0031] Calculate the actual change step size between adjacent sample points of the same parameter in the steady-state dataset. If the actual change step size is greater than the preset maximum allowable sampling step size, it is determined that there is sampling sparsity.

[0032] For any details not covered in this invention, please refer to the prior art.

[0033] The beneficial effects of this invention are as follows:

[0034] This invention proposes a sensitivity analysis method for device main circuit parameters based on an LLM-Agent. This method assigns specific expert roles to the LLM through cue word engineering to clarify task boundaries, injects knowledge from the power electronics field, and enforces standardized output formats for structured data, transforming it from a general-purpose large language model into a specialized LLM for specific analysis. This method effectively replaces the traditional analysis and design cycle that heavily relies on human experience, significantly reducing human error and subjective bias, and achieving standardization and high reusability in the analysis process of complex parameter systems.

[0035] In data processing, this invention effectively decouples parameters by using "control variable" samples and performing normalization calculations. The resulting sensitivity indicators are direct and accurate, significantly superior to the rough judgments of traditional single-factor experiments. To further overcome the limitations of traditional passive analysis, this invention also includes a data self-checking module, enabling the system to proactively diagnose the limitations of existing data and generate targeted supplementary testing suggestions with clear physical and mathematical significance, transforming the optimization process from passive analysis to active exploration.

[0036] The intelligent agents in this invention form an efficient iterative closed loop through standardized instruction interaction, enabling rapid focusing on the key parameters and optimal design range that have the greatest impact on performance constraints. Compared to traditional trial-and-error or mesh scanning methods, it achieves the optimization goal with orders of magnitude fewer simulations or experiments. Furthermore, this method is highly scalable; it can be quickly and easily transferred to the design optimization problems of other multi-parameter, multi-constraint, and highly nonlinear power electronic devices or complex industrial systems simply by adjusting the domain knowledge definition and parameter set of the LLM. Attached Figure Description

[0037] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0038] Figure 1 This is an overall framework diagram of the device main circuit parameter sensitivity analysis method based on LLM-Agent of the present invention;

[0039] Figure 2 This is a flowchart of the device main circuit parameter sensitivity analysis method based on LLM-Agent of the present invention;

[0040] Figure 3 This is a schematic diagram of the sensitivity level matrix under a certain operating condition.

[0041] Figure 4 This is a schematic diagram of the sensitivity correlation matrix under a certain working condition.

[0042] Figure 5 A comparison of the phase-to-phase second harmonic circulating current waveforms when the bridge arm inductance is 40mH and 50mH;

[0043] Figure 6 The diagram shows a comparison of the bridge arm capacitor voltage waveforms when the bridge arm inductance is 40mH and 50mH. Detailed Implementation

[0044] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. However, this is not the only description; all aspects not described in detail herein are based on conventional techniques in the art.

[0045] Example 1

[0046] A method for sensitivity analysis of device main circuit parameters based on LLM-Agent, such as Figure 1 and Figure 2 As shown, it includes the following steps:

[0047] S1, the user inputs necessary grid operating condition information and system design constraints through the interactive interface. This information will serve as the baseline input for subsequent parameter generation LLM, defining strict physical boundaries and feasible regions for reasonable parameter estimation. The system acquires the user-inputted grid operating conditions and generates an initial parameter set through parameter generation LLM.

[0048] S2, input the initial parameter set into the steady-state analysis tool for solution, and obtain a steady-state dataset containing the safety margins of each set of parameters;

[0049] S3. Based on sensitivity analysis LLM, decoupling analysis is performed on the steady-state dataset to generate a preliminary sensitivity analysis matrix;

[0050] S4. The sensitivity analysis matrix and steady-state dataset are tested for confidence using the data self-checking LLM. If the test is successful, the final sensitivity analysis matrix is ​​output. If the test fails, a supplementary test suggestion is generated and fed back to the parameter generation LLM. The parameter generation LLM updates the parameter set according to the supplementary test suggestion and returns to step S2, thus forming a closed-loop iteration to achieve automated and intelligent evaluation of parameter sensitivity.

[0051] In this invention, the architecture of parameter generation LLM, sensitivity analysis LLM, and data self-testing LLM is the same. The difference lies in the input prompt words. Through prompt word engineering, different role settings, task backgrounds, processing rules, and output format requirements are given to the same large model, thereby instantiating it into three independent and collaborative functional modules.

[0052] The parameter generation LLM configuration is set to switch between two modes: initial generation and closed-loop optimization. It is used to generate a structured parameter set based on the initial operating conditions of the system or self-test and supplementary test suggestions.

[0053] The sensitivity analysis LLM configuration is used to perform deep decoupling analysis on steady-state datasets. By matching control variable sample pairs, it calculates the quantitative impact of changes in each parameter on the safety margin of key performance constraints of the equipment.

[0054] The data self-checking LLM configuration performs comprehensive verification on steady-state datasets, automatically diagnoses sample defects, and generates structured retesting suggestions.

[0055] Example 2

[0056] A device main circuit parameter sensitivity analysis method based on LLM-Agent, as described in Example 1, differs in that the parameter generation LLM corresponds to an initial generation mode and a closed-loop optimization mode when generating the initial parameter set and updating the parameter set, respectively. Specifically:

[0057] In the initial generation mode, multiple sets of benchmark parameter sets with different design preferences are generated based on the power electronics experience inference rules, and a single-factor comprehensive scan is performed on each set of benchmark parameter sets, i.e., single-variable step adjustment, to generate the initial parameter set.

[0058] In this embodiment, to prevent a single initial value from causing the optimization process to get stuck in a local optimum, the parameter generation LLM constructs m sets of benchmark parameters, each focusing on different engineering objectives such as high efficiency, low cost, and high reliability, to ensure that the initial samples are widely distributed across different regions of the solution space. Subsequently, a strict univariate step adjustment is performed on each parameter set, that is, only a single parameter is changed by a preset step size each time, while the other parameters remain unchanged from their benchmark values, generating an initial parameter set containing multiple sets of samples.

[0059] In closed-loop optimization mode, the preferred parameter combination specified in the supplementary test suggestion is extracted as the new benchmark center to generate an updated parameter set.

[0060] Example 3

[0061] A device main circuit parameter sensitivity analysis method based on LLM-Agent, as described in Example 2, differs in that, in step S2, relying on steady-state analysis tools, the m initial parameter sets generated in step S1 are batch calculated, and the steady-state constraint electrical quantities such as the peak voltage of the submodule capacitor, the modulation signal and the interphase circulating current corresponding to each set of parameters are calculated by the following three formulas respectively;

[0062]

[0063]

[0064]

[0065] in, This represents the DC component of the capacitor voltage. Indicates the capacitance value of the submodule. This represents the current flowing through the capacitor of the submodule. This represents the DC component of the modulated signal. and This represents the amplitude and initial phase angle of the l-th harmonic component. This represents the DC component of the bridge arm current. and These represent the amplitude and phase angle of the k-th harmonic component, respectively. The fundamental angular frequency of the system, For time variables, This represents the harmonic order. This represents the modulation function of the upper arm of phase A. This indicates the internal circulation flowing through the A-phase bridge arm.

[0066] Subsequently, the obtained steady-state constraint electrical quantities are compared with the safety physical thresholds of the current system operating conditions. The difference between the safety physical thresholds and the actually calculated steady-state constraint electrical quantities is calculated to obtain the safety margin Mar of each key performance indicator, thus forming a steady-state dataset.

[0067] Example 4

[0068] A device main circuit parameter sensitivity analysis method based on LLM-Agent, as described in Example 3, differs in that, in step S3, the sensitivity analysis LLM filters the steady-state dataset to find two sets of data. and It can satisfy the condition that only one parameter is different, and the rest are different. -1 parameters are exactly the same. The total number of parameters is expressed by the following formula:

[0069]

[0070] in, Indicates the index of the target parameter that has changed only, while Indicates except Other parameters besides these, which act as control variables and remain strictly consistent across the two sets of data, i.e., only in position... Above, vector and The elements are different, but all other positions are the same.

[0071] To eliminate the influence of different parameter dimensions and variation ranges, the sensitivity analysis LLM normalizes the safety margin based on a preset minimum step unit, as shown in the following formula:

[0072]

[0073] in, This represents the normalized sensitivity index. , These represent the safety margins of the key performance indicators calculated under data A and data B, respectively. , These represent the specific values ​​of the target parameter that undergoes univariate step adjustment in data A and data B, respectively. This indicates the minimum step unit preset for the target parameters.

[0074] Based on the normalized sensitivity index, the sensitivity intensity and influence direction of each parameter are comprehensively judged to obtain the sensitivity analysis matrix. The process is as follows:

[0075] Regarding sensitivity intensity, multiple threshold values ​​for change are preset, including a lower threshold value. T 1. Intermediate threshold T n / 2 and upper limit threshold T n When the sensitivity index is less than 0.01, it is judged as having no sensitivity; when the sensitivity index is greater than or equal to 0.01 and less than... T A value of 1 indicates weak sensitivity; when the sensitivity index is greater than or equal to... T 1 and less than T n / 2 When the sensitivity is determined to be moderate; when the sensitivity index is greater than or equal to T n / 2 and less than T n When the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity; when the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity. T n The time-sensitive determination is extremely high; in this embodiment, T 1 is 0.1. T n / 2 Take 5. T n Take 10.

[0076] Regarding the direction of influence, if the increase in parameters leads to an increase in the safety margin of the performance index, i.e., the sensitivity index is greater than zero, it is judged as a positive correlation, i.e., the system safety is improved; if the increase in parameters leads to a decrease in the safety margin of the performance index, i.e., the sensitivity index is less than zero, it is judged as a negative correlation, i.e., the system safety is reduced. Finally, a preliminary quantitative sensitivity analysis matrix is ​​output.

[0077] Example 5

[0078] A device main circuit parameter sensitivity analysis method based on LLM-Agent, as described in Example 4, differs in that in step S4, the data self-check LLM first performs a defect self-check on the steady-state dataset from three aspects, including missing control variables, conflict, and sparse sampling. When any one of the defects is triggered, it is determined that the self-check has failed and is marked as a low confidence region.

[0079] For low-confidence areas, the data self-checking LLM will generate a structured supplementary test suggestion containing specific reasons and resampling instructions, and feed it back to the parameter generation LLM. After receiving the supplementary test suggestion, the parameter generation LLM triggers a closed-loop optimization mode, extracts the preferred parameter combination specified in the supplementary test suggestion as a new baseline center, reduces the step size or densifies the sampling to regenerate the supplementary parameter set.

[0080] If none of the three defects are triggered, the self-test is considered passed, meaning the sensitivity analysis matrix is ​​converged and stable with high confidence. The iteration then terminates, and the final sensitivity analysis matrix is ​​output, providing a reliable decision-making basis for equipment parameter optimization and trade-off design.

[0081] Sensitivity analysis matrix quantification reveals the intensity and direction of the influence of each main circuit parameter on key performance constraints, providing a reliable basis for parameter optimization design of the equipment, including sensitivity level matrix and sensitivity correlation matrix.

[0082] The sensitivity level matrix is ​​obtained by dividing the absolute values ​​of normalized sensitivity indices into intervals; the sensitivity correlation matrix is ​​obtained by determining the sign of the normalized sensitivity indices. The sensitivity level matrix transforms complex matrix values ​​into intuitive "correlation" and "level," aiming to provide R&D personnel with direct and actionable guidance for parameter optimization. R&D personnel do not need to analyze specific numerical values; they only need to identify the core parameters with the greatest impact on the system based on the "level" and prioritize their adjustment, and determine whether to increase or decrease the parameters based on the "correlation," thereby significantly improving the efficiency of equipment parameter optimization design.

[0083] Example 6

[0084] A device main circuit parameter sensitivity analysis method based on LLM-Agent, as described in Example 5, except that when there are parameters in the steady-state dataset that cannot be independently decoupled for analysis, it is considered that the control variable is missing.

[0085] A conflict exists when the same parameter exhibits significant differences in sensitivity indices or nonlinear abrupt changes that reverse the direction of influence in different parameter sets. Specifically, multiple sensitivity indices for the same parameter are extracted from different sample pairs, and the variance or relative standard deviation of the set of sensitivity indices is calculated. If the variance exceeds a preset nonlinear tolerance threshold, the sensitivity indices are considered to have significant differences. At the same time, the signs of the set of sensitivity indices are checked. If both positive and negative values ​​are included, the direction of influence is considered to have reversed, i.e., there is a nonlinear abrupt change that reverses the direction of influence.

[0086] Calculate the actual change step size between adjacent sample points of the same parameter in the steady-state dataset. If the actual change step size is greater than the preset maximum allowable sampling step size, it is determined that there is sampling sparsity.

[0087] To further verify the effectiveness and engineering applicability of the method proposed in this invention, a detailed explanation is provided below with reference to specific MMC design examples.

[0088] A 500kV flexible DC transmission system was selected as the specific design object. The user first inputs the key grid operating condition information required for designing the MMC (Multi-Level Control) into the system via the human-machine interface. The specific input operating condition information and system parameters are shown in Table 1.

[0089] Table 1 shows the input operating condition information and system parameters.

[0090]

[0091] The user inputs the power grid operating condition information shown in Table 1 into the parameter generation LLM. Based on pre-trained knowledge in the field of power electronics, the agent automatically plans and outputs the first round of 60 sets of initial experimental parameter matrices. Subsequently, the system inputs these 60 sets of initial parameters into a steady-state analysis tool. After rigorous physical constraint verification and electrical quantity calculation, a steady-state dataset containing 60 sets of steady-state constraint electrical quantities and safety margins is obtained, and then submitted to the sensitivity analysis LLM to generate a preliminary sensitivity analysis matrix.

[0092] To avoid "local optima" and "data blind spots" that are prone to occur in traditional analysis, the system sends the sensitivity analysis matrix and steady-state dataset to the data self-testing LLM for confidence diagnosis. The data self-testing LLM successfully identified two low-confidence areas in the first round of data and provided structured suggestions for targeted supplementary testing: 1) Conflict exists: The diagnosis found that the bridge arm inductance L m Within the 55mH to 65mH range, the interphase circulating current exhibits extreme nonlinear sensitivity, with dramatic abrupt changes in value. The system recommends an increase in bridge arm inductance L... m Within the interval [50, 100], encrypted sampling with a step size of 5 was performed to accurately locate the inflection point of the circulating current suppression curve. 2) Sparse sampling: Diagnosis revealed that the transformer turns ratio K T This is a key parameter determining whether the maximum modulation ratio exceeds the limit, but the current step size of 0.05 is too large to accurately locate the feasible region boundary. The system suggests K. T Local encrypted sampling is performed in the interval [1.0, 1.1] with a small step size of 0.01.

[0093] The system automatically feeds back the above supplementary testing suggestions and historical data to the parameter generation LLM. Upon receiving the instruction, the parameter generation LLM uses the specified parameter combination as the new reference center for the bridge arm inductance L. m and K TThe sensitive intervals generated a second set of 60 supplementary parameters. These 60 supplementary parameters were then recalculated using a steady-state analysis tool, and the results were merged with the first set of data to form a steady-state dataset containing 120 high-quality samples. The sensitivity analysis matrix was then updated using sensitivity analysis LLM.

[0094] The updated analysis results and 120 sets of steady-state datasets were sent again to the data self-checking LLM for a second self-check. The diagnostic results showed that the system still needed further optimization and pinpointed a deeper gradient resolution problem: the diagnosis found that in U dc =800V region, bridge arm inductance L m Within the 50mH to 60mH range, the interphase circulating current exhibits extreme sensitivity, plummeting from approximately 5936A to approximately 2338A. The system determines that the current sampling step size of 5mH is too coarse to capture this steep gradient, resulting in poor resolution at the feasible region boundary. The system automatically outputs a new instruction: fix other parameters, and adjust the bridge arm inductance L... m Locally refined sampling with a step size of 2mH is performed within the interval [50,70]. Based on this instruction, the parameter generation LLM generates a third round of supplementary parameter sets and completes steady-state calculations. The latest results are then incorporated into the total dataset for a third round of sensitivity analysis.

[0095] The final dataset and analysis results, after multiple rounds of iterative expansion, were sent to the data self-checking LLM for the third time. This self-check confirmed that the steady-state dataset had high confidence, and no issues such as missing control variables, conflicts, or sparse sampling were found. The sensitivity analysis matrix had converged and stabilized. Finally, the system directly output the final sensitivity analysis matrix under this high confidence level, providing a decision-making basis for subsequent design.

[0096] In generating the final sensitivity analysis matrix, the sensitivity analysis LLM performs a deep traversal of all data, precisely matching sample pairs in the dataset that differ only in a single parameter, thus eliminating coupling effects and independently analyzing the capacitance C of the seven input parameter submodules. sm Bridge arm inductor L m Short-circuit impedance percentage Uk, transformer rated capacity S Tn Transformer turns ratio K T DC side voltage U dcThe specific impact of changes in the number of bridge arm submodules N on the margins of 12 constraint indicators is investigated. After normalization calculations, the final outputs are a sensitivity level matrix and a sensitivity correlation matrix. The sensitivity correlation matrix focuses on "direction," extracting and visualizing the sign of the values ​​in the "sensitivity analysis matrix." It considers only the sign, not the magnitude of the values, to guide the direction of parameter adjustments for researchers. The sensitivity level matrix focuses on "intensity," performing thresholding and visualization of the absolute values ​​in the "sensitivity analysis matrix." It considers only the absolute value, not the sign, to guide the priority of parameter adjustments for researchers, such as... Figure 3 and Figure 4 As shown.

[0097] Figure 3 and Figure 4 Two sets of matrices comprehensively reveal the strength and direction of the influence of main circuit parameters on key performance constraints in a clear and quantitative manner. The sensitivity level matrix normalizes the influence strength of parameters and divides it into discrete magnitudes of 0 to 10, with darker colors representing higher sensitivity. The sensitivity correlation matrix precisely identifies positive correlation, negative correlation, and no correlation using +1, -1, and 0, respectively. Figure 3 and Figure 4 It can accurately determine the positive and negative impacts and magnitudes of each main circuit parameter on each constraint index under the input power grid operating conditions.

[0098] Compared to traditional methods that rely on manual trial and error and grid traversal, this invention successfully avoids nonlinear mutation regions and sparse sampling blind spots through multi-round active diagnosis and closed-loop iteration using an LLM-Agent. This method not only significantly reduces unnecessary computation but also produces sensitivity quantification results that highly align with the underlying physical laws of power electronics, providing a direct and reliable decision-making basis for multi-parameter decoupling and collaborative optimization of complex converter equipment.

[0099] To further verify the effectiveness and accuracy of the sensitivity level matrix and sensitivity correlation matrix obtained by the method proposed in this invention, the following verification is carried out in conjunction with the embodiments. The fixed main circuit parameters of MMC in the embodiments are shown in Table 2.

[0100] Table 2 MMC Fixed Main Circuit Parameters

[0101]

[0102] Figure 5 This demonstrates the bridge arm inductance L under the aforementioned main circuit parameters. m The waveforms of the phase-to-phase second harmonic circulating current are shown at 40mH and 50mH, respectively. Under this operating condition, the safety constraint value of the second harmonic circulating current is 776.13A. As can be seen from the figure, when the bridge arm inductance L... mWhen the inductance is increased from 40mH to 50mH, the peak value of the second harmonic circulating current decreases significantly from 774A to 562A. This indicates that the bridge arm inductance L... m The increase was only 10mH, while the amplitude of the interphase second harmonic circulating current decreased by 212A, a relative change rate of 27.39%. Simultaneously, the circulating current margin of the MMC increased significantly from 2.13A to 214.13A. Taking the bridge arm inductor L... m The absolute sensitivity to circulating current margin is 21.2 when calculated with a unit improvement step size of 1mH. Since this value is greater than 10, the sensitivity level of the bridge arm inductance change to interphase second harmonic circulating current is rated as 10. In addition, the increase in inductance effectively suppresses circulating current and significantly increases the circulating current margin, and the two are strongly positively correlated.

[0103] Figure 6 This demonstrates the bridge arm inductance L under the same main circuit parameters. m The waveforms of the bridge arm capacitor voltage at 40mH and 50mH are shown respectively. Under this operating condition, the safe limit value for the peak value of the bridge arm capacitor voltage is 2523V. As can be seen from the figure, when the bridge arm inductance L... m When the inductance L increases from 40mH to 50mH, the peak voltage of the bridge arm capacitor increases slightly from 2423V to 2438V. This shows that the bridge arm inductance L... m The capacitance increased by 10mH, while the bridge arm capacitor voltage amplitude only increased by 15V, a very small relative change of only 0.62%. Correspondingly, the safety margin of the MMC capacitor voltage peak decreased from 100V to 85V, with an absolute change in margin of 15V. If L... m Calculated using a unit improvement step size of 1mH, its absolute sensitivity to the peak margin of the capacitor voltage is 1.5. Since this value falls between 1 and 2, and the overall variation is small, the overall sensitivity level of the bridge arm inductor to the peak margin of the capacitor voltage is rated as 1. This shows that the bridge arm inductor and the peak margin of the capacitor voltage are negatively correlated, but the overall impact is very weak.

[0104] The quantitative analysis results of the aforementioned time-domain simulation waveforms are in perfect agreement with the sensitivity level matrix and sensitivity correlation matrix obtained by the method proposed in this invention. Specifically, in the sensitivity level matrix, the sensitivity level of the bridge arm inductance to the interphase circulating current is evaluated as the highest level "10", corresponding to a significant change of 27.39%; while the sensitivity level to the peak capacitor voltage is evaluated as "1", corresponding to a tiny change of only 0.62%. The sensitivity correlation matrix also accurately reflects the strong positive correlation between the increase in bridge arm inductance and the increase in circulating current margin, as well as the weak negative correlation with the peak capacitor voltage. This fully demonstrates that the sensitivity level matrix and sensitivity correlation matrix proposed in this invention can accurately and intuitively quantify the degree and direction of influence of each main circuit parameter on the key electrical quantities of the system, verifying the effectiveness and accuracy of the evaluation method, thus providing a reliable theoretical basis for the multi-objective optimization design of MMC systems.

[0105] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for sensitivity analysis of device main circuit parameters based on LLM-Agent, characterized in that, Includes the following steps: S1: Obtain the power grid operating conditions input by the user, and generate an initial parameter set through parameter generation LLM; S2, input the initial parameter set into the steady-state analysis tool for solution, and obtain a steady-state dataset containing the safety margins of each set of parameters; S3. Based on sensitivity analysis LLM, decoupling analysis is performed on the steady-state dataset to generate a preliminary sensitivity analysis matrix; S4, perform confidence diagnosis on the sensitivity analysis matrix and steady-state dataset through data self-checking LLM; If the diagnosis is successful, the final sensitivity analysis matrix will be output. If the diagnosis fails, a supplementary test suggestion is generated and fed back to the parameter generation LLM. The parameter generation LLM updates the parameter set according to the supplementary test suggestion and returns to step S2, thus forming a closed loop iteration to realize the automated and intelligent evaluation of parameter sensitivity. In the parameter generation LLM, the initial parameter set generation and parameter set update correspond to the initial generation mode and closed-loop optimization mode, respectively, as follows: In the initial generation mode, multiple sets of baseline parameter sets with different design preferences are estimated based on physical experience, and a single-variable step adjustment is performed on each set of baseline parameter sets to generate the initial parameter set. In closed-loop optimization mode, the preferred parameter combination specified in the supplementary test suggestion is extracted as the new benchmark center to generate an updated parameter set; In step S3, the sensitivity analysis LLM filters the steady-state dataset to find two sets of data. and It can satisfy the condition that only one parameter is different, and the rest are different. -1 parameters are exactly the same; Indicates the total number of parameters; To eliminate the influence of different parameter dimensions and variation ranges, the sensitivity analysis LLM normalizes the safety margin based on a preset minimum step unit, as shown in the following formula: in, This represents the normalized sensitivity index. , These represent the safety margins of the key performance indicators calculated under data A and data B, respectively. , These represent the specific values ​​of the target parameter that undergoes univariate step adjustment in data A and data B, respectively. This indicates the minimum step unit preset for the target parameters; Based on the normalized sensitivity index, the sensitivity intensity and influence direction of each parameter are comprehensively judged to obtain the sensitivity analysis matrix. The process is as follows: Regarding sensitivity intensity, multiple threshold values ​​for change are preset, including a lower threshold value. T 1. Intermediate threshold T n / 2 and upper limit threshold T n When the sensitivity index is less than 0.01, it is judged as having no sensitivity; when the sensitivity index is greater than or equal to 0.01 and less than... T A value of 1 indicates weak sensitivity; when the sensitivity index is greater than or equal to... T 1 and less than T n / 2 When the sensitivity is determined to be moderate; when the sensitivity index is greater than or equal to T n / 2 and less than T n When the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity; when the sensitivity index is greater than or equal to the specified value, it is considered to have high sensitivity. T n The time-sensitive determination is extremely high. Regarding the direction of influence, if the increase in parameters leads to an increase in the safety margin of the performance index, i.e., the sensitivity index is greater than zero, it is judged as a positive correlation, i.e., the system safety is improved; if the increase in parameters leads to a decrease in the safety margin of the performance index, i.e., the sensitivity index is less than zero, it is judged as a negative correlation, i.e., the system safety is reduced. Finally, a preliminary sensitivity analysis matrix is ​​output. In step S4, the data self-check LLM first performs a defect self-check on the steady-state dataset from three aspects, including missing control variables, conflict, and sparse sampling. When any of these defects is triggered, the self-check is deemed to have failed and is marked as a low-confidence region. For low-confidence areas, the data self-checking LLM will generate a structured supplementary test suggestion containing specific reasons and resampling instructions, and feed it back to the parameter generation LLM. After receiving the supplementary test suggestion, the parameter generation LLM triggers a closed-loop optimization mode, extracts the preferred parameter combination specified in the supplementary test suggestion as a new baseline center, reduces the step size or densifies the sampling to regenerate the supplementary parameter set. If none of the three defects are triggered, the self-test is considered passed, the iteration terminates, and the final sensitivity analysis matrix is ​​output.

2. The method for sensitivity analysis of device main circuit parameters based on LLM-Agent according to claim 1, characterized in that, In step S2, relying on the steady-state analysis tool, the m sets of initial parameters generated in step S1 are calculated in batches to obtain the steady-state constraint electrical quantities corresponding to each set of parameters. Subsequently, the obtained steady-state constraint electrical quantities are compared with the safety physical threshold of the current system operating condition, and the difference between the safety physical threshold and the actual calculated steady-state constraint electrical quantities is calculated to obtain the safety margin Mar of each key performance indicator, thus forming a steady-state dataset.

3. The method for sensitivity analysis of device main circuit parameters based on LLM-Agent according to claim 2, characterized in that, When there are parameters in the steady-state dataset that cannot be independently decoupled for analysis, the control variables are missing. A conflict exists when the same parameter exhibits significant differences in sensitivity indices or nonlinear abrupt changes in the direction of influence across different parameter sets. Calculate the actual change step size between adjacent sample points of the same parameter in the steady-state dataset. If the actual change step size is greater than the preset maximum allowable sampling step size, it is determined that there is sampling sparsity.

Images