Adaptive testing method based on large model and digital twinning and intelligent simulation system

Abstract

The application provides an adaptive testing method based on a large model and digital twinning and an intelligent simulation system, and the method comprises the following steps: constructing a digital twinning model of a power electronic system, checking the compliance and safety of a modeling instruction, ensuring that the model parameter structure meets the requirements, and generating a standardized digital twinning body for simulation testing; collecting key operation data of the power electronic system, matching potential root causes through a feature mode database of the digital twinning model when an abnormal working condition is detected, and generating comprehensive diagnosis suggestions in combination with the system structure and the test scene; summarizing test data to generate a test report, extracting structured knowledge units containing fault mode labels, fault feature sets, root cause suggestion lists, and verification test conclusions, inputting the structured knowledge units into an enterprise knowledge base and associating the corresponding system modules, and automatically prompting corresponding fault risks and mitigation solutions when subsequent similar system modeling is performed. The application realizes intelligent enhancement of the whole process from design modeling, online monitoring to automatic testing.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to an adaptive testing method and intelligent simulation system based on large models and digital twins. Background Technology

[0002] Hardware-in-the-loop simulation is a core component of the research, verification, and certification of complex power electronic systems. While mainstream simulation platforms, such as the Typhoon HIL series, offer a full-stack toolchain from graphical modeling (Schematic Editor) and real-time monitoring (HILSCADA) to automated testing (TyphoonTest IDE), their level of intelligence is insufficient to cope with the increasing complexity, primarily due to the following bottlenecks: Low modeling efficiency: Modeling complex systems relies heavily on expert experience, and parameter selection and topology construction are time-consuming and labor-intensive, with a high barrier to entry for beginners.

[0003] Data deluge and lack of insights: SCADA interfaces provide massive amounts of data, but lack automatic analysis and diagnostic capabilities. Engineers need to manually filter and correlate the data, which can easily lead to missing key signs.

[0004] Limitations of automated testing: Existing tests rely heavily on pre-written static scripts, making it difficult to autonomously explore performance boundaries and discover potential defects under non-preset operating conditions. The adequacy of tests depends on the engineer's experience.

[0005] Knowledge silos: Valuable debugging experience, fault models, and optimization strategies in projects often exist in the form of unstructured documents, which are difficult to reuse directly in new projects, resulting in a waste of knowledge.

[0006] Meanwhile, artificial intelligence technologies, represented by Large Language Models (LLM), have demonstrated powerful command-following, code generation, and reasoning capabilities after being fine-tuned with high-quality data in specialized fields; while digital twin technology provides a framework for data fusion and knowledge accumulation between physical objects and virtual models. However, how to deeply and systematically integrate the two to form a new paradigm for simulation testing that combines efficiency, security, and self-evolution remains a pressing technical challenge. Summary of the Invention

[0007] This invention addresses the technical problems existing in the prior art by providing an adaptive testing method and intelligent simulation system based on large models and digital twins. By constructing a collaborative framework between an AI large model and a holographic digital twin kernel, it achieves intelligent enhancement throughout the entire process from design modeling and online monitoring to automated testing.

[0008] According to a first aspect of the present invention, an adaptive testing method based on large models and digital twins is provided, comprising the following steps: Digital twin models of power electronic systems are generated through natural language interaction, and the modeling instructions are verified for compliance and security to ensure that the model parameters and structure conform to engineering safety boundaries and physical laws, thereby generating standardized digital twins for simulation testing. Collect key operating data of power electronic systems, and when abnormal operating conditions are detected, match potential root causes through the feature pattern database of the digital twin model, and generate comprehensive diagnostic suggestions by combining system structure and test scenario; The test data is aggregated to generate a test report. Structured knowledge units containing fault mode labels, fault feature sets, root cause suggestion lists, and verification test conclusions are extracted, entered into the enterprise knowledge base, and associated with the corresponding system modules. When modeling similar systems in the future, the corresponding fault risks and mitigation solutions will be automatically prompted.

[0009] Based on the above technical solution, the present invention can also be improved as follows.

[0010] Optionally, the compliance and security verification of the modeling instructions includes: The digital twin model performs semantic parsing of natural language modeling instructions, verifies whether the model parameters meet the physical constraints and engineering safety tolerances of the power electronic system, and automatically provides correction suggestions and remodels if there are parameter out-of-bounds or structural conflicts, until a digital twin that meets the requirements is generated.

[0011] Optionally, the key operating data of the power electronic system collected includes: load step response, harmonic spectrum analysis, and current imbalance; the process of matching potential root causes through the feature pattern database of the digital twin model and generating comprehensive diagnostic suggestions based on the system structure and test scenario includes: By retrieving matching feature patterns from the performance library of the digital twin model, the root cause analysis covers potential reasons such as insufficient control loop bandwidth, excessively high values ​​of the current loop integral term, and phase-locked loop frequency offset. Combined with the coupling characteristics of the filter components and the test scenario, precise improvement suggestions are given.

[0012] Optionally, the digital twin model includes: The L3 planning layer is used to develop a structured test outline in conjunction with the digital twin kernel, clarifying the tasks, priorities, and security baselines for baseline testing and exploratory testing. The L2 scheduling layer is used to execute standard test scripts in batches and track the results. The L1 restricted exploration layer uses a reinforcement learning algorithm with safety barriers to explore within a limited parameter space. When a safety threshold is triggered, the exploration is forcibly interrupted and the strategy is guided to avoid the danger zone, thus uncovering potential problems under non-standard working conditions.

[0013] Optionally, the baseline test includes routine operating condition test and fault injection test; wherein, the routine operating condition test examines the impact of grid frequency changes on total harmonic distortion rate; the fault injection test simulates three-phase voltage drops of different degrees and durations, records recovery response characteristics and monitors device thermal loss.

[0014] Optionally, the safety baseline includes hard protection triggered by DC voltage exceeding ±15% of the rated voltage and control loop key parameter variations not exceeding 20% ​​of the engineering tolerance.

[0015] Optionally, the exploratory test includes boundary search and anomalous data injection; the boundary search, within the 20% tolerance of the control loop parameters, fine-tunes the controller parameters through reinforcement learning to test the impact of small changes in grid frequency on the total harmonic distortion rate; the anomalous data injection sets up asymmetrical grid voltage disturbances to explore the changes in the internal circulating current of the control loop under unbalanced voltage.

[0016] Optionally, the reinforcement learning algorithm with safety guardrail includes: defining a state space containing grid frequency offset, filter capacitor voltage harmonic values, and controller bandwidth parameters; setting a safety limiter to forcibly interrupt the current test and impose a penalty on the reinforcement learning strategy when the DC voltage change rate exceeds the limit or the operating condition enters an unstable region, thereby guiding the strategy to avoid dangerous operating conditions.

[0017] Optionally, the method further includes: For anomaly diagnosis and testing suggestions, an interactive button is provided to confirm the start of parameter optimization or verification scripts with one click, supporting engineers to approve and execute as needed, and realizing the collaboration between intelligent decision-making and manual intervention.

[0018] According to a second aspect of the present invention, an intelligent simulation system is provided, comprising: The semantic digital twin modeling module is used to generate digital twin models of power electronic systems through natural language interaction, and to verify the compliance and security of modeling instructions to ensure that the model parameters and structure conform to the engineering safety boundaries and physical laws, and to generate standardized digital twins for simulation testing. The real-time monitoring and diagnosis module includes a real-time feature extraction unit, an anomaly detection unit, and a root cause analysis unit. It is used to collect key operating data such as load step response and harmonic spectrum, match feature patterns to locate anomalies, and generate comprehensive diagnostic suggestions based on system structure and scenario. The knowledge management and closed-loop iteration module is used to summarize test data to generate special reports, extract structured knowledge units and enter them into the enterprise knowledge base, and associate them with corresponding system modules. When modeling similar systems in the future, it will automatically prompt the corresponding fault risks and mitigation solutions, realizing knowledge reuse and a closed-loop process.

[0019] The technical effects and advantages of this invention are as follows: This invention provides an adaptive testing method and intelligent simulation system based on large-scale models and digital twins. By constructing a collaborative framework between an AI large-scale model and a holographic digital twin kernel, it achieves intelligent enhancement across the entire process, from design modeling and online monitoring to automated testing. The synergy between the large-scale model and the digital twin runs through the entire process: compliance verification in the modeling stage, root cause identification in the monitoring stage, task planning and security exploration in the testing stage, and structured extraction and association in the knowledge accumulation stage, realizing full-link intelligentization of power electronic system design and testing. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an adaptive testing method based on a large model and digital twin, provided as an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Understandably, given the deficiencies in the background technology, this invention proposes an adaptive testing method based on large models and digital twins, specifically as follows: Figure 1 As shown, it includes the following steps: Step S1: Construct a semantic digital twin model and complete modeling safety verification; Generate a digital twin model of the power electronic system through natural language interaction, and verify the compliance and safety of the modeling instructions to ensure that the model parameters and structure conform to the engineering safety boundaries and physical laws, and generate a standardized digital twin for simulation testing; In this embodiment, a digital twin model of a power electronic system is generated through natural language interaction. The digital twin model includes: an L3 planning layer, an L2 scheduling layer, and an L1 restricted exploration layer; wherein, The L3 planning layer, combined with the digital twin kernel, formulates a structured test outline, clarifying the tasks, priorities, and safety baselines for baseline testing and exploratory testing; the L2 scheduling layer is used to batch execute standard test scripts and track the results; the L1 restricted exploration layer uses a reinforcement learning algorithm with safety barriers to explore within a limited parameter space, forcibly interrupting and guiding the strategy to avoid dangerous areas when a safety threshold is triggered, thus uncovering potential problems under non-standard operating conditions.

[0023] Furthermore, the L3 planning layer, combined with the digital twin kernel, formulates a structured test outline, clarifying the tasks and priorities of baseline testing and exploratory testing, as well as the safety baselines: DC voltage exceeding ±15% of the rated voltage triggers hard protection, and key control loop parameter variations do not exceed the 20% engineering tolerance. The baseline testing includes routine operating condition testing and fault injection testing: routine operating condition testing examines the impact of grid frequency changes on the total harmonic distortion (THD); fault injection testing simulates three-phase voltage drops of varying degrees and durations, records recovery response characteristics, and monitors device thermal losses. The safety baselines include DC voltage exceeding ±15% of the rated voltage triggering hard protection and key control loop parameter variations not exceeding the 20% engineering tolerance. The exploratory testing includes boundary search and anomalous data injection: boundary search, within the ±20% tolerance of control loop parameters, fine-tunes controller parameters through reinforcement learning to test the impact of small grid frequency mutations on the THD; anomalous data injection sets up asymmetrical grid voltage disturbances to explore the changes in the internal circulating current of the control loop under unbalanced voltage.

[0024] The L1 restricted exploration layer employs a reinforcement learning algorithm with safety barriers. It defines a state space containing parameters such as grid frequency offset and an action space containing parameters such as frequency disturbances. Exploring within the limited parameter space, when the DC voltage change rate exceeds the limit or the operating condition enters an unstable region, the current test is forcibly interrupted and a penalty is imposed on the strategy, guiding it to avoid dangerous areas and uncover potential problems under non-standard operating conditions. The reinforcement learning algorithm with safety barriers includes: defining a state space containing parameters such as grid frequency offset, filter capacitor voltage harmonic values, and controller bandwidth; and an action space containing parameters such as frequency disturbances, parameter adjustment, and test type selection; setting a safety limiter, which forcibly interrupts the current test and imposes a penalty on the reinforcement learning strategy when the DC voltage change rate exceeds the limit or the operating condition enters an unstable region, guiding the strategy to avoid dangerous operating conditions.

[0025] It should be noted that the digital twin model described in this embodiment is a semantic layered architecture model that integrates the capabilities of large models. It supports natural language interactive modeling and completes compliance and security verification. It constructs a layered test management system consisting of an L3 planning layer (for developing structured test outlines), an L2 scheduling layer (for batch execution of test scripts), and an L1 restricted exploration layer (for reinforcement learning exploration with safety barriers). It can also connect to real-time feature monitoring, knowledge accumulation loop, and human-computer interaction collaboration, accurately addressing the pain points of difficulty in non-standard operating condition discovery, low modeling efficiency, and weak safety boundary control in power electronic system testing. It is specifically designed for adaptive testing scenarios of power electronic systems.

[0026] Existing digital twin models in the field of power electronics mostly focus on the simulation mapping of physical systems or the testing support of single functions, while the digital twin model described in this embodiment of the invention has the following characteristics: Semantic generation capability: Generate models through natural language interaction and conduct compliance and security verification on modeling instructions to ensure that parameters and structures conform to engineering safety boundaries and physical laws; A three-tiered collaborative test management architecture: L3-L2-L1 layers enable test outline formulation, batch script execution, and exploration of non-standard working conditions with safety barriers, covering all scenarios of baseline testing and exploratory testing; End-to-end collaborative capability: Deeply integrated with large models, real-time fault diagnosis, knowledge accumulation, and human-computer interaction, it constructs a closed loop of design-monitoring-testing-learning, realizing end-to-end intelligent design and testing of power electronic systems.

[0027] The compliance and security verification of the modeling instructions includes: The digital twin model performs semantic parsing of natural language modeling instructions, verifies whether the model parameters meet the physical constraints and engineering safety tolerances of the power electronic system, and automatically provides correction suggestions and remodels if there are parameter out-of-bounds or structural conflicts, until a digital twin that meets the requirements is generated.

[0028] Step S2: Real-time feature monitoring and anomaly root cause diagnosis; Collect key operating data of the power electronic system. When an abnormal operating condition is detected, match potential root causes through the feature pattern database of the digital twin model, and generate comprehensive diagnostic suggestions by combining system structure and test scenario. Key operational data of the power electronic system are collected, including load step response, harmonic spectrum analysis, and current imbalance. During anomaly detection, the performance library of the digital twin kernel is used to search for matching feature patterns. Root cause analysis covers potential causes such as insufficient control loop bandwidth, excessively high value of current loop integral term, and phase-locked loop frequency offset. Precise improvement suggestions are given in combination with the coupling characteristics of filter components and test scenarios.

[0029] The comprehensive diagnostic recommendations include improvements such as problem localization, parameter adjustment, and design optimization.

[0030] Step S3: Knowledge Accumulation and Closed-Loop Iteration; Summarize test data to generate test reports, extract structured knowledge units containing fault mode labels, fault feature sets, root cause suggestion lists, and verification test conclusions, enter them into the enterprise knowledge base and associate them with corresponding system modules. When modeling similar systems in the future, the system will automatically prompt the corresponding fault risks and mitigation solutions.

[0031] The test report includes baseline performance, key issues, and optimization suggestions; it extracts structured knowledge units of failure modes, features, root cause suggestions, and verification conclusions. These structured knowledge units are automatically associated with corresponding functional modules in the digital twin model. When modeling similar systems in the future, the report automatically prompts corresponding failure risks and mitigation solutions, enabling continuous accumulation and reuse of knowledge and forming a closed loop of design-monitoring-testing-learning.

[0032] Step S4, Human-Computer Interaction Collaboration: For anomaly diagnosis and testing suggestions, provide an interactive button for one-click confirmation to start parameter optimization or verification scripts, supporting engineers to approve and execute as needed, realizing intelligent decision-making and human intervention collaboration.

[0033] Furthermore, the collaboration between large models and digital twins runs through the entire process, playing a role in compliance verification, root cause correlation, task planning and security exploration, and structured extraction and correlation in each stage of modeling, monitoring, testing, and knowledge accumulation, thereby realizing full-link intelligence in power electronic system design and testing.

[0034] The following specific embodiments, from design to verification, further illustrate this technical solution. This invention provides a photovoltaic inverter simulation and testing method integrating AI and digital twins, including the following steps: Step 1: Intelligent modeling assistance to quickly build a baseline model; Engineers log into the platform and enter their natural language requirements in the "Smart Assistant" dialog box of the Schematic Editor: "Please design a three-phase two-level grid-connected photovoltaic inverter for a solar power plant. Design requirements: rated power 150kW, input DC voltage 800V, output AC grid-connected voltage 380V / 50Hz. The system should include maximum power point tracking, LCL filter, voltage / current dual closed-loop control, complete grid-connected / off-grid switching logic, and overvoltage, overcurrent, and islanding protection. The control loop should be stable and the current harmonics should meet national standards." 1.1. System Identification and Decomposition: Command parsing: The platform's core AI big model (customized for the power electronics field) immediately parses the commands, identifies and extracts: "Topology: Three-phase two-level", "Power: 150kW", "Input: DC800V", "Output: 3AC380V / 50Hz", "Key functions: MPPT, LCL filtering, dual-loop control, mode switching, protection function", "Optimization goal: Loop stability, THDi compliance".

[0035] Semantic encoding and mapping: The AI ​​model maps these keywords to a predefined semantic tag library. Power: 150,000 Topology: 2L_VSI, Application: grid-tied_PV_inverter Core modules: MPPT, LCL_Filter, Dual_Loop_Control, Protection_Logic Performance metrics: overshoot: (<5%), THDi (<3%).

[0036] 1.2. Scheme Planning and Search: The AI ​​engine, based on the modular knowledge graph of power electronics, retrieves and activates node clusters related to the 2L_VSI topology, such as “DC Link Capacitor”, “IGBT 6-Pack”, “LCL Filter”, “MPPT”, “PWM Generator”, “Dual-loop Control (DQ Domain)”, and “Islanding Detection”.

[0037] Using a graph-based matching and combination algorithm, a basic intermediate graph representation (IGR) is automatically constructed based on the interface definition (voltage_in: DC, voltage_out: AC) and connection constraints (DC side connected to DC capacitor, PWM output connected to IGBT driver). The graph includes the main circuit structure and the block diagram connections of each control module.

[0038] 1.3. Dual verification of structure and parameters (critical guardrail mechanism): Static rule validation: The system checks the IGR based on built-in topology-level validation rules. For example, it ensures that: The DC bus has at least one capacitor bank, all IGBT gates have drive signal inputs, current feedback signals flow back to the controller from the correct acquisition point, and grid voltage and frequency signals are connected to the phase-locked loop.

[0039] Dynamic Simulation Preview (Sandbox Mode): The system calls its Fast Reduced-Order Simulation Model (FRM) to generate a simplified simulation model based on IGR. It automatically performs a microsecond-level fast simulation of power-on startup and steady-state operation. This simulation not only checks for direct short circuits that could cause a crash, but also automatically calculates some key initial operating point parameters (such as rated current estimation) and monitors the transient stability of the DC bus voltage. For example, if an oscillation risk is detected at the moment of bus voltage startup, a prompt will be given: "The DC side capacitor value may be too small; it is recommended to increase it to XXuF to enhance stability." Model Delivery and Parameter Recommendation: After confirming the safety baseline is acceptable, the AI ​​engine, combining built-in empirical formulas and physical constraints, outputs a complete circuit model with pre-filled parameters. It generates a detailed design report listing all core parameters and their calculation logic, for example: DC bus capacitor: Recommended value is 2200µF. Calculated based on constraints of energy storage time constant t=0.2ms and ripple voltage <5%.

[0040] LCL filter: Recommended values ​​are L_f (grid-side inductance) = 0.5mH, C_f (filter capacitor) = 15µF, and L_g (grid-side inductance) = 0.1mH. The design is based on a switching frequency f_sw = 10kHz and a resonant frequency f_res = 2.8kHz (between f_sw / 6 and f_sw / 2).

[0041] PI controller parameters: Voltage outer loop Kp_v=0.8, Ki_v=2; Current inner loop Kp_i=3, Ki_i=50. Based on model reference method (pole placement) and empirical tuning values, specific online optimization is available.

[0042] Monitoring interface solution: Automatically provides 8 charts to monitor DC bus voltage / current, three-phase grid voltage / current, active / reactive power, switching device temperature rise, THDi and PF. The interface is laid out according to the commonly used dashboard layout for engineers.

[0043] 1.4. Parameter Optimization and Confirmation: After reviewing the system's provided solution and initial parameters, the engineer determined that the switching frequency needed fine-tuning to reduce losses. The following input was then made: The switching frequency was changed to 8kHz, the filter parameters were recalculated, and grid disturbances (voltage ±10%) and frequency fluctuations (±0.5Hz) were added to the benchmark test procedure.

[0044] AI assistant accepts instructions: The filter values ​​(L_f = 0.6mH, C_f = 20µF) were recalculated to avoid the resonant frequency entering the bandgap, and the corresponding parameters in the circuit diagram were automatically updated.

[0045] This condition was automatically added as one of the primary items for subsequent adaptive testing, and the parameters of interest were annotated synchronously with the digital twin kernel: "grid_frequency_deviation: [49.5, 50.5] Hz".

[0046] After the engineer confirms the parameter modifications, they can click "Accept Design Model and Generate SCADA" with one click.

[0047] 1.5. Automatic Deployment of SCADA Interface: The AI ​​engine quickly creates an operable SCADA monitoring interface for the model, integrating all necessary monitoring panels, instruments, and control buttons (such as "start / stop", "parallel / offline mode switching", and "manual / automatic settings").

[0048] In this step, AI plays the role of "expert designer and efficient assistant," quickly creating a working HIL simulation model from scratch, significantly reducing the workload in the early stages of design.

[0049] Step Two: Cognitive Monitoring and Autonomous Fault Origination; 2.1. Intelligent Operation Monitoring: The engineer enters the following in the command window of the SCADA interface: Run the system and set it to automatic power ramp-up. After startup, run at full power for 15 seconds, then perform a 90% DC voltage sag (lasting 300 milliseconds) to test the grid voltage adaptability. Monitor THDi and overload current.

[0050] The system executes instructions, while the real-time feature extraction module monitors all key data, such as "load step response", "harmonic spectrum analysis", and "current imbalance".

[0051] 2.2. Proactive Early Warning and Root Cause Analysis: During the test, a potential harmonic exceedance scenario was detected. High-dimensional feature extraction at a certain moment showed that the sum of the 5th and 7th harmonics of the grid current suddenly exceeded the preset allowable range of 5%.

[0052] Multi-level anomaly detection and root cause correlation: The system detected that the harmonic spike occurred precisely at the moment of the simulated brief fluctuations in wind power generation and the slight change in grid frequency. Therefore, the performance library of the digital twin kernel retrieved matches from the feature pattern database and inferred its possible matching patterns: Possible cause A: Insufficient control loop bandwidth leads to slow response to frequency fluctuations, resulting in spectrum leakage.

[0053] Possible cause B: An excessively high value for the integral term in the current loop may cause oscillations when there is a specific load or voltage imbalance.

[0054] Possible cause C: A small offset occurred when the phase-locked loop was tracking the grid frequency, causing PWM synchronization misalignment and introducing harmonics.

[0055] Multi-level suggestion generation: Upon receiving a potential match, the cognitive reasoning module, combining the digital twin kernel's analysis of the simulated component structure (e.g., the coupled resonance effect between the filter inductor and the grid-side equivalent impedance) and the user's current test scenario ("grid adaptability scenario"), provides a comprehensive diagnosis: Harmonic spikes (THD exceeding limits) are detected during grid frequency fluctuations. The main harmonic components are the 5th and 7th harmonics (possibly related to grid impedance coupling).

[0056] (Diagnostic and Simulation Recommendations) Check and enhance the PLL controller bandwidth parameters for grid frequency adaptability. It is recommended to adjust the current parameter ω_n = 150 rad / s to 200 rad / s to optimize tracking.

[0057] (Design Optimization) If the grid impedance characteristics are known, the resonant point of the LCL filter should be rechecked to see if it is too close to the multiple of the switching frequency (or may cause an increase in frequency-dependent harmonics).

[0058] (Safety Verification) Under the premise of ensuring safety boundaries (such as keeping it within ±10% tolerance), adjust the phase-locked loop (PLL) parameters. If necessary, automatically start the "PLL parameter adaptive test" script to verify that the adjusted system will not trigger overvoltage or loop resonance within a range of ±0.75Hz.

[0059] The system then displays a "One-click confirmation to start PLL parameter optimization" button, awaiting engineer approval.

[0060] In the above monitoring process, AI and digital twins act as "on-site diagnosticians and senior engineers," not only quickly "observing, listening, questioning, and diagnosing" to pinpoint the possible causes of the problem, but also providing accurate and verifiable professional improvement suggestions for users.

[0061] Step 3: Deep validation based on adaptive testing and knowledge accumulation; The engineer decided to test the recommended PLL parameters in depth, and entered the following in the test editor: Initiate "PLL parameter adaptive verification" and "Low voltage ride-through (LVRT) standard compliance test".

[0062] Strategic planning and security baseline determination: AI model (L3 planning layer) collaborates with digital twin kernel: Based on the current model characteristics, develop a structured test outline: Baseline testing (standard driven): Normal operating conditions: Stable rated power operation, and investigate the impact of grid frequency variations in the 0-100 Hz range on THDi.

[0063] Fault injection: Simulate three-phase voltages simultaneously dropping to 30%, 50%, and 70% of the grid nominal voltage, lasting for 200ms, 500ms, and 1s respectively. Record recovery response characteristics and monitor whether IGBT thermal loss exceeds limits.

[0064] Safety constraints: Hardware shutdown (hard protection) is triggered when the DC voltage exceeds ±15% of the rated voltage. It is also stipulated that changes in key control loop parameters (such as bandwidth ratio) are not allowed to exceed 20% (engineering tolerance).

[0065] Exploratory Testing (Under Safety Limits): Boundary search: Within a tolerance range of ±20% within the PLL control loop, AI is allowed to fine-tune Kp or Ki using a restricted reinforcement learning (RL) algorithm, while testing the impact of small grid frequency abrupt changes (such as a 0.3 Hz instantaneous offset) on THDi.

[0066] Abnormal data injection: Randomly set several slightly asymmetrical grid voltage disturbances (e.g., only phase B drops by 65%, while the rest remain at 80%) to explore whether the control loop causes excessive internal circulating current under unbalanced voltage conditions.

[0067] At this stage, the L3 AI in the planning layer specifies a macro-level task list for subsequent testing, clarifying priorities and key test points.

[0068] Layered execution framework (safety fence): The system will perform the above tasks for verification based on a three-tier execution strategy: L2 scheduling layer: performs batch execution of standard test scripts (benchmark performance, fault injection), and manages timing and result tracking.

[0069] L1 Restricted Exploration: When the exploration entry "Response under LCL filter resonant frequency offset" is executed, the AI's test module starts the RL agent and gives it a limited whitelisted exploration task space; State space: instantaneous offset of grid frequency (±0.2~±0.7 Hz), harmonic value of filter capacitor voltage (2~5kHz), controller bandwidth adjustment parameters (±2% range).

[0070] Action space: {Disturbance frequency value Δf, controller bandwidth ΔBW, increase L filter inductance ΔL, select to perform "resonant frequency sweep test" or "step transient response test"}. Safety limiter: When the DC voltage change rate exceeds the engineering limit, or the current operating condition enters a known unstable region (such as when the frequency is close to the inherent resonant frequency), the system will forcibly interrupt the current operating condition, determine it as invalid data and impose a "large penalty", and then automatically guide the RL strategy to avoid the danger zone.

[0071] Through limited risk exploration, AI discovered that under a specific combination of grid short-circuit ratio and LCL resonant damping settings, although the system meets the LVRT (Low Voltage Ride-Through) requirements under standard operating conditions, RL found that after a slight imbalance fault in phases A and C (phase shift after a single-phase voltage drop), the inner loop integrator will saturate slightly, resulting in a brief circulating current shift (which does not threaten the hardware but affects the output power quality).

[0072] AI-generated report segment: LVRT Anomaly Analysis (Adaptive Exploratory Discovery): The system performed normally under the standard single-phase -60% voltage drop +100 ms scenario, but the adaptive RL found that there was slight integrator saturation (output error was clamped) for unbalanced faults (phase A -65%, phase deviation 5°).

[0073] Potential consequences: Long-term operation under high harmonics may cause components to heat up too quickly (close to but not exceeding the safety limit), and the harmonic spectrum may expand to the intermediate frequency (THDi needs to be checked).

[0074] Recommended measures (feasibility verified): Slightly increasing the current controller gain Kp_i in the DSP control loop by 2.5% has been verified to improve the saturation effect, while having an impact of less than 0.2% on other operating conditions. 2) Alternatively, adjust the sampling strategy or add an anti-saturation loop to the inductor current under uncontrolled conditions.

[0075] This demonstrates that even after the platform has undergone benchmark performance and functional testing in accordance with standards, it is still able to proactively discover micro-problems based on security constraints, providing additional insights for enhanced security and reliability.

[0076] Results summary, root cause analysis, and knowledge accumulation: The system summarizes the above test steps and generates a report: Test Summary (Partial) The baseline performance meets the national standard: steady-state THDi≤3%, average power factor above 0.99, and has three standard LVR fault ride-through capabilities of 60%, 80%, and 90% (test results are shown in Appendix A.1 to A.3).

[0077] Key issues (discovered through adaptive exploration): The response delay of the PLL controller affects the 5th / 7th harmonics exceeding the standard due to frequency fluctuations.

[0078] For certain non-standard unbalanced drop faults, the controller integrator may experience momentary saturation (slight current offset). Adjustments to the control parameters and reconfirmation of the DSP interrupt update frequency are required.

[0079] The overall system design has sufficient margin and requires no structural modifications. However, in power grids operating for extended periods and containing significant harmonics, it is recommended to include a 3% margin in the hardware thermal design.

[0080] Fault mode: Grid frequency tracking imbalance => Insufficient controller gain; Fault characteristics: A sudden increase in total harmonic distortion (THD) at the 5th and 7th harmonics; observation of internal periodic current offset. Root cause recommendations: Adjust the phase-locked loop gain (kp, ki) to match the grid impedance; slight integral saturation => adjust Kp_i; hardware derating may be required. Verification test results: Increasing Kp_i by 2.5% alleviated the problem (Test ID: 20230330_CaseXYZ). The aforementioned knowledge units are automatically entered into the knowledge base and automatically associated with the "Grid PLLModule" and the control module "Current Inner Loop" in the original design scheme. Now, in all of the company's grid-connected photovoltaic projects, if the system model uses similar LCL filter and PWM parameter configurations, the digital twin will automatically prompt this "non-standard imbalance fault" risk and its mitigation solution.

[0081] Thus, this method completes a closed loop from "intelligent modeling - diagnostic monitoring - deep testing - knowledge accumulation". When designing a similar system, the next engineer can directly obtain knowledge guidance based on this complete process, forming a continuously iterative, evidence-based enterprise-level expert system and collaboration platform. This is precisely the core objective and practical application scenario of this invention.

[0082] On the other hand, embodiments of the present invention also provide an intelligent simulation system applied to the aforementioned adaptive testing method based on large models and digital twins, comprising: The semantic digital twin modeling module is used to generate digital twin models of power electronic systems through natural language interaction, and to verify the compliance and security of modeling instructions to ensure that the model parameters and structure conform to the engineering safety boundaries and physical laws, and to generate standardized digital twins for simulation testing. The real-time monitoring and diagnostic module is used to collect key operational data such as load step response and harmonic spectrum, match characteristic patterns to locate anomalies, and generate comprehensive diagnostic suggestions based on system structure and scenario. The knowledge management and closed-loop iteration module is used to summarize test data to generate special reports, extract structured knowledge units and enter them into the enterprise knowledge base, and associate them with corresponding system modules. When modeling similar systems in the future, it will automatically prompt the corresponding fault risks and mitigation solutions, realizing knowledge reuse and a closed-loop process.

[0083] The digital twin model is divided into an L3 planning layer, an L2 scheduling layer, and an L1 restricted exploration layer. The L3 layer formulates the test outline and safety baseline, the L2 layer executes standard test scripts in batches and tracks the results, and the L1 layer uses a reinforcement learning algorithm with safety barriers for restricted exploration.

[0084] The real-time monitoring and diagnosis module includes a real-time feature extraction unit, an anomaly detection unit, and a root cause analysis unit; wherein... The feature extraction unit is responsible for filtering and extracting key features that are strongly correlated with the system status from multi-source heterogeneous business operation data, system logs, monitoring indicators and other data, and completing data noise reduction and dimensionality compression to provide an accurate and efficient analysis basis for subsequent anomaly detection.

[0085] The anomaly detection unit, based on the key features output by the feature extraction unit, uses algorithms such as rule matching, statistical analysis, and machine learning to identify abnormal states that deviate from expectations by comparing them with the normal baseline. It outputs information such as the type, occurrence time, and scope of impact of the abnormal event in real time or near real time, enabling rapid perception of system anomalies.

[0086] The root cause analysis unit, in response to the abnormal events identified by the anomaly detection unit, locates the root cause of the anomaly by associating feature data, tracing the call chain, and analyzing causal relationships, and outputs the root cause location results and related evidence, providing a direct basis for problem repair.

[0087] The system also includes: The human-computer interaction module provides an interactive interface for one-click confirmation of execution parameter optimization and script verification, supporting engineer approval and manual intervention.

[0088] In summary, the embodiments of the present invention provide an adaptive testing method and intelligent simulation system based on large models and digital twins. By integrating AI large models, semantic digital twins, and a multi-layered collaborative framework of safety constraints, it completely changes the working mode of power electronics simulation, realizing a highly automated and intelligent closed loop of design, monitoring, testing, learning, and knowledge management, enabling engineers to gain in-depth and forward-looking engineering insights while ensuring safety boundaries.

[0089] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0090] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0091] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0092] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive testing method based on large models and digital twins, characterized in that, Includes the following steps: Digital twin models of power electronic systems are generated through natural language interaction, and the modeling instructions are verified for compliance and security to ensure that the model parameters and structure conform to engineering safety boundaries and physical laws, thereby generating standardized digital twins for simulation testing. Collect key operating data of power electronic systems, and when abnormal operating conditions are detected, match potential root causes through the feature pattern database of the digital twin model, and generate comprehensive diagnostic suggestions by combining system structure and test scenario; The test data is aggregated to generate a test report. Structured knowledge units containing fault mode labels, fault feature sets, root cause suggestion lists, and verification test conclusions are extracted, entered into the enterprise knowledge base, and associated with the corresponding system modules. When modeling similar systems in the future, the corresponding fault risks and mitigation solutions will be automatically prompted.

2. The adaptive testing method based on large models and digital twins according to claim 1, characterized in that, The compliance and security verification of the modeling instructions includes: The digital twin model performs semantic parsing of natural language modeling instructions, verifies whether the model parameters meet the physical constraints and engineering safety tolerances of the power electronic system, and automatically provides correction suggestions and remodels if there are parameter out-of-bounds or structural conflicts, until a digital twin that meets the requirements is generated.

3. The adaptive testing method based on large models and digital twins according to claim 1, characterized in that, The key operational data collected from the power electronic system include: load step response, harmonic spectrum analysis, and current imbalance. The process of matching potential root causes using a digital twin model's feature pattern database, combined with system structure and test scenarios, to generate comprehensive diagnostic suggestions includes: By retrieving matching feature patterns from the performance library of the digital twin model, the root cause analysis covers potential reasons such as insufficient control loop bandwidth, excessively high values ​​of the current loop integral term, and phase-locked loop frequency offset. Combined with the coupling characteristics of the filter components and the test scenario, precise improvement suggestions are given.

4. The adaptive testing method based on large models and digital twins according to claim 1, characterized in that, The digital twin model includes: The L3 planning layer is used to develop a structured test outline in conjunction with the digital twin kernel, clarifying the tasks, priorities, and security baselines for baseline testing and exploratory testing. The L2 scheduling layer is used to execute standard test scripts in batches and track the results. The L1 restricted exploration layer uses a reinforcement learning algorithm with safety barriers to explore within a limited parameter space. When a safety threshold is triggered, the exploration is forcibly interrupted and the strategy is guided to avoid the danger zone, thus uncovering potential problems under non-standard working conditions.

5. The adaptive testing method based on large models and digital twins according to claim 4, characterized in that, The baseline test includes routine operating condition test and fault injection test; the routine operating condition test examines the impact of grid frequency changes on total harmonic distortion rate; the fault injection test simulates three-phase voltage drops of different degrees and durations, records recovery response characteristics and monitors device thermal loss.

6. The adaptive testing method based on large models and digital twins according to claim 4, characterized in that, The safety baseline includes hard protection triggered when the DC voltage exceeds ±15% of the rated voltage, and control loop key parameter changes not exceeding the 20% engineering tolerance limit.

7. The adaptive testing method based on large models and digital twins according to claim 4, characterized in that, The exploration test includes boundary search and anomaly data injection. Boundary search involves fine-tuning the controller parameters through reinforcement learning within a 20% tolerance range for control loop parameters, testing the impact of small changes in grid frequency on total harmonic distortion. Anomaly data injection involves setting up asymmetrical grid voltage disturbances to explore the changes in the internal circulating current of the control loop under unbalanced voltage.

8. The adaptive testing method based on large models and digital twins according to claim 4, characterized in that, The reinforcement learning algorithm with safety guardrails includes: defining a state space containing grid frequency offset, filter capacitor voltage harmonic values, and controller bandwidth parameters; setting a safety limiter to forcibly interrupt the current test and penalize the reinforcement learning strategy when the DC voltage change rate exceeds the limit or the operating condition enters an unstable region, thereby guiding the strategy to avoid dangerous operating conditions.

9. The adaptive testing method based on large models and digital twins according to claim 1, characterized in that, The method further includes: For anomaly diagnosis and testing suggestions, an interactive button is provided to confirm the start of parameter optimization or verification scripts with one click, supporting engineers to approve and execute as needed, and realizing the collaboration between intelligent decision-making and manual intervention.

10. An intelligent simulation system applied to the adaptive testing method based on large models and digital twins as described in any one of claims 1 to 9, characterized in that, The system includes: The semantic digital twin modeling module is used to generate digital twin models of power electronic systems through natural language interaction, and to verify the compliance and security of modeling instructions to ensure that the model parameters and structure conform to the engineering safety boundaries and physical laws, and to generate standardized digital twins for simulation testing. The real-time monitoring and diagnosis module includes a real-time feature extraction unit, an anomaly detection unit, and a root cause analysis unit. It is used to collect key operating data such as load step response and harmonic spectrum, match feature patterns to locate anomalies, and generate comprehensive diagnostic suggestions based on system structure and scenario. The knowledge management and closed-loop iteration module is used to summarize test data to generate special reports, extract structured knowledge units and enter them into the enterprise knowledge base, and associate them with corresponding system modules. When modeling similar systems in the future, it will automatically prompt the corresponding fault risks and mitigation solutions, realizing knowledge reuse and a closed-loop process.

Images