A method and system for setting control parameters of an inverter

Abstract

The application provides a control parameter setting method and system of an inverter, applied to the technical field of inverters, and obtains a waveform by running a DC-AC simulation model; generates an atlas image by using the waveform; identifies a response type of the DC-AC simulation model according to the atlas image by using a first intelligent model; identifies a tail trend type of the DC-AC simulation model according to the atlas image by using a second intelligent model; if the DC-AC simulation model does not satisfy a termination condition, determines a Kp recommended adjustment ratio and a Ki recommended adjustment ratio according to the response type and the tail trend type; performs primary parameter correction on Kp and Ki of the DC-AC simulation model by using the Kp recommended adjustment ratio and the Ki recommended adjustment ratio; determines fine tuning parameters according to current Kp, current Ki and historical simulation information of the DC-AC simulation model after primary parameter correction; and performs fine tuning on Kp and Ki of the DC-AC simulation model after primary parameter correction by using the fine tuning parameters.

Description

Technical Field

[0001] This invention relates to the field of inverter technology, and more specifically, to a method and system for tuning control parameters of an inverter. Background Technology

[0002] With the continuous development of power electronics technology, DC-AC inverters, as an important power conversion device, are widely used in photovoltaic power generation, wind power generation, electric vehicles, and other fields. The main function of an inverter is to convert direct current (DC) power into alternating current (AC) power to meet the needs of different loads.

[0003] In inverter control systems, PI controllers (proportional-integral controllers) are commonly used to adjust control parameters such as the waveforms of output voltage and current to ensure the stability and efficiency of the control system under various operating environments. However, tuning the control parameters of PI controllers has always been a challenge in the field of power electronics, especially in complex environments with multiple operating conditions. How to adjust the Kp (proportional coefficient) and Ki (integral coefficient) of the PI controller in real time and accurately to maintain the optimal performance of the system remains a technical problem that urgently needs to be solved.

[0004] Existing PI parameter tuning methods mainly include empirical methods, numerical optimization methods, and model comparison methods. However, these methods suffer from low adjustment accuracy and robustness, especially when dynamic anomalies exist, making it difficult to respond quickly and compromising the stability and response speed of the control system. Furthermore, existing methods struggle to achieve end-to-end closed-loop automated control, often requiring manual intervention and frequent parameter adjustments. This not only reduces the system's automation level but also increases operational complexity. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for tuning control parameters of an inverter to achieve automated parameter tuning, thereby improving the accuracy and robustness of parameter tuning, and ensuring the stability and response speed of the system.

[0006] The first aspect of this application provides a method for tuning control parameters of an inverter, the method comprising:

[0007] Run the DC-AC simulation model to obtain the corresponding waveforms; wherein the waveforms are voltage waveforms or current waveforms.

[0008] A standardized spectral image is generated using the waveform;

[0009] The response type of the DC-AC simulation model is identified based on the graph image using a first intelligent model; wherein, the first intelligent model is obtained by training a neural network intelligent model using a first dataset;

[0010] The tail trend type of the DC-AC simulation model is identified based on the graph image using a second intelligent model; wherein, the second intelligent model is obtained by training the neural network intelligent model using a second dataset;

[0011] If it is determined that the DC-AC simulation model does not meet the termination conditions, the recommended adjustment ratios for Kp and Ki are determined based on the response type and the tail trend type.

[0012] The initial parameter correction of Kp and Ki in the DC-AC simulation model is performed using the proposed adjustment ratios for Kp and Ki.

[0013] Based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, determine the fine-tuning parameters;

[0014] The process involves fine-tuning Kp and Ki of the DC-AC simulation model after the primary parameters have been corrected using the fine-tuning parameters, and then returning to run the DC-AC simulation model to obtain the corresponding waveform.

[0015] Optionally, the DC-AC simulation model is the main circuit topology of a three-phase two-level inverter. The DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

[0016] Optionally, the first intelligent model is used to identify the response type of the DC-AC simulation model based on the spectral image, including:

[0017] The spectrum image is input into the first intelligent model, which extracts the peak features and oscillation time features of the waveform from the spectrum image, and uses the peak features and oscillation time features to identify the response type of the DC-AC simulation model.

[0018] Optionally, the second intelligent model is used to identify the tail trend type of the DC-AC simulation model based on the spectral image, including:

[0019] The spectrum image is input into the second intelligent model, which extracts waveform stationarity features from the stationary portion of the spectrum image and uses these features to identify the tail trend type of the DC-AC simulation model.

[0020] Optionally, the method further includes:

[0021] Determine whether the response type is a specified response type and whether the tail trend type is a specified tail trend type;

[0022] If the response type is a specified response type and the tail trend type is a specified tail trend type, the DC-AC simulation model is determined to meet the termination condition, and the control parameter tuning process of the DC-AC simulation model is terminated.

[0023] If the response type is not the specified response type, and / or the tail trend type is not the specified tail trend type, the DC-AC simulation model is determined not to meet the termination condition.

[0024] Optionally, determining the Kp suggested adjustment ratio and the Ki suggested adjustment ratio based on the response type and the tail trend type includes:

[0025] Determine the response type label to which the response type belongs, and determine the Kp suggestion adjustment ratio that matches the response type label from a set of preset Kp suggestion adjustment ratios;

[0026] Determine the tail trend type label to which the tail trend type belongs, and determine the Ki suggested adjustment ratio that matches the tail trend type label from the preset Ki suggested adjustment ratios.

[0027] Optionally, generating a standardized spectral image using the waveform includes:

[0028] The corresponding imaging frame is configured according to the model information of the DC-AC simulation model and the waveform information of the waveform; wherein, the leftmost side of the imaging frame is based on the initial response point of the waveform, the central axis of the imaging frame is the target given value, the height of the imaging frame is configured according to the current working condition of the DC-AC simulation model, and the width of the imaging frame is configured according to the maximum response time of the waveform.

[0029] An initial spectral image is acquired from the waveform according to the imaging frame, and the initial spectral image is standardized to obtain a standardized spectral image.

[0030] Optionally, fine-tuning parameters are determined based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, including:

[0031] Obtain the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction. The historical simulation information includes the current tuning round, target number, historical response type label and historical tail trend type label of the DC-AC simulation model in the previous tuning. The target number is the number of tuning rounds in which the response type label and tail trend type label remain unchanged for consecutive cycles within a preset number of rounds.

[0032] The current state vector and reward are determined by the policy network based on the current Ki, the current Kp, the historical response type label, the historical tail trend type label, the current tuning round, and the target number.

[0033] The policy network outputs a corresponding target action based on the state vector and reward, and determines fine-tuning parameters based on the target action. The target action includes adjusting the target number of steps up / down by Kp / Ki, and the fine-tuning parameters include adding / subtracting the target number of steps by Kp / Ki.

[0034] A second aspect of this application provides a control parameter tuning system for an inverter, the system comprising:

[0035] The running unit is used to run the DC-AC simulation model to obtain the corresponding waveforms; wherein the waveforms are voltage waveforms or current waveforms.

[0036] The spectral image generation unit is used to generate a standardized spectral image using the waveform;

[0037] The response type identification unit is used to identify the response type of the DC-AC simulation model based on the graph image using a first intelligent model; wherein, the first intelligent model is obtained by training a neural network intelligent model using a first dataset;

[0038] The tail trend type identification unit is used to identify the tail trend type of the DC-AC simulation model based on the graph image using a second intelligent model; wherein, the second intelligent model is obtained by training the neural network intelligent model using a second dataset;

[0039] The preliminary parameter correction unit is used to determine the suggested adjustment ratios of Kp and Ki based on the response type and the tail trend type if it is determined that the DC-AC simulation model does not meet the termination condition, and to perform preliminary parameter correction on Kp and Ki of the DC-AC simulation model using the suggested adjustment ratios of Kp and Ki.

[0040] The fine-tuning unit is used to determine the fine-tuning parameters based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, and to fine-tune Kp and Ki of the DC-AC simulation model after primary parameter correction using the fine-tuning parameters before returning to the execution unit.

[0041] Optionally, the DC-AC simulation model is the main circuit topology of a three-phase two-level inverter. The DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

[0042] This application provides a method and system for tuning control parameters of an inverter. The method involves running a DC-AC simulation model to obtain corresponding waveforms, where the waveforms are voltage or current waveforms. A standardized spectral image is generated using the waveforms. A first intelligent model identifies the response type of the DC-AC simulation model based on the spectral image. The first intelligent model is obtained by training a neural network intelligent model using a first dataset. A second intelligent model identifies the tail trend type of the DC-AC simulation model based on the spectral image. The second intelligent model is obtained by training the neural network intelligent model using a second dataset. If the DC-AC simulation model is determined... If the termination condition is not met, the suggested adjustment ratios for Kp and Ki are determined based on the response type and the tail trend type. The suggested adjustment ratios for Kp and Ki in the DC-AC simulation model are then used to perform primary parameter correction. Based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, fine-tuning parameters are determined. These fine-tuning parameters are then used to fine-tune the Kp and Ki of the DC-AC simulation model after primary parameter correction, and the process returns to running the DC-AC simulation model to obtain the corresponding waveform. This process continues until the DC-AC simulation model meets the termination condition, thereby achieving end-to-end closed-loop automated control. Therefore, this application utilizes pre-trained first and second intelligent models to identify the response type and tail trend type of the DC-AC simulation model, enabling primary parameter correction based on these characteristics. This transforms the control performance, which traditionally requires precise quantification, into a discrete label representation. Furthermore, after primary parameter correction, corresponding fine-tuning parameters are determined to fine-tune the PI control parameters of the DC-AC simulation model after primary parameter correction. This improves the adjustment accuracy and robustness of the PI control parameters. Moreover, the entire process requires no manual intervention or frequent parameter adjustments, which not only enhances the system's automation level but also reduces operational complexity. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a method for tuning control parameters of an inverter provided in an embodiment of this application;

[0045] Figure 2 A structural diagram of the main circuit topology of a three-phase two-level inverter provided in this application embodiment;

[0046] Figure 3 This application provides a control block diagram for introducing a feedforward compensation term to decouple the parameters of a dual closed-loop controller.

[0047] Figure 4 An example diagram of an imaging frame provided in an embodiment of this application;

[0048] Figure 5 A structural diagram of a neural network intelligent model provided in an embodiment of this application;

[0049] Figure 6 A structural diagram of a C2f layer provided in an embodiment of this application;

[0050] Figure 7 A three-phase voltage waveform diagram provided in an embodiment of this application;

[0051] Figure 8 A three-phase current waveform diagram provided in an embodiment of this application;

[0052] Figure 9 A voltage total harmonic distortion (THD) diagram provided for embodiments of this application;

[0053] Figure 10 A total harmonic distortion (THD) diagram of current provided in an embodiment of this application;

[0054] Figure 11 This is a schematic diagram of the control parameter tuning system for an inverter provided in an embodiment of this application. Detailed Implementation

[0055] 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.

[0056] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0057] As can be seen from the background above, existing PI parameter tuning methods mainly include empirical methods, numerical optimization methods, and model comparison methods. Among these, empirical methods, such as the Ziegler-Nichols method, while providing preliminary tuning results under certain conditions, require multiple manual trials, resulting in low efficiency and difficulty adapting to complex operating conditions, especially in dynamically changing environments. Numerical optimization methods, such as genetic algorithms and particle swarm optimization, can effectively optimize PI control parameters, but these methods are prone to getting trapped in local optima and have high computational costs, failing to meet real-time requirements. Model comparison methods, such as model reference adaptive control, require accurate system models; however, in practical applications, the inverter's operating environment is often complex and dynamically changing, making it difficult to establish accurate mathematical models and thus hindering precise tuning of PI control parameters.

[0058] Therefore, it is evident that existing parameter tuning methods suffer from low adjustment accuracy and robustness, especially when dynamic anomalies exist, making it difficult to respond quickly and compromising the stability and response speed of the control system. Furthermore, existing methods struggle to achieve end-to-end closed-loop automated control, often requiring manual intervention and frequent parameter adjustments. This not only reduces the system's automation level but also increases operational complexity.

[0059] Therefore, this application provides a method and system for tuning the control parameters of an inverter. By utilizing a pre-trained first and second intelligent model to identify the response type and tail trend type of the DC-AC simulation model, the primary parameter correction can be completed based on the response type and tail trend type, transforming the control performance that traditionally requires precise quantization into a discrete label representation. Furthermore, after completing the primary parameter correction, the corresponding fine-tuning parameters are further determined to fine-tune the PI control parameters of the DC-AC simulation model after the primary parameter correction, thereby improving the adjustment accuracy and robustness of the PI control parameters. Moreover, the entire process does not require manual intervention or frequent parameter adjustments, which not only improves the automation level of the system but also reduces the operational complexity.

[0060] See Figure 1 The diagram shows a flowchart of a control parameter tuning method for an inverter according to an embodiment of this application. The control parameter tuning method for the inverter specifically includes the following steps:

[0061] S101: Run the DC-AC simulation model to obtain the corresponding waveforms; where the waveforms are voltage waveforms or current waveforms.

[0062] In this embodiment, a corresponding DC-AC simulation model can be pre-built in the Simulink simulation environment. The DC-AC simulation model is the main circuit topology of a three-phase two-level inverter, such as... Figure 2 As shown; and the DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

[0063] In practical applications, while proportional-integral (PI) controllers exhibit excellent steady-state and dynamic regulation performance for DC quantities, their control characteristics inherently mismatch with sinusoidal AC quantities. Therefore, the DC-AC simulation model provided in this application employs a Park transform-based synchronous rotating coordinate system (dq) decoupling control strategy (preset decoupling strategy) to convert three-phase AC quantities into DC quantities in the dq coordinate system, thus addressing this issue. Furthermore, the DC-AC simulation model also incorporates a dual-loop controller with a cascaded control structure of "voltage outer loop—current inner loop." To effectively eliminate the coupling effect between the dq axes, this application introduces a feedforward compensation term to decouple the parameters of the dual-loop controller. The control block diagram for parameter decoupling of the dual-loop controller with the feedforward compensation term is shown below. Figure 3 As shown.

[0064] It should be noted that, from Figure 3From this, we can understand that the parameter K of the feedforward compensation can be specifically interpreted as follows: the feedforward compensation terms for the outer voltage loop are UdωC and UqωC, and the feedforward compensation terms for the inner current loop are IdωL and IqωL. Parameter decoupling is achieved through feedforward compensation to realize independent control of the dq-axis voltage and current. Where Udref and Uqref are the voltage command values ​​for the dq-axis, Uod and Uoq are the actual dq-axis voltages after passing through the LC filter in the DC-AC simulation model, Iod and Ioq are the actual dq-axis currents before passing through the LC filter in the DC-AC simulation model, and Id and Iq are the actual dq-axis currents before the LC filter in the DC-AC simulation model.

[0065] In this embodiment of the application, before building and running the DC-AC simulation model in the Simulink simulation environment, the PI control parameters of the DC-AC simulation model can be initialized first; wherein, the initialization of the PI control parameters of the DC-AC simulation model includes the initialization of the PI control parameters of the inner current loop and the initialization of the PI control parameters of the outer voltage loop.

[0066] It should be noted that the initialization process of the PI control parameters of the current inner loop is as shown in formulas (1) to (9). The specific process can be as follows: In the dq coordinate system, after complete decoupling by the feedforward compensation term, the controlled object of the current inner loop can be simplified to obtain the simplified object as shown in formula (1); After simplification, the corresponding current inner loop bandwidth can be further configured (as shown in formula (2)), the corresponding angular frequency can be configured (as shown in formula (3)), the transfer function of the current inner loop PI controller can be configured (as shown in formula (4)), the poles of the controlled object and the zeros of the controller can be canceled (the cancellation method is shown in formula (5)), the open-loop transfer function of the current inner loop IP controller can be configured (as shown in formula (6)), the gain can be set to 1 at the corresponding bandwidth (the gain configuration method is shown in formula (7)), and the initial proportional coefficient of the current inner loop can be configured. and integral coefficient (As shown in formulas (8) and (9)).

[0067] (1)

[0068] in, This represents the response characteristics of the circuit in the complex frequency domain (s-domain), i.e., the transfer function of the controlled object, where s is the complex frequency variable, L represents the inductance value in the circuit, and R represents the resistance value in the circuit.

[0069] (2)

[0070] in, The 0dB crossover frequency of the inner current loop, which is the frequency at which the gain of the inner current loop drops to 1 (0dB), determines the bandwidth and response speed of the inner current loop; f sw This indicates the switching frequency of the switching transistor.

[0071] (3)

[0072] in, ω is the angular frequency.

[0073] (4)

[0074] in, This is the transfer function of the current inner-loop PI controller. This is the proportionality coefficient of the inner current loop. is the integral coefficient of the inner current loop. This indicates the zero-point angular frequency of the PI controller.

[0075] (5)

[0076] In this circuit, L represents the inductance value and R represents the resistance value. This is the proportionality coefficient of the inner current loop. is the integral coefficient of the inner current loop. This indicates the zero-point angular frequency of the PI controller.

[0077] (6)

[0078] in, This represents the open-loop transfer function of the current inner-loop PI controller. This is the transfer function of the current inner-loop PI controller. Pass functions to the controlled object.

[0079] (7)

[0080] in, Let j be the complex gain of the open-loop transfer function of the inner current loop at frequency ω, where j is the imaginary unit to avoid confusion with the current symbol i. The high-frequency equivalent inductance determines the proportional gain and mainly affects the bandwidth.

[0081] (8)

[0082] in, This is the proportional coefficient of the inner current loop.

[0083] (9)

[0084] in, is the integral coefficient of the inner current loop. The high-frequency equivalent resistance determines the integral gain and mainly affects the speed of steady-state error elimination.

[0085] It should be noted that the initialization process of the PI control parameters of the voltage outer loop is as shown in formulas (10)-(18). The specific process can be as follows: the current inner loop is regarded as an ideal gain of 1 to simplify the controlled object of the voltage outer loop, and the simplified controlled object is shown in formula (10); after simplification, the corresponding voltage outer loop bandwidth can be further configured (as shown in formula (11)), the corresponding angular frequency can be configured (as shown in formula (12)), the transfer function of the voltage outer loop PI controller can be configured (as shown in formula (13)), and the integrator poles and controller zeros can be canceled (i.e., the integrator poles and controller zeros are canceled). Set to 0), retain the integral element in actual engineering to eliminate steady-state error (as shown in formula (14)), configure the open-loop transfer function of the voltage outer loop IP controller (as shown in formula (15)), set the gain to 1 at the corresponding bandwidth (the gain configuration method is shown in formula (16)), and configure the initial proportional coefficient of the voltage outer loop. and integral coefficient (As shown in formulas (17) and (18)).

[0086] (10)

[0087] in, It represents the voltage-to-current transfer function, where s is a complex frequency variable and C represents the capacitance value.

[0088] (11)

[0089] in, Indicates the voltage outer loop crossover frequency. This indicates the 0dB crossover frequency of the inner current loop.

[0090] (12)

[0091] in, This indicates the angular frequency at which the outer voltage loop crosses.

[0092] (13)

[0093] in, This represents the transfer function of the voltage outer-loop PI controller. This is the initial proportional gain of the outer voltage loop. is the initial integral coefficient of the outer voltage loop, and s is the complex frequency variable.

[0094] (14)

[0095] in, The initial integral coefficient of the outer voltage loop is denoted as . This is the initial proportional gain of the outer voltage loop. This indicates the angular frequency at which the outer voltage loop crosses.

[0096] (15)

[0097] in, This represents the open-loop transfer function of the voltage outer loop. This represents the transfer function of the voltage outer-loop PI controller. This represents the voltage-to-current transfer function. is the initial proportionality coefficient of the outer voltage loop, s is the complex frequency variable, and C represents the capacitance value.

[0098] (16)

[0099] in, Let be the complex gain of the transfer function at frequency ω. Angular frequency The amplitude of the "capacitive admittance" at the location (unit: S). is the initial proportionality coefficient of the outer voltage loop, and j is the imaginary unit.

[0100] (17)

[0101] in, The initial proportional gain of the outer voltage loop is given by C, where C represents the capacitance value. Indicates the crossing angular frequency of the outer voltage loop. This indicates the voltage outer loop crossover frequency.

[0102] (18)

[0103] in, This is the initial proportional gain of the outer voltage loop. The initial integral coefficient of the outer voltage loop.

[0104] During the specific execution of step S101, the DC-AC simulation model in MATLAB can be run to make the DC-AC simulation model output the corresponding waveform (current waveform or voltage waveform).

[0105] It should be noted that if this is the first time running the DC-AC simulation model, the DC-AC simulation model being run will be the initialized DC-AC simulation model; if this is not the first time running the DC-AC simulation model, the DC-AC simulation model being run will be the DC-AC simulation model after the control parameters were tuned last time.

[0106] S102: Generate standardized spectral images using waveforms.

[0107] In the specific execution step S102, after acquiring the waveform output by the DC-AC simulation model, an initial spectrum image can be extracted from the waveform based on the model information of the currently running DC-AC simulation model and the waveform information of the currently acquired waveform. The initial spectrum image is then standardized to obtain a standardized spectrum image.

[0108] It should be noted that by using the acquired waveforms to generate standardized spectral images, the required DC data can be converted into standardized spectral images that can be input into the pre-trained first and second intelligent models.

[0109] Optionally, the process of generating a standardized spectral image from the waveform can be as follows: configure the corresponding imaging frame according to the model information of the DC-AC simulation model and the waveform information; wherein, the leftmost side of the imaging frame is based on the initial response point of the waveform, the central axis of the imaging frame is the target given value, the height of the imaging frame is configured according to the current working condition of the DC-AC simulation model, and the width of the imaging frame is configured according to the maximum response time of the waveform; the initial spectral image is acquired from the waveform according to the imaging frame, and the initial spectral image is standardized to obtain a standardized spectral image.

[0110] It should be noted that the waveform information can include the initial response point and maximum response time of the waveform; the model information of the DC-AC simulation model can include the current operating condition of the DC-AC simulation model; the target setpoint can be a pre-set given voltage or given current.

[0111] In practical applications, after acquiring the corresponding waveform, the corresponding imaging frame is configured based on the model information of the DC-AC simulation model and the waveform information, such as... Figure 4 As shown; from Figure 4 As can be seen, the leftmost part of the imaging frame uses the initial response point of the waveform as a reference, and the pre-set given voltage is used as the central axis of the imaging graph. The height of the imaging frame is 2.4 times the maximum allowable overshoot (σ) of the current operating condition of the DC-AC simulation model (1.2 times the maximum overshoot below the central axis is used as the height of the imaging frame, i.e., the height of the imaging frame = 2.4σ). The width of the imaging frame is the maximum response time t. s Twice as much.

[0112] In some embodiments, size information that conforms to the standardization of the first and second intelligent models can be pre-configured so that after the initial spectral image is extracted from the waveform using the imaging frame, the pre-configured standardized size information can be used to standardize the initial spectral image to obtain an input spectral image that conforms to the first and second intelligent models.

[0113] It should be noted that the standardized size information enables the first / second intelligent model to efficiently and accurately complete the corresponding recognition tasks based on the input spectral images, regardless of the DC-AC simulation model under any working condition. Furthermore, to improve the model's discriminative power, simple sharpening enhancement processing can be applied to the spectral images.

[0114] S103: Use the first intelligent model to identify the response type of the DC-AC simulation model based on the graph image; wherein, the first intelligent model is obtained by training the neural network intelligent model using the first dataset.

[0115] In the specific execution step S103, a first dataset can be pre-collected, which includes multiple historical spectrogram images of the DC-AC simulation model and their corresponding target response types. The collected first dataset is used to train the neural network intelligent model to obtain a first intelligent model. After obtaining the spectrogram images, the spectrogram images are input into the first intelligent model so that the first intelligent model can extract the first key dynamic features from the spectrogram images and use the first key dynamic features to identify the response type.

[0116] It should be noted that the first key dynamic feature may include the peak feature of the waveform (the peak feature of the first peak) and the oscillation time feature.

[0117] It should also be noted that the response type can be type A, type B, type C, type D, etc. Type A can indicate that the DC-AC simulation model may have obvious overshoot; type B indicates that the DC-AC simulation model may have slow response / undershoot; type C indicates that the DC-AC simulation model may have continuous oscillation; and type D indicates that the DC-AC simulation model is currently in an ideal response.

[0118] In this embodiment, a first dataset can be pre-collected, which includes multiple historical spectrum images of the DC-AC simulation model and their corresponding target response types. The historical spectrum images are input into the neural network intelligent model, which extracts the historical peak features and historical oscillation time features of the corresponding historical waveforms from the historical spectrum images. Based on the historical peak features and historical oscillation time features, the corresponding response types are identified. A corresponding loss function is constructed based on the identified response types and the corresponding target response types. The parameters of the neural network intelligent model are adjusted using the loss function until the neural network intelligent model converges, thus obtaining the first intelligent model.

[0119] It should be noted that the first intelligent model can be used to initially determine the matching between the PI control parameters and the corresponding behaviors of the DC-AC simulation model.

[0120] It should also be noted that the structure of the neural network intelligent model is as follows: Figure 5 As shown, from Figure 5 From this, we can understand that the neural network intelligent model has 5 Conv layers (convolutional layers), 4 C2f layers (cross-stage fusion layers), and 1 Classify layer (classifier layer). The structure of each C2f layer is as follows: Figure 6 As shown; correspondingly, the structure of the first intelligent model obtained by training the neural network intelligent model using the first dataset is the same as the structure of the neural network intelligent model.

[0121] Optionally, the process of using the first intelligent model to identify the response type of the DC-AC simulation model based on the spectrum image can be as follows: input the spectrum image into the first intelligent model, so that the first intelligent model extracts the peak features and oscillation time features of the waveform from the spectrum image, and uses the peak features and oscillation time features to identify the response type of the DC-AC simulation model.

[0122] In this embodiment, since the structure of the first intelligent model is the same as the structure of the neural network intelligent model, it can be referred to... Figure 5The obtained spectrum image is input into the first intelligent model; the first intelligent model extracts the first key dynamic features (peak features and oscillation time features) of the waveform from the spectrum image; the first key dynamic features are processed by convolution through two convolutional layers (Conv_1 and Conv_2); the result of the convolution is processed by the C2f_1 layer, and then convolved by Conv_3; the result of the convolution is processed by the C2f_2 layer, and then convolved by Conv_4; the result of the convolution is processed by the C2f_3 layer, and then convolved by Conv_5; the result of the convolution is processed by the C2f_4 layer, and finally processed by the Classify layer to obtain the response type of the DC-AC simulation model.

[0123] In some embodiments, the C2f layer processing includes the following stages: See Figure 6 First, the result of convolutional processing of the first key dynamic feature through two convolutional layers is then subjected to a combined operation consisting of a 1×1 convolution (Conv2d (2D convolutional layer)), batch normalization (BatchNorm2d (2D regularization layer)), and SiLU activation function to uniformly set the number of channels to C. Second, the result obtained after the combined operation is divided into two parts proportionally (1:1) in the channel dimension by splitting the channels, with each part having C / 2 channels. One part remains unchanged for subsequent skip connections, while the other part is sequentially input into n Bottleneck structures (a Bottleneck residual module includes a 1×1 convolutional module (composed of a convolutional layer, BN, and SiLU activation)) to compress the number of channels to C / 2×e (where e represents the channel expansion factor, defaulting to 0.5), and then passed through a 3×3 convolution. The layer restores it to C / 2, and finally the input and output are superimposed by element-wise addition to form a residual connection. Concat (feature concatenation) is used to concatenate the outputs of all Bottleneck modules with the initially retained shorted branches in the channel dimension to generate a new feature tensor with C×(1+n / 2) channels. Then, a 1×1 convolutional layer (containing convolution, BN and SiLU) is used to compress the number of channels of the new feature tensor back to C. Finally, the result is added to the original 1×1 convolution output to obtain the final output, keeping the channel dimension C.

[0124] In some embodiments, the calculation process of the Classify layer can be as follows: First, extract C channels from the end of the channel dimension of the input feature tensor (the result of the C2f_4 layer output) to form a sub-tensor 1 with dimension (B,A,C); feed this sub-tensor 1 into a 1×1 convolutional layer (containing a C×C convolutional kernel) without bias term and normalization to perform the corresponding linear transformation of the channel dimension, and the output shape is still a sub-tensor 2 with the shape of (B,A,C). Next, apply the Sigmoid function (in the form of 1 / (1+e^(-x))) to each value in this sub-tensor 2 to map it to the interval [0,1], thereby obtaining a probability tensor with unchanged shape. Finally, this output will be directly used as the classification result (i.e., to obtain the corresponding response type) for subsequent processing without any further operations.

[0125] Therefore, this application obtains a first intelligent model by pre-training a neural network intelligent model using a first dataset containing historical spectral images and their target response types, so as to accurately identify the current response type of the DC-AC simulation model based on the currently obtained spectral images using the first intelligent model.

[0126] S104: The tail trend type of the DC-AC simulation model is identified based on the graph image using the second intelligent model; the second intelligent model is obtained by training the neural network intelligent model using the enemy dataset.

[0127] In the specific execution step S104, a second dataset can be pre-collected, which includes multiple historical spectrogram images of the DC-AC simulation model and their corresponding target waveform stationary features. The collected second dataset is used to train the neural network intelligent model to obtain a second intelligent model. After obtaining the spectrogram images, the spectrogram images are input into the second intelligent model, so that the second intelligent model can extract the second key dynamic features from the spectrogram images and use the second key dynamic features to identify the corresponding tail trend type.

[0128] It should be noted that the second key dynamic feature can be the waveform stationarity feature.

[0129] It should also be noted that the tail trend type can be P type, Pd type, or Ok type, etc. Among them, P type can indicate that the DC-AC simulation model may have a large steady-state deviation; Pd type indicates that the DC-AC simulation model may have severe integral oscillation; Ok type indicates that the DC-AC simulation model is currently in stable convergence.

[0130] In this embodiment, a second dataset can be pre-collected, which includes multiple historical graph images of the DC-AC simulation model and their corresponding target tail trend types. The historical graph images are input into the neural network intelligent model, which extracts the corresponding historical waveform stationary features from the historical graph images and identifies the corresponding tail trend types based on the historical waveform stationary features. A corresponding loss function is then constructed based on the identified tail trend type and the corresponding target tail trend type. The parameters of the neural network intelligent model are adjusted using the loss function until the neural network intelligent model converges, thus obtaining the second intelligent model.

[0131] It should be noted that the second intelligent model can evaluate the steady-state performance benchmark brought about by the integral element.

[0132] It should also be noted that the structure of the neural network intelligent model is as follows: Figure 5 As shown, from Figure 5 From this, we can understand that the neural network intelligent model has 5 Conv layers, 4 C2f layers, and 1 Classify layer. The structure of each C2f layer is as follows: Figure 6 As shown; correspondingly, the structure of the second intelligent model obtained by training the neural network intelligent model using the second dataset is the same as the structure of the neural network intelligent model.

[0133] Optionally, the process of using the second intelligent model to identify the tail trend type of the DC-AC simulation model based on the spectral image can be as follows: input the spectral image into the second intelligent model, so that the second intelligent model extracts the waveform stationary features from the waveform stationary part of the spectral image, and uses the waveform stationary features to identify the tail trend type of the DC-AC simulation model.

[0134] In this embodiment, since the structure of the second intelligent model is the same as that of the neural network intelligent model, the structure of the second intelligent model is the same as that of the first intelligent model. Therefore, the execution process of the second intelligent model extracting waveform stationary features from the waveform stationary part of the spectrum image and using the waveform stationary features to identify the tail trend type of the DC-AC simulation model can refer to the execution process of the first intelligent model described above. This embodiment does not limit it here.

[0135] Therefore, this application obtains a second intelligent model by pre-training a neural network intelligent model using a second dataset containing historical spectral images and their target tail trend types, so as to accurately identify the current tail trend type of the DC-AC simulation model based on the currently obtained spectral images using the second intelligent model.

[0136] S105: Determine whether the DC-AC simulation model meets the termination condition; if not, proceed to step S106.

[0137] In this embodiment of the application, a corresponding termination condition can be pre-configured, wherein the termination condition indicates that the response type is a specified response type and the tail trend type is a specified tail trend type, and determines that the DC-AC simulation model meets the termination condition; otherwise, it is determined that the DC-AC does not meet the termination condition.

[0138] During the specific execution of step S105, after obtaining the response type and tail trend type of the DC-AC simulation model, it can be determined whether the response type is the specified response type and whether the tail trend type is the specified tail trend type. If the response type is the specified response type and the tail trend type is the specified tail trend type, it is determined that the DC-AC simulation model meets the termination condition, and the control parameter tuning process of the DC-AC simulation model is terminated. If the response type is not the specified response type and / or the tail trend type is not the specified tail trend type, it is determined that the DC-AC simulation model does not meet the termination condition, and step S106 is executed.

[0139] It should be noted that the specified response type can be type D, and the specified tail trend type can be type Ok. Since a type D response and a type Ok tail trend indicate that the DC-AC simulation model is currently in an ideal response and stable convergence state, there is no need to tune the PI control parameters of the DC-AC simulation model in this state, and the tuning process can be terminated directly. Conversely, if the model is not in this state, the PI control parameters of the DC-AC simulation model need to be adjusted, and step S106 can be executed to perform the corresponding PI control parameter tuning operation.

[0140] S106: Determine the recommended adjustment ratios for Kp and Ki based on the response type and tail trend type.

[0141] In the specific execution step S106, the response type label corresponding to each response type can be pre-configured, and the Kp suggested adjustment ratio matching each response type label can be configured; the tail trend type label corresponding to each tail trend type can be pre-configured, and the Ki suggested adjustment ratio matching each tail trend type label can be configured; after determining the current response type and tail trend type of the DC-AC simulation model, the response type label to which the response type belongs can be determined, and the Kp suggested adjustment ratio matching the response type label can be determined from the pre-set Kp suggested adjustment ratios; the tail trend type label to which the tail trend type belongs can be determined, and the Ki suggested adjustment ratio matching the tail trend type label can be determined from the pre-set Ki suggested adjustment ratios.

[0142] It should be noted that the suggested adjustment ratios for the response type labels and their matching Kp values ​​for each pre-configured response type are shown in Table 1.

[0143] Table 1:

[0144]

[0145] Table 1 shows the label for each response category of the first intelligent model and the corresponding suggested adjustment ratio for Kp. In the actual iteration process, based on the actual operation of the code, these types were mathematically numbered (with corresponding response type labels). The initial value of Kp was the system default value or the result of the previous tuning round. The adjustment range was matched according to the classification results, that is, the suggested adjustment ratio for Kp was determined based on the response type label corresponding to the response type.

[0146] It should be noted that the tail trend type labels corresponding to each pre-configured tail trend type and their matching Ki suggested adjustment ratios are shown in Table 2.

[0147] Table 2:

[0148]

[0149] Table 2 shows the label for each tail trend type in the second intelligent model and the corresponding suggested adjustment ratio for Ki. During the actual iteration process, these types were mathematically numbered based on the actual operation of the code. The initial value of Ki is set by the system or passed from the previous iteration, and the tail trend type label is used to determine whether there are problems with insufficient or excessive integration.

[0150] S107: Perform primary parameter correction on Kp and Ki of the DC-AC simulation model using the proposed adjustment scales Kp and Ki.

[0151] In the specific execution step S107, after determining the initial suggested adjustment ratios for Kp and Ki, these ratios can be used to adjust Kp and Ki of the DC-AC simulation model to complete the initial parameter correction of the DC-AC simulation model. Specifically, this adjustment can be applied to the DC-AC simulation model's... and Adjustments need to be made to the DC-AC simulation model. and Adjustments will be made.

[0152] For example, if Kp is recommended to be adjusted to -10% and Ki is recommended to be adjusted to +8%, then Kp of the DC-AC simulation model can be adjusted downward by 10%, and Ki of the DC-AC simulation model can be adjusted upward by 8%.

[0153] It should be noted that the adjustment scales suggested by Kp and Ki are used for the DC-AC simulation model. and Perform initial parameter calibration, or adjust the DC-AC simulation model. and Primary parameter correction is performed, but this embodiment of the application is not limited to this.

[0154] S108: Determine the fine-tuning parameters based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction.

[0155] In this embodiment, based on the initial parameter correction, to improve the adjustment accuracy of the PI control parameters, further optimization of the PI control parameters can be performed. Specifically, a Proximal Policy Optimization (PPO) algorithm in reinforcement learning can be used. By training the agent to learn the optimal parameter adjustment strategy based on feedback from corresponding response image types (response type and tail trend type), the system response is driven towards "ideal response" and "stable convergence," resulting in the final policy network. Unlike traditional algorithms based on precise indicators (such as overshoot, steady-state error, etc.), this application uses "improvement of image classification labels" as the sole feedback criterion, constructing a label (response type and tail trend type) driven reinforcement learning strategy optimization adapted to unobservable physical quantity systems, thus obtaining an adapted policy network.

[0156] During the specific execution step S108, after completing the primary parameter calibration, the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter calibration can be obtained. This allows the policy network to output the corresponding target action based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter calibration, and to determine the fine-tuning parameters based on the target action. The target action includes adjusting Kp by the first step number upward / downward and Ki by the second step number upward / downward. The fine-tuning parameters can include + / - Kp by the first step number and + / - Ki by the second step number. Here, "+" indicates an upward adjustment, and "-" indicates a downward adjustment.

[0157] It should be noted that the corresponding action space A can be predefined in the policy network and the policy network outputs only one parameter each time, with a fixed direction and a step size that changes proportionally. The action space A includes 8 discrete actions, as shown in formula (19).

[0158] (19)

[0159] Where δ is the base step size, set to 0.05, and the action selection is obtained by sampling after the probability distribution is generated by the policy network.

[0160] It should also be noted that the structure of the policy network is shown in Table 3.

[0161] Table 3:

[0162]

[0163] Optionally, the process of determining the fine-tuning parameters based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction can be as follows: Obtain the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction. The historical simulation information includes the current tuning round, target number of times, historical response type label from the previous tuning, and historical tail trend type label. The target number of times is the number of tuning rounds within a preset range where the response type label and tail trend type label remain unchanged. The policy network determines the current state vector and reward based on the current Ki, current Kp, historical response type label, historical tail trend type label, current tuning round, and target number of times. The policy network outputs the corresponding target action based on the state vector and reward, and determines the fine-tuning parameters based on the target action. The target action includes adjusting the target number of steps up / down by Kp / Ki, and the fine-tuning parameters may include + / - the target number of steps by Kp / Ki.

[0164] In practical applications, the current Kp, current Ki, current tuning round, target number, historical response type label of the previous tuning, and historical tail trend type label of the DC-AC simulation model are input into the policy network. The policy network constructs a state vector based on the current Kp, current Ki, current tuning round, target number, historical response type label of the previous tuning, and historical tail trend type label, as shown in formula (20). The response type reward, tail trend reward, and plateau penalty are determined based on the state vector, as shown in formulas (21)-(23). The corresponding reward is calculated using the reward function based on the response type reward, tail trend reward, and plateau penalty, as shown in formula (24). The advantage function is constructed based on the reward and the state vector, as shown in formula (25). The probability ratio of each predefined action is calculated based on the state vector, as shown in formula (26). The optimization objective function based on the PPO pruning strategy is used to iteratively update the policy network parameters based on the joint calculation of the state vector, the advantage function, and the probability ratio of each action. In this process, the objective function Lclip(θ) is maximized as the optimization criterion, enabling the policy network to gradually learn to select the action with better performance (i.e., determine the target action) with higher probability under different states, thereby achieving adaptive optimization of the control parameters. The way to maximize the objective function is as shown in formula (27).

[0165] (20)

[0166] in, For state vectors, For the current Kp, For the current Ki; The historical response type label for the previous step (the last adjustment); The previous step's historical tail trend type label; This is the current tuning round for this tuning; This is the number of times that the consecutive response type labels and tail trend type labels remain unchanged within a preset round range. For example, the number of times that the consecutive labels (historical response type labels and historical tail trend type labels) remain unchanged in the previous few rounds.

[0167] (twenty one)

[0168] in, Rewards are given for response types.

[0169] (twenty two)

[0170] in, Rewards are given for tail trends.

[0171] (twenty three)

[0172] in, This is a penalty for the plateau period.

[0173] It should be noted that in formula (23), if the condition in parentheses is true, its value is 1; otherwise, it is 0. That is, when the label remains unchanged for 3 consecutive rounds, a micro-penalty is given to encourage exploratory behavior.

[0174] (twenty four)

[0175] in, As a reward.

[0176] (25)

[0177] in, The dominant function; This is a discount factor that controls the importance of future rewards (e.g., 0.99). The attenuation factor (GAE parameter, 0~1) controls the smoothness of the estimation; This refers to the timing difference error (TDerror), where... , , The current state The value estimate, In strategy The expectations below The immediate environmental reward obtained at step t+k; l represents the offset from the current time t to the future step l, where each l corresponds to the TD error (i.e., δ) at a future time. t And by accumulating the exponentially decaying weights of (γλ)^l, a balancing advantage estimate of short-term and long-term information is obtained.

[0178] (26)

[0179] in, Represents the state vector Next, the current strategy selects an action. The probability of; The parameters representing the old strategy (the parameters used to sample the data are fixed); Represents the same state vector Select this action under the old strategy The probability of; For action The probability ratio.

[0180] (27)

[0181] in, Let the objective function of the policy network be... These are the current parameters of the policy network; This represents the expected value of the data sampled at time step t (current time). The dominance function measures the action. The "good" or "bad" of a balancing strategy; The hyperparameter for the shear range is set to 0.1; This means limiting the probability ratio to Within the range; The probability ratio represents the action under the current policy. The probability of this action under the old strategy The ratio of the probability of .

[0182] It should be noted that after obtaining the current state vector, reward, and target action, the policy network adds the current state vector, reward, and target action to the corresponding experience buffer.

[0183] It should also be noted that after each round of interaction between the policy network and the simulation environment, the generated state vectors, target actions, rewards, and other data are stored in an experience buffer. Subsequently, the advantage function and probability ratio are calculated using the sample data in this buffer, and the policy network parameters are updated accordingly to achieve continuous policy optimization.

[0184] S109: Fine-tune Kp and Ki of the DC-AC simulation model after primary parameter correction using fine-tuning parameters.

[0185] In the specific execution step S109, after obtaining the fine-tuning parameters, the fine-tuning parameters can be used to fine-tune Kp and Ki of the DC-AC simulation model after the primary parameters are corrected, and then return to execute the DC-AC simulation model to obtain the corresponding waveform, until the DC-AC simulation model meets the termination condition.

[0186] It should be noted that fine-tuning parameters can include + / - the target number of steps, Kp / Ki. For example, fine-tuning parameters include... Then according to The calculated Kp is used to adjust the Kp of the DC-AC simulation model.

[0187] It should also be noted that after fine-tuning Kp and Ki of the DC-AC simulation model, it can be considered that one round of PI parameter tuning of the DC-AC simulation model has been completed. At this time, the corresponding tuning round can be recorded. For example, if the current tuning round is T, then after tuning the PI parameters of the DC-AC simulation model, 1 can be added to the current tuning round T to update the corresponding tuning round.

[0188] It is worth noting that a corresponding maximum number of rounds can also be set so that after updating the corresponding tuning rounds, it can be determined whether the updated tuning rounds are greater than the maximum number of rounds. If they are greater, the corresponding tuning process is terminated. If they are not greater, the process can return to step S101 to continue tuning the PI control parameters of the DC-AC simulation model.

[0189] Furthermore, in this application, after tuning the IP control parameters of the DC-AC simulation model, the DC-AC simulation model can be simulated and verified, and the verification results are as follows: Figures 7-10 As shown. Among them, from Figures 7-8 The displayed three-phase voltage and current waveforms show that the DC-AC simulation model outputs ideal sinusoidal voltage and current after tuning; from Figures 9-10 The voltage total harmonic distortion (THD) and current total harmonic distortion (THD) diagrams shown indicate that the total harmonic distortion (THD) of both the three-phase voltage and current is below 0.1%, and the system performance indicators meet the design requirements.

[0190] This application provides a method for tuning control parameters of an inverter. The method involves running a DC-AC simulation model to obtain corresponding waveforms, where the waveforms are voltage or current waveforms. A standardized spectrogram image is generated from the waveforms. A first intelligent model is used to identify the response type of the DC-AC simulation model based on the spectrogram image. The first intelligent model is obtained by training a neural network intelligent model using a first dataset. A second intelligent model is used to identify the tail trend type of the DC-AC simulation model based on the spectrogram image. The second intelligent model is obtained by training a neural network intelligent model using a second dataset. If it is determined that the DC-AC simulation model does not meet the termination criteria... The process involves determining the suggested adjustment ratios for Kp and Ki based on the response type and tail trend type; using these ratios, performing primary parameter correction on Kp and Ki of the DC-AC simulation model; determining fine-tuning parameters based on the current Kp and Ki of the DC-AC simulation model after primary parameter correction and historical simulation information; fine-tuning Kp and Ki of the DC-AC simulation model after primary parameter correction using these parameters; and then returning to execute the DC-AC simulation model to obtain the corresponding waveform. This process continues until the DC-AC simulation model meets the termination conditions, thereby achieving end-to-end closed-loop automated control. Therefore, this application utilizes pre-trained first and second intelligent models to identify the response type and tail trend type of the DC-AC simulation model, enabling primary parameter correction based on these characteristics. This transforms the control performance, which traditionally requires precise quantization, into a discrete label representation. Furthermore, after primary parameter correction, a label feedback mechanism in reinforcement learning can be used to further determine fine-tuning parameters. These parameters are then used to fine-tune the PI control parameters of the DC-AC simulation model after primary parameter correction, thereby improving the adjustment accuracy and robustness of the PI control parameters. This label feedback mechanism overcomes the high dependence of existing methods on system modeling accuracy, numerical observability, and index function analytical properties. It is particularly suitable for real-world scenarios or data simulation environments where only spectral images or response type feedback can be obtained. Moreover, the entire process does not require nested structure recognition, feature engineering, or accuracy matching modeling, significantly simplifying the tuning path. In other words, the entire process requires no manual intervention or frequent parameter adjustments, which not only improves the system's automation level but also reduces operational complexity.

[0191] Based on the inverter control parameter tuning method provided in the embodiments of this application, correspondingly, the inverter control parameter tuning system is also provided in the embodiments of this application, such as... Figure 11 As shown, the system includes:

[0192] Run Unit 1 is used to run the DC-AC simulation model to obtain the corresponding waveforms; where the waveforms are voltage waveforms or current waveforms.

[0193] The spectral image generation unit 2 is used to generate a standardized spectral image using waveforms;

[0194] The response type identification unit 3 is used to identify the response type of the DC-AC simulation model based on the graph image using the first intelligent model; wherein, the first intelligent model is obtained by training the neural network intelligent model using the first dataset;

[0195] Tail trend type identification unit 4 is used to identify the tail trend type of the DC-AC simulation model based on the graph image using the second intelligent model; wherein, the second intelligent model is obtained by training the neural network intelligent model using the second dataset;

[0196] The preliminary parameter correction unit 5 is used to determine the suggested adjustment ratios of Kp and Ki based on the response type and tail trend type if it is determined that the DC-AC simulation model does not meet the termination conditions, and to perform preliminary parameter correction on Kp and Ki of the DC-AC simulation model using the suggested adjustment ratios of Kp and Ki.

[0197] Fine-tuning unit 6 is used to determine fine-tuning parameters based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, and to fine-tune Kp and Ki of the DC-AC simulation model after primary parameter correction using the fine-tuning parameters before returning to the execution unit.

[0198] This application provides a control parameter tuning system for an inverter. It obtains corresponding waveforms by running a DC-AC simulation model, where the waveforms are voltage or current waveforms. A standardized spectrogram image is generated from the waveforms. A first intelligent model identifies the response type of the DC-AC simulation model based on the spectrogram image. The first intelligent model is obtained by training a neural network intelligent model using a first dataset. A second intelligent model identifies the tail trend type of the DC-AC simulation model based on the spectrogram image. The second intelligent model is obtained by training a neural network intelligent model using a second dataset. If it is determined that the DC-AC simulation model does not meet the termination condition... The process involves determining the suggested adjustment ratios for Kp and Ki based on the response type and tail trend type; using these ratios, performing primary parameter correction on Kp and Ki of the DC-AC simulation model; determining fine-tuning parameters based on the current Kp and Ki of the DC-AC simulation model after primary parameter correction and historical simulation information; fine-tuning Kp and Ki of the DC-AC simulation model after primary parameter correction using these parameters; and then returning to execute the DC-AC simulation model to obtain the corresponding waveform. This process continues until the DC-AC simulation model meets the termination conditions, thereby achieving end-to-end closed-loop automated control. Therefore, this application utilizes pre-trained first and second intelligent models to identify the response type and tail trend type of the DC-AC simulation model, enabling primary parameter correction based on these characteristics. This transforms the control performance, which traditionally requires precise quantization, into a discrete label representation. Furthermore, after primary parameter correction, a label feedback mechanism in reinforcement learning can be used to further determine fine-tuning parameters. These parameters are then used to fine-tune the PI control parameters of the DC-AC simulation model after primary parameter correction, thereby improving the adjustment accuracy and robustness of the PI control parameters. This label feedback mechanism overcomes the high dependence of existing methods on system modeling accuracy, numerical observability, and index function analytical properties. It is particularly suitable for real-world scenarios or data simulation environments where only spectral images or response type feedback can be obtained. Moreover, the entire process does not require nested structure recognition, feature engineering, or accuracy matching modeling, significantly simplifying the tuning path. In other words, the entire process requires no manual intervention or frequent parameter adjustments, which not only improves the system's automation level but also reduces operational complexity.

[0199] Optionally, the DC-AC simulation model is the main circuit topology of a three-phase two-level inverter. The DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

[0200] Optionally, a response type identification unit is used to identify the response type of the DC-AC simulation model based on the spectral image using the first intelligent model. Specifically, this unit is used for:

[0201] The spectrum image is input into the first intelligent model, which extracts the peak characteristics and oscillation time characteristics of the waveform from the spectrum image, and uses the peak characteristics and oscillation time characteristics to identify the response type of the DC-AC simulation model.

[0202] Optionally, a tail trend type identification unit is used to identify the tail trend type of the DC-AC simulation model based on the spectral image using the second intelligent model. Specifically, this is used for:

[0203] The spectrum image is input into the second intelligent model, which extracts the waveform stationarity features from the stationary portion of the spectrum image and uses these features to identify the tail trend type of the DC-AC simulation model.

[0204] Optionally, this application implements a control parameter tuning system for an inverter, and further includes a determination unit for:

[0205] Determine whether the response type is the specified response type and whether the tail trend type is the specified tail trend type; if the response type is the specified response type and the tail trend type is the specified tail trend type, determine that the DC-AC simulation model meets the termination condition and terminate the control parameter tuning process of the DC-AC simulation model; if the response type is not the specified response type and / or the tail trend type is not the specified tail trend type, determine that the DC-AC simulation model does not meet the termination condition.

[0206] Optionally, a primary parameter correction unit is used to determine the suggested adjustment ratios for Kp and Ki based on the response type and tail trend type, specifically for:

[0207] Determine the response type label to which the response type belongs, and determine the Kp suggested adjustment ratio that matches the response type label from the preset Kp suggested adjustment ratios; determine the tail trend type label to which the tail trend type belongs, and determine the Ki suggested adjustment ratio that matches the tail trend type label from the preset Ki suggested adjustment ratios.

[0208] Optionally, a spectral image generation unit that generates a standardized spectral image using waveforms is specifically used for:

[0209] The corresponding imaging frame is configured based on the model information of the DC-AC simulation model and the waveform information of the waveform. The leftmost part of the imaging frame is based on the initial response point of the waveform, the central axis of the imaging frame is the target given value, the height of the imaging frame is configured according to the current working condition of the DC-AC simulation model, and the width of the imaging frame is configured according to the maximum response time of the waveform. The initial spectral image is acquired from the waveform according to the imaging frame, and the initial spectral image is standardized to obtain a standardized spectral image.

[0210] Optionally, based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, the fine-tuning unit for fine-tuning parameters is determined, specifically for:

[0211] The system acquires the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after initial parameter correction. Historical simulation information includes the current tuning round, target number of iterations, historical response type label from the previous tuning, and historical tail trend type label. The target number of iterations is the number of tuning iterations within a preset range where the response type and tail trend type label remain unchanged. The policy network determines the current state vector and reward based on the current Ki, current Kp, historical response type label, historical tail trend type label, current tuning round, and target number of iterations. The policy network outputs the corresponding target action based on the state vector and reward, and determines fine-tuning parameters based on the target action. The target action includes adjusting the target number of steps up / down by Kp / Ki, and the fine-tuning parameters include increasing / decreasing the target number of steps by Kp / Ki.

[0212] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0213] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0214] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0215] The above are merely 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 principle 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 tuning control parameters of an inverter, characterized in that, The method includes: Run the DC-AC simulation model to obtain the corresponding waveforms; wherein the waveforms are voltage waveforms or current waveforms. A standardized spectral image is generated using the waveform; The response type of the DC-AC simulation model is identified based on the graph image using a first intelligent model; wherein, the first intelligent model is obtained by training a neural network intelligent model using a first dataset; the response type is characterized by the following: the DC-AC simulation model currently exhibits obvious overtuning, the DC-AC simulation model currently exhibits slow / undertuned response, the DC-AC simulation model currently exhibits continuous oscillation, or the DC-AC simulation model currently exhibits an ideal response. The tail trend type of the DC-AC simulation model is identified based on the graph image using a second intelligent model; wherein, the second intelligent model is obtained by training the neural network intelligent model using a second dataset; the tail trend type is characterized by a large steady-state deviation in the DC-AC simulation model, a severe integral oscillation in the DC-AC simulation model, or a stable convergence of the DC-AC simulation model. If it is determined that the DC-AC simulation model does not meet the termination conditions, the recommended adjustment ratio of Kp is determined according to the response type, and the recommended adjustment ratio of Ki is determined according to the tail trend type. The initial parameter correction of Kp and Ki in the DC-AC simulation model is performed using the proposed adjustment ratios for Kp and Ki. Based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, determine the fine-tuning parameters; The process involves fine-tuning Kp and Ki of the DC-AC simulation model after the primary parameters have been corrected using the fine-tuning parameters, and then returning to run the DC-AC simulation model to obtain the corresponding waveform. The step of generating a standardized spectral image using the waveform includes: The corresponding imaging frame is configured according to the model information of the DC-AC simulation model and the waveform information of the waveform; wherein, the leftmost side of the imaging frame is based on the initial response point of the waveform, the central axis of the imaging frame is the target given value, the height of the imaging frame is configured according to the current working condition of the DC-AC simulation model, and the width of the imaging frame is configured according to the maximum response time of the waveform. An initial spectral image is acquired from the waveform according to the imaging frame, and the initial spectral image is standardized to obtain a standardized spectral image.

2. The method according to claim 1, characterized in that, The DC-AC simulation model is the main circuit topology of a three-phase two-level inverter. The DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

3. The method according to claim 1, characterized in that, The first intelligent model identifies the response type of the DC-AC simulation model based on the spectral image, including: The spectrum image is input into the first intelligent model, which extracts the peak features and oscillation time features of the waveform from the spectrum image, and uses the peak features and oscillation time features to identify the response type of the DC-AC simulation model.

4. The method according to claim 1, characterized in that, The second intelligent model is used to identify the tail trend type of the DC-AC simulation model based on the spectral image, including: The spectrum image is input into the second intelligent model, which extracts waveform stationarity features from the stationary portion of the spectrum image and uses these features to identify the tail trend type of the DC-AC simulation model.

5. The method according to claim 1, characterized in that, The method further includes: Determine whether the response type is a specified response type and whether the tail trend type is a specified tail trend type; If the response type is a specified response type and the tail trend type is a specified tail trend type, the DC-AC simulation model is determined to meet the termination condition, and the control parameter tuning process of the DC-AC simulation model is terminated. If the response type is not the specified response type, and / or the tail trend type is not the specified tail trend type, the DC-AC simulation model is determined not to meet the termination condition.

6. The method according to claim 1, characterized in that, Determining the recommended adjustment ratios for Kp and Ki based on the response type and the tail trend type includes: Determine the response type label to which the response type belongs, and determine the Kp suggestion adjustment ratio that matches the response type label from a set of preset Kp suggestion adjustment ratios; Determine the tail trend type label to which the tail trend type belongs, and determine the Ki suggested adjustment ratio that matches the tail trend type label from the preset Ki suggested adjustment ratios.

7. The method according to claim 1, characterized in that, Based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, fine-tuning parameters are determined, including: Obtain the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction. The historical simulation information includes the current tuning round, target number, historical response type label and historical tail trend type label of the DC-AC simulation model in the previous tuning. The target number is the number of tuning rounds in which the response type label and tail trend type label remain unchanged for consecutive cycles within a preset number of rounds. The current state vector and reward are determined by the policy network based on the current Ki, the current Kp, the historical response type label, the historical tail trend type label, the current tuning round, and the target number. The policy network outputs a corresponding target action based on the state vector and reward, and determines fine-tuning parameters based on the target action. The target action includes adjusting the target number of steps up / down by Kp / Ki, and the fine-tuning parameters include adding / subtracting the target number of steps by Kp / Ki.

8. A control parameter tuning system for an inverter, characterized in that, The system includes: The running unit is used to run the DC-AC simulation model to obtain the corresponding waveforms; wherein the waveforms are voltage waveforms or current waveforms. The spectral image generation unit is used to generate a standardized spectral image using the waveform; The response type identification unit is used to identify the response type of the DC-AC simulation model based on the spectral image using a first intelligent model; wherein, the first intelligent model is obtained by training a neural network intelligent model using a first dataset; the response type is characterized by the following: the DC-AC simulation model currently exhibits obvious overtuning, the DC-AC simulation model currently exhibits slow / undertuned response, the DC-AC simulation model currently exhibits continuous oscillation, or the DC-AC simulation model currently exhibits an ideal response. The tail trend type identification unit is used to identify the tail trend type of the DC-AC simulation model based on the graph image using a second intelligent model; wherein, the second intelligent model is obtained by training the neural network intelligent model using a second dataset; the tail trend type is characterized by a large steady-state deviation in the DC-AC simulation model, a severe integral oscillation in the DC-AC simulation model, or a stable convergence of the DC-AC simulation model. The preliminary parameter correction unit is used to determine the recommended adjustment ratio of Kp based on the response type and the recommended adjustment ratio of Ki based on the tail trend type if it is determined that the DC-AC simulation model does not meet the termination condition, and to perform preliminary parameter correction on Kp and Ki of the DC-AC simulation model using the recommended adjustment ratio of Kp and the recommended adjustment ratio of Ki. The fine-tuning unit is used to determine the fine-tuning parameters based on the current Kp, current Ki, and historical simulation information of the DC-AC simulation model after primary parameter correction, and to fine-tune the Kp and Ki of the DC-AC simulation model after primary parameter correction using the fine-tuning parameters, and then return to the execution unit. The spectral image generation unit, which generates a standardized spectral image using waveforms, is specifically used for: The corresponding imaging frame is configured based on the model information of the DC-AC simulation model and the waveform information of the waveform. The leftmost part of the imaging frame is based on the initial response point of the waveform, the central axis of the imaging frame is the target given value, the height of the imaging frame is configured according to the current working condition of the DC-AC simulation model, and the width of the imaging frame is configured according to the maximum response time of the waveform. The initial spectral image is acquired from the waveform according to the imaging frame, and the initial spectral image is standardized to obtain a standardized spectral image.

9. The system according to claim 8, characterized in that, The DC-AC simulation model is the main circuit topology of a three-phase two-level inverter. The DC-AC simulation model adopts a preset decoupling control strategy and a dual closed-loop controller, and uses a feedforward compensation term to decouple the parameters of the dual closed-loop controller.

Images