Converter wide frequency resonance suppression method and system based on generalized regression neural network impedance identification
By identifying grid impedance through a generalized regression neural network and constructing an adaptive parameter adjustment model, the resonance problem of LCL converters was solved, the stable grid connection of the converter system was achieved, and the grid connection quality was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID JIANGSU ELECTRIC POWER CO LIANYUNGANG POWER SUPPLY CO
- Filing Date
- 2025-08-19
- Publication Date
- 2026-07-21
AI Technical Summary
The resonance problem of LCL-type grid-connected converters leads to system instability, affects grid connection quality, and hinders the development of distributed generation technology and microgrids.
An impedance identification method based on a generalized regression neural network is adopted. By adaptively adjusting the parameters of the inner and outer loop controls, a dataset of harmonics and grid impedance is constructed. The neural network is trained to predict the grid impedance in real time, and the transfer function parameters of the inner and outer loops are adjusted to achieve resonance suppression.
It improves the resonance suppression effect of the converter system, ensures stable system operation, optimizes controller parameters, and enhances grid connection quality.
Smart Images
Figure CN120914885B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for suppressing wideband resonance in converters, specifically to a method and system for suppressing wideband resonance in converters based on impedance identification using a generalized regression neural network. Background Technology
[0002] Against the backdrop of "carbon peaking and carbon neutrality," sustainable new energy sources, represented by wind and solar power, are widely used across various regions. As the global energy structure accelerates its transition to cleaner energy, the installed capacity of new energy power generation such as wind and solar power is experiencing explosive growth. With the large-scale grid connection of wind and solar power using power electronic equipment as interfaces, LCL-type grid-connected converters are being used extensively. LCL filters, based on L-type and LC-type filters, possess superior high-frequency attenuation characteristics while retaining the low-frequency characteristics of L-type filters, and also offer smaller size and lower cost. However, due to the large-scale use of LCL-type grid-connected converters, the resulting resonance problem is becoming increasingly significant. The primary factor stems from the fact that the LCL filter is a third-order oscillating system. When the system reaches a certain frequency, it will be in a resonant state, generating severe spikes at the resonant frequency, which can easily lead to system instability and significantly affect grid connection quality. This resonance problem of LCL-type grid-connected converters, to some extent, hinders the large-scale application of distributed generation technology and the development of microgrids. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for suppressing wideband resonance in converters based on impedance identification using a generalized regression neural network. This method solves the problems of difficulty and insufficient accuracy in calculating the grid impedance of converter-connected systems, and achieves wideband resonance suppression of converters.
[0004] To achieve the above objectives, the present invention employs the following technical solution:
[0005] A converter broadband resonance suppression method based on impedance identification using a generalized regression neural network includes:
[0006] A converter system based on inner and outer loop control is constructed. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point.
[0007] Based on the transfer functions of the inner and outer loops, an adaptive parameter adjustment model for the parameters in the inner and outer loop transfer functions relative to the grid impedance is constructed.
[0008] A generalized regression neural network for power grid impedance prediction was constructed, and a dataset corresponding to harmonics and power grid impedance was established to train the generalized regression neural network.
[0009] The grid impedance is predicted in real time using a trained generalized regression neural network. The parameters in the inner and outer loop transfer functions are adjusted by using the predicted grid impedance and the adaptive parameter adjustment formula to suppress converter resonance.
[0010] Furthermore, the inner and outer loops employ closed-loop control and open-loop control, respectively.
[0011] Furthermore, the closed-loop transfer function of the inner loop is:
[0012]
[0013] in, and For adaptive adjustment parameters.
[0014] Furthermore, adaptive adjustment parameters and The adaptive parameter adjustment model relative to the grid impedance is as follows:
[0015]
[0016] in, For converter filter inductance, For the grid-side inductance of the converter, For grid impedance, For filter capacitors, parameters .
[0017] Furthermore, the closed-loop transfer function of the outer loop is:
[0018]
[0019] in, To adaptively adjust parameters, To set parameters.
[0020] Furthermore, adaptive adjustment parameters The adaptive parameter adjustment model relative to the grid impedance is as follows:
[0021]
[0022] In the formula, = , For converter filter inductance, For the grid-side inductance of the converter, This represents the power grid impedance.
[0023] Furthermore, a generalized regression neural network for power grid impedance prediction is constructed, and a dataset corresponding to harmonics and power grid impedance is established to train the generalized regression neural network: the neural network is trained, the accuracy of the output value is verified during the training process, and the generalized regression theory is introduced to estimate the conditional probability density function through the kernel function method. At the same time, combined with the expert database and expert rules, the optimized power grid impedance prediction value is output.
[0024] Furthermore, the data in the dataset is acquired by installing current sensors on the AC side of the converter and arranging voltage sensors at the common coupling point.
[0025] Furthermore, the conditional probability density function is estimated using the kernel function method: the inter-layer transformation of the neural network is represented by a kernel function, and the matrix multiplication of the fully connected layer is replaced by the kernel function to achieve kernelized linear transformation.
[0026] A converter broadband resonance suppression system based on impedance identification using a generalized regression neural network includes:
[0027] The converter system building unit is used to build a converter system based on inner and outer loop control. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point.
[0028] An adaptive parameter adjustment model building unit is constructed based on the transfer functions of the inner and outer loops to build an adaptive parameter adjustment model of the parameters in the inner and outer loop transfer functions relative to the grid impedance.
[0029] The generalized regression neural network construction and training unit constructs a generalized regression neural network for power grid impedance prediction and establishes a dataset corresponding to harmonics and power grid impedance to train the generalized regression neural network.
[0030] The parameter adjustment unit predicts the grid impedance in real time using a trained generalized regression neural network; it adjusts the parameters in the inner and outer loop transfer functions based on the predicted grid impedance and the adaptive parameter adjustment formula to suppress converter resonance.
[0031] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0032] (1) This invention introduces expert rules and generalized regression theory into the traditional neural network, which solves the problems of difficulty in calculating grid impedance and insufficient accuracy in converter grid-connected systems, and the problem that the system instability caused by resonance at a specific frequency significantly affects the grid connection quality; (2) By reading and collecting the current harmonics and grid voltage harmonics output by the converter, a database corresponding to the harmonics and grid impedance is established, and the sample data is processed and calculated through the GRNN structure. The accuracy of the predicted grid impedance data is calibrated, and the neural network is repeatedly trained. Finally, the data in the database is used as prior knowledge for learning and reasoning, and the predicted grid impedance is optimized, which greatly improves the accuracy of the predicted data; (3) The parameters of the resonance compensation link are designed. and Parameters of the current controller The adaptive parameter adjustment system structure, through the adaptive adjustment of three parameters, plays a role in suppressing resonance, ensuring stable system operation, and providing a theoretical basis for the optimized design of controller parameters. Attached Figure Description
[0033] Figure 1 This is a system block diagram of a resonance suppression scheme based on adaptive parameter adjustment;
[0034] Figure 2 This is a block diagram of the control structure for the resonant compensation circuit;
[0035] Figure 3 These are the parameters of the resonant compensation stage. With grid impedance Relationship diagram;
[0036] Figure 4 Resonance compensation parameters With grid impedance Relationship diagram;
[0037] Figure 5 It is the proportional coefficient of the current controller. With grid impedance Relationship diagram;
[0038] Figure 6 These are experimental waveforms of grid-connected current and grid-connected voltage of a converter without adaptive parameter adjustment method;
[0039] Figure 7 These are experimental waveforms of grid-connected current and grid-connected voltage of a converter using an adaptive parameter adjustment method.
[0040] Figure 8 This is a flowchart of a generalized regressive neural network;
[0041] Figure 9 This is a diagram of the generalized regression neural network structure. Detailed Implementation
[0042] The technical solution of the present invention will be described in detail below with reference to the principle block diagram and specific embodiments. It should be understood that the specific features of the embodiments of this application are a detailed description of the technical solution of this application, rather than a limitation on the technical solution of this application. In the absence of conflict, the embodiments of this application and the technical features in the embodiments can be combined with each other.
[0043] This invention provides a broadband resonance suppression method for converters based on impedance identification using a generalized regression neural network. The method directly identifies the equivalent grid impedance using a generalized regression neural network, which includes harmonic reading and data acquisition, database establishment, neural network training, prediction accuracy, generalized regression theory, and expert rule data optimization. By identifying the grid impedance, accurate data can be provided for controller parameter design, improving the resonance suppression effect. Specifically, this suppression method includes:
[0044] Step 1: Build a converter system based on inner and outer loop control. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point.
[0045] Step 2: Based on the transfer functions of the inner and outer loops, construct an adaptive parameter adjustment model for the parameters in the inner and outer loop transfer functions relative to the grid impedance;
[0046] Step 3: Construct a generalized regression neural network for power grid impedance prediction, and establish a dataset corresponding to harmonics and power grid impedance to train the generalized regression neural network.
[0047] Step 4: Predict the grid impedance in real time using the trained generalized regression neural network; adjust the parameters in the inner and outer loop transfer functions using the predicted grid impedance and the adaptive parameter adjustment formula to suppress converter resonance.
[0048] The system block diagram of the converter resonance suppression scheme based on adaptive impedance reshaping is as follows: Figure 1 As shown, the converter is connected to the power grid through the Point of Common Coupling (PCC). This is the converter-side current value. For the converter grid-connected current, For filtering capacitors; For converter filter inductance, For the grid-side inductance of the converter, For grid impedance, This is the output voltage at PCC. for The magnitude of the current command value in the coordinate system. The phase angle of the voltage at PCC.
[0049] The LCL filter structure, inner and outer loop control loops in the resonance suppression scheme system are all well-known structures in the art. The present invention mainly designs the transfer functions of the inner and outer loops. Therefore, the topology of the LCL filter structure, inner and outer loop control loops will not be described in detail in this embodiment.
[0050] Step 2, based on the transfer functions of the inner and outer loops, construct an adaptive parameter adjustment model for the parameters in the inner and outer loop transfer functions relative to the grid impedance, specifically including:
[0051] The closed-loop transfer function of the inner loop of the resonant compensation circuit is designed as follows: ,in, and For adaptive adjustment parameters.
[0052] The open-loop transfer function of the converter system is designed as follows: ,in, To adaptively adjust parameters, To set parameters.
[0053] Equivalent control structure diagram of resonant compensation circuit Figure 2 As shown, the characteristic equation of the internal current loop system is as follows:
[0054]
[0055] In the formula, A = It is the dominant coefficient in the high-frequency band and is related to the LC filtering stage; B= It is the damping modulation term, which mainly reflects the effect of the controller; C= It is the low-frequency inertia coefficient, determined by the inductance of the main circuit; D= It is a feedback compensation term that includes the controller's parameter gain. ;
[0056] Since a converter system typically has one inherent origin pole, which corresponds to the DC steady-state characteristic, the remaining three poles in the system are determined by... and This is a joint decision. To avoid overdamping of the system and resulting in an excessively slow dynamic response, and to maintain system stability, a pair of conjugate complex poles needs to be configured. And there must be a real pole. The characteristic equation for the reconstruction is further derived as follows:
[0057]
[0058] In the formula, The natural angular frequency of the system. The damping coefficient is... for Distance from the imaginary axis and the real pole The ratio;
[0059] Combining the characteristic equations of the internal current loop system, we can obtain the energy representation. , same , and The expression for the relationship between them is:
[0060]
[0061] The correlation analysis between the system's open-loop frequency characteristics and dynamic performance shows that the resonant frequency... The conditions that must be met are:
[0062]
[0063] The next relation can be derived as follows:
[0064]
[0065] when As it gets closer and closer to 1, The value is getting closer and closer to 0, and the value is too small. In order to obtain a normal value, Take 1.18.
[0066] when When the damping ratio decreases, it can be seen from the system damping constraint relationship that the damping ratio... It will decay in the opposite direction. At this time, in the system's pole distribution, there is a pair of conjugate complex poles. It gradually gains dominance and becomes the core factor determining the system's dynamic response. However, it is also important to note... The value of , if If the damping is too small, the system will exhibit an underdamped state, with violent oscillations in the output response and a surge in overshoot, making it difficult to effectively suppress the inherent resonance of the LCL converter, and potentially even leading to system instability due to continuous oscillation. Conversely, if the damping is too large... Increase It will become the dominant pole, but the dynamic response speed will be severely limited, and the delay in converter output adjustment and grid disturbance compensation will increase.
[0067] It can be deduced that:
[0068]
[0069] when When =0.5, It will be a very small value. It is necessary to make... exist The frequency response at a certain point must be close to that of a second-order differential filter to satisfy the condition for suppressing LCL filter resonance in the system. And when... When =0.5, Less than It approximately conforms to the frequency characteristics of a second-order differential filter;
[0070] Based on the above inferences, the following can be obtained: and The adaptive parameter adjustment formula is:
[0071]
[0072] in:
[0073]
[0074] The parameters of the resonance compensation element can be obtained from this. With grid impedance Relationship diagram as follows Figure 3 As shown, the parameters of the resonant compensation circuit With grid impedance Relationship diagram as follows Figure 4 As shown. By Figure 3 It can be seen that the parameters of the resonant compensation circuit are... With grid impedance The increase is synchronous with the increase, and there is a large range of variation, while the increase is caused by Figure 4 The parameters of the resonant compensation element can be obtained. With grid impedance The increase was accompanied by a simultaneous decrease, showing a trend of rapid increase followed by a gradual decrease, eventually stabilizing to a relatively stable level.
[0075] When it is below the cutoff frequency In the low-frequency range, the capacitor branch can be ignored, and the LCL filter can be simplified to... Therefore, the open-loop transfer function of the system at this point can be simplified to:
[0076]
[0077] Further derivation yields the system's open-loop transfer function. The amplitude-frequency characteristic at that point is:
[0078]
[0079] In the formula, = .
[0080] The system open-loop transfer function is The phase frequency characteristic at this point is:
[0081]
[0082] Next, when the system is in the low-frequency range, the open-loop transfer function of the system is further derived. The phase frequency characteristic at this point is:
[0083]
[0084] Combining the first aspect, we can further obtain the phase margin. Further calculations yielded Through derivation, we obtain The adaptive adjustment formula relative to the grid impedance is:
[0085]
[0086] In the formula, =
[0087] The parameters of the current controller can be obtained from this. With grid impedance Relationship diagram as follows Figure 5 As shown in the figure. The parameters of the current controller can be seen from the figure. With grid impedance The value gradually decreases as the value increases, and the range of variation is relatively large, indicating that the resonance coefficient... The necessity of regulation.
[0088] Step 3, the generalized regression neural network, mainly consists of the following parts: harmonic reading and data acquisition, database establishment, neural network training, prediction accuracy, generalized regression theory, and expert rule data optimization. Specifically:
[0089] (1) Harmonic reading and data acquisition: A high-precision current sensor is installed on the AC side of the converter to collect the output current in real time. A voltage sensor is placed at the point of common coupling (PCC) to synchronously collect the grid voltage. Furthermore, the Fast Fourier Transform (FFT) is used to decompose the current and voltage signals and extract harmonic information.
[0090]
[0091] in, and For harmonic amplitude, , The phase angle, This is the fundamental frequency of the converter output.
[0092] Construct feature vectors.
[0093] Input vector: , dimension
[0094] Target vector: , dimension
[0095] (2) Database Establishment: Establish a database corresponding to harmonics and grid impedance. First, preprocess the data on harmonics and grid impedance. Perform outlier detection on the read harmonics and collected data. Based on the 3σ criterion, identify and remove outliers. Label the harmonic features after Fast Fourier Transform decomposition and store them in the harmonic feature table. Calculate the corresponding grid impedance and store it in the impedance calculation table; specifically including:
[0096] First, data preprocessing is performed. The read harmonics and acquired data are compared to detect outliers. Based on the 3σ criterion, outliers (such as amplitude and phase abrupt changes) are identified and removed.
[0097]
[0098] in, The mean of the data. For a single data point. For missing harmonic data, linear interpolation is used to fill in the gaps:
[0099]
[0100] in, For missing data, , Data from adjacent time points, This is a timestamp. Next, the data is imported into the database: the preprocessed data... and Store the data in the original signal table according to the time series. Then, decompose the data using the Fast Fourier Transform. and Harmonic characteristics are labeled and stored in the harmonic characteristic table. The corresponding grid impedance is calculated and stored in the impedance calculation table. A strict one-to-one correspondence is maintained between harmonic characteristics and grid impedance values.
[0101] After establishing a database corresponding to harmonics and grid impedance, the process includes: training the neural network, which outputs predicted grid impedance values through an input layer, mode layer, summation layer, and output layer, followed by accuracy verification and repeated training. Generalized regression theory and expert rules are applied, and the conditional probability density function is estimated based on the kernel function method. Non-parametric regression modeling of local data allows the model to better adapt to complex data distributions and nonlinear relationships. The kernel function method represents the inter-layer transformation of the neural network using kernel functions, replacing matrix multiplication in fully connected layers to achieve "kernelized linear transformation." The expert rules use the database data corresponding to harmonics and grid impedance as "prior knowledge," comparing the predicted grid impedance values output by the neural network with the "prior knowledge," verifying the rationality of the results through expert rules, optimizing the predicted data, and finally outputting the predicted grid impedance values. Specifically, this includes:
[0102] (3) Training the neural network: First, adapt the GRNN structure.
[0103] Input layer: Number of nodes Receive current harmonics .
[0104] Pattern layer: Introduces a Gaussian kernel function to calculate the similarity between the input sample and the training set samples. The Gaussian kernel function is denoted as:
[0105]
[0106] in, An adaptive smoothing factor is optimized through cross-validation.
[0107] Summation layer:
[0108]
[0109] Output layer:
[0110]
[0111] GRNN model training: Sample selection involves evenly sampling from the database according to operating conditions to ensure coverage of the entire operating range. An incremental learning mode is adopted, dynamically updating the GRNN parameters as new data is added to the database.
[0112]
[0113] in For model parameters, For learning rate, This is the loss function.
[0114] Based on predicted voltage harmonics and measured current harmonics Calculate the power grid impedance under each harmonic:
[0115]
[0116] (4) Prediction accuracy: Verify whether the accuracy requirements are met by using the mean square error (MSE). ) and mean absolute error ( )check:
[0117]
[0118] in, The number of samples in the validation / test set. For the actual impedance, To predict impedance
[0119] If the accuracy requirements are met (MSE and MAE are below the preset threshold), the process proceeds to the generalized regression theory and expert rule data optimization stage.
[0120] If the requirements are not met, return to (1) to reread the acquired harmonic current, supplement the diverse data, and retrain the GRNN until the accuracy is met.
[0121] (5) Generalized regression theory and expert rule data optimization: The performance index function after introducing the kernel function is:
[0122]
[0123] in, For the first The input for the next iteration; For the first The expected input for the next iteration; For the first The error function of the next iteration.
[0124] The parameters were adjusted using the negative gradient descent method as shown below:
[0125]
[0126] in, This is the negative gradient of the performance index function with respect to the error function.
[0127] Generalized regression function Choose an approximate sign function, whose expression is as follows:
[0128]
[0129] Therefore, the formula for the weight parameters can be obtained as follows:
[0130]
[0131] To improve the parameter performance and convergence of generalized neural networks, an expert system is introduced. Expert rules are based on the knowledge and experience of domain experts and are intuitive and easy to understand. A database comparison method is used, treating the database as prior knowledge. The output predicted grid impedance value is compared with the grid impedance values stored in similar harmonics in the database, thereby further optimizing the predicted grid impedance data.
[0132] Generally, expert rules are set up in two cases:
[0133] 1) When the absolute error is large:
[0134]
[0135] in, This represents the maximum value of the error.
[0136] Therefore, when the absolute value of the error function is not less than the set maximum error value, the expression of the generalized regression function is:
[0137]
[0138] 2) When the absolute error is small:
[0139]
[0140] in, This is the median of the error; This represents the minimum value of the error.
[0141] Therefore, when the absolute value of the error function is greater than or equal to the set minimum error and less than the median, the generalized regression function can be expressed as:
[0142]
[0143] Finally, after comparison and optimization using expert rules, the predicted power grid impedance value optimized by expert rules is output.
[0144] This invention also provides a converter broadband resonance suppression system based on generalized regression neural network impedance identification, comprising:
[0145] The converter system building unit is used to build a converter system based on inner and outer loop control. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point.
[0146] An adaptive parameter adjustment model building unit is constructed based on the transfer functions of the inner and outer loops to build an adaptive parameter adjustment model of the parameters in the inner and outer loop transfer functions relative to the grid impedance.
[0147] The generalized regression neural network construction and training unit constructs a generalized regression neural network for power grid impedance prediction and establishes a dataset corresponding to harmonics and power grid impedance to train the generalized regression neural network.
[0148] The parameter adjustment unit predicts the grid impedance in real time using a trained generalized regression neural network; it adjusts the parameters in the inner and outer loop transfer functions based on the predicted grid impedance and the adaptive parameter adjustment formula to suppress converter resonance.
[0149] The embodiments of the present invention were experimentally verified, and the grid-connected converter parameter settings are shown in Table 1.
[0150] Table 1 Parameters of Grid-Connected Converter
[0151]
[0152] To verify the resonance suppression effect based on the adaptive impedance reshaping method, when a grid-connected converter operates independently without parameter adaptive adjustment in the control loop, the experimental results are as follows: Figure 6 As shown. By Figure 6 The waveform of the converter output current shows that the harmonic current is amplified, the resonance phenomenon still exists in the converter grid-connected system, and the grid-connected current quality is poor.
[0153] When a grid-connected converter operates independently in a grid-connected manner, and the control loop includes a parameter adaptive component, the experimental results are as follows: Figure 7 As shown. By Figure 7 It can be seen that when the parameters of the resonant compensation circuit are... and Parameters of the current controller After the adaptive parameter adjustment was started, the waveforms of the grid-connected current and grid-connected voltage were significantly improved, and the THD decreased from 6.32% to 2.03%, indicating that the resonance phenomenon in the grid-connected current of the system was effectively suppressed.
[0154] Figure 8A flowchart of a generalized regression neural network (GRNN) is presented. This neural network plays a crucial role in power grid data processing. By reading and collecting the current harmonics from the converter output and the grid voltage harmonics, a database corresponding to harmonics and grid impedance is established. This database allows for the collection of a large amount of relevant data, which can be used as prior knowledge for optimizing prediction data in subsequent stages. The GRNN structure can approximate arbitrarily complex nonlinear function relationships. Through learning from sample data, it can accurately capture these nonlinear relationships, thus providing effective support for prediction and analysis. Furthermore, by introducing expert rules, the model combines expert knowledge in the relevant field with the learning ability of the GRNN, improving the interpretability of data processing and enabling more accurate modeling and prediction of complex data. This enhances the model's ability to handle multivariate and nonlinear data. Simultaneously, by utilizing the stored data in the database to optimize the predicted values, not only is the accuracy of data processing improved, but the precision and computational efficiency of the algorithm are also enhanced.
[0155] Figure 9 The generalized regression neural network architecture is presented. In the input layer, the raw data is received; each node corresponds to one input dimension and performs only signal transmission without complex computation, its role being to "access external information." In the pattern layer, the input is mapped to a new feature space through connection weights with the input layer. In the summation layer, aggregation operations are performed on the output of the pattern layer to further integrate features, preparing for the output layer. In the output layer, based on the result of the summation layer, the final result is output after weight mapping.
[0156] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for suppressing wideband resonance in converters based on impedance identification using a generalized regression neural network, characterized in that, include: A converter system based on inner and outer loop control is constructed. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point. Based on the transfer functions of the inner and outer loops, an adaptive parameter adjustment model for the parameters in the inner and outer loop transfer functions relative to the grid impedance is constructed. A generalized regression neural network for power grid impedance prediction was constructed, and a dataset corresponding to harmonics and power grid impedance was established to train the generalized regression neural network. The grid impedance is predicted in real time using a trained generalized regression neural network; the parameters in the inner and outer loop transfer functions are adjusted by using the predicted grid impedance and adaptive parameter adjustment formula to suppress converter resonance. The inner and outer loops respectively employ closed-loop control and open-loop control. The closed-loop transfer function of the inner loop is: in, and For adaptive adjustment parameters; Adaptive adjustment parameters and The adaptive parameter adjustment model relative to the grid impedance is as follows: in, For converter filter inductance, For the grid-side inductance of the converter, For grid impedance, For filter capacitors, parameters .
2. The converter broadband resonance suppression method based on generalized regression neural network impedance identification according to claim 1, characterized in that, The closed-loop transfer function of the outer loop is: in, To adaptively adjust parameters, To set parameters.
3. The converter broadband resonance suppression method based on generalized regression neural network impedance identification according to claim 2, characterized in that, Adaptive adjustment parameters The adaptive parameter adjustment model relative to the grid impedance is as follows: In the formula, = , For converter filter inductance, For the grid-side inductance of the converter, This represents the power grid impedance.
4. The converter broadband resonance suppression method based on generalized regression neural network impedance identification according to claim 1, characterized in that, A generalized regression neural network for power grid impedance prediction is constructed, and a dataset corresponding to harmonics and power grid impedance is established to train the generalized regression neural network. During the training process, the accuracy of the output value is verified, and the generalized regression theory is introduced to estimate the conditional probability density function through the kernel function method. At the same time, combined with the expert database and expert rules, the optimized power grid impedance prediction value is output.
5. The converter broadband resonance suppression method based on generalized regression neural network impedance identification according to claim 4, characterized in that, The data in the dataset is acquired by installing current sensors on the AC side of the converter and placing voltage sensors at the common coupling point.
6. The converter broadband resonance suppression method based on generalized regression neural network impedance identification according to claim 4, characterized in that, The conditional probability density function is estimated using the kernel function method: the inter-layer transformation of the neural network is represented by a kernel function, and the matrix multiplication of the fully connected layer is replaced by the kernel function to achieve kernelized linear transformation.
7. A converter broadband resonance suppression system implementing the method of any one of claims 1-6, characterized in that, include: The converter system building unit is used to build a converter system based on inner and outer loop control. The converter in the system adopts an LCL filter structure and is connected to the power grid through a common coupling point. An adaptive parameter adjustment model building unit is constructed based on the transfer functions of the inner and outer loops to build an adaptive parameter adjustment model of the parameters in the inner and outer loop transfer functions relative to the grid impedance. The generalized regression neural network construction and training unit constructs a generalized regression neural network for power grid impedance prediction and establishes a dataset corresponding to harmonics and power grid impedance to train the generalized regression neural network. The parameter adjustment unit predicts the grid impedance in real time using a trained generalized regression neural network; it adjusts the parameters in the inner and outer loop transfer functions based on the predicted grid impedance and the adaptive parameter adjustment formula to suppress converter resonance.