A grid-forming grid-connected inverter instability prevention positive design and device

Abstract

The application discloses a network-structured grid-connected inverter instability prevention and positive design method, comprising the following steps: establishing an impedance model and a control loop model of the network-structured grid-connected inverter; defining a multi-target constraint condition and confirming an impedance to be designed based on the impedance model and the control loop model; drawing an impedance curve Bode diagram of each sample and finding out a main difference frequency band, sampling the impedance curve in the frequency band to obtain a labeled impedance frequency response data point set; performing data screening processing on the labeled impedance frequency response point; training the processed labeled data set to obtain a feasible region and a forbidden region of an impedance plane and a decision boundary thereof; and searching and optimizing a surrounding neighborhood of a preliminary feasible sample based on the obtained impedance feasible region and forbidden region to form a parameter feasible region. The application further provides a network-structured grid-connected inverter instability prevention and positive design device. The method provided by the application can realize instability prevention and positive design.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of grid-connected inverter design, and particularly relates to a grid-forming grid-connected inverter instability prevention and positive design and device. BACKGROUND

[0002] With the increasing penetration of new energy generation forms such as photovoltaic and wind power in the power grid, the dynamic characteristics of the power system have undergone profound changes. In the traditional power grid, the system inertia and voltage support are provided by synchronous generators, while the traditional grid-following (GFL) inverter mainly connects to the grid as a current source and lacks active support capability, resulting in a decrease in the stability of the power grid when facing disturbances, i.e., the problem of "weak power grid" is increasingly prominent.

[0003] To solve this problem, grid-forming (GFM) inverter technology has emerged. By simulating the operating mechanism of a synchronous generator, for example, by using virtual synchronous machine (VSG) control, the grid-forming inverter can actively establish the amplitude and frequency of its output voltage, providing key support such as virtual inertia and virtual damping for the power grid, significantly enhancing the stability of the power grid.

[0004] However, the VSG control introduces multiple coupled control parameters, such as virtual rotational inertia J, virtual damping coefficient D, active-frequency loop, reactive-voltage loop, voltage loop, and current loop PI controller parameters, which significantly increase the complexity of system design. These parameters jointly determine the steady-state and dynamic performance of the inverter when operating in parallel, such as cascade stability, loop stability, dynamic response speed, grid inertia and damping characteristics, etc.

[0005] Patent document CN116054233A discloses a grid-forming inverter switching control method with phase support capability under fault, the grid-forming inverter normally works in voltage source control mode, when detecting that the current limit under fault, the angular velocity compensation control link based on the grid phase change is put into operation, the phase difference between the inverter output inner electric potential phase and the actual grid phase is reduced, and the phase support capability is provided at the same time. A PI control link is introduced in the current limiting control module to improve the situation that the inner electric potential amplitude changes sharply due to current limiting under fault.

[0006] Patent document CN120150227A discloses a strong series compensation grid-connected stability regulation method and related device of grid-forming energy storage converter, by using harmonic linearization method to establish complete sequence impedance model of three-phase grid-connected inverter controlled by virtual synchronous generator, and analyzing the reason of sub-synchronous oscillation caused by double closed loop control strategy. SUMMARY

[0007] The application aims to provide a grid-structured grid-connected inverter instability prevention positive design method and device, which can optimize the prediction model to realize instability prevention positive design.

[0008] In order to realize the first object of the application, the following technical scheme is provided: a grid-structured grid-connected inverter instability prevention positive design method, comprising the following steps:

[0009] Based on the grid-structured grid-connected inverter, a corresponding impedance model and a control loop model are constructed, the impedance model comprising an inverter output impedance and a grid impedance;

[0010] Sample data is obtained, comprising control parameters of the grid-structured grid-connected inverter and corresponding target impedance;

[0011] Multi-objective constraint conditions are constructed and brought into the control loop model and the impedance model, and the sample data is screened to divide into feasible sample data and non-feasible sample data meeting the multi-objective constraint conditions;

[0012] Based on the feasible sample data and the non-feasible sample data, a corresponding impedance curve Bode diagram is drawn, and the difference frequency bands in the impedance curve Bode diagram are collected and labeled with sample sources, and the collected difference frequency bands and labels constitute an impedance frequency response data point set;

[0013] Impedance amplitude and phase response points in the difference frequency bands in the impedance frequency response data point set are normalized, and the normalized impedance amplitude and phase response points are respectively taken as dd axis, dq axis, qd axis and qq axis impedance for screening to construct independent data sets corresponding to each impedance;

[0014] The independent data sets corresponding to each impedance are trained by using a deep neural network algorithm to output the feasible region and forbidden area of each impedance plane and the decision boundary thereof;

[0015] Based on the feasible region and forbidden area of all impedance planes and the decision boundary thereof, the neighborhood of the feasible sample data is searched and optimized to output a parameter combination meeting the multi-objective constraint conditions.

[0016] The application quickly trains and generates an impedance feasible region under a specific frequency band based on a small amount of parameter samples, and then quickly searches a parameter sample space by using the trained impedance feasible region to obtain the feasible region of each parameter.

[0017] Specifically, the control loop model comprises an active-frequency loop, a reactive-voltage loop, a voltage control loop and a current control loop.

[0018] Specifically, the impedance frequency response data point set is expressed as follows: ; ;

[0019] wherein, M denotes a set of sampled magnitude response points, P denotes a set of sampled phase response points, i denotes that the sampling point belongs to the i th impedance curve, flag denotes a data label, including feasible and infeasible.

[0020] Specifically, the multi-objective constraint condition includes one or more of a magnitude and phase margin of a grid-connected inverter-weak grid level system, an active-frequency loop magnitude and phase margin, a reactive-voltage loop magnitude and phase margin, a voltage loop magnitude and phase margin, a current loop magnitude and phase margin, a voltage loop bandwidth, a current loop bandwidth, virtual inertia, and virtual damping.

[0021] Specifically, the construction process of the independent data set is as follows:

[0022] Determine the algorithm initialization parameters, and normalize the impedance magnitude and phase response points respectively to unify the scales of the coordinates of each dimension.

[0023] Design the target axis impedance: regarding each frequency response point on the magnitude plane and the phase plane as a newly placed data point, find K the nearest sample point and count the number of feasible labels as K 1, the number of infeasible labels as K 2, and determine the label attribution of the newly placed data point according to the following formula: ; In the formula, the labels 1, 0, and -1 represent feasible, infeasible, and pending labels, respectively, K err denotes the allowed error; count the number of magnitude response points and the number of frequency response points defined as infeasible labels in each parameter sample curve, when the number of magnitude response points or the number of frequency response points meets the threshold, then remove the points defined as pending labels in the corresponding parameter sample impedance curve, and update the total magnitude response point data set and the phase data set, reduce K and increase K err , until K is reduced to meet the iteration termination condition, to obtain the processed independent data set.

[0024] Specifically, the deep neural network includes a support vector machine for generating a binary classification boundary and a Gaussian process regression joint model for predicting a feasibility probability distribution of an unlabeled region.

[0025] Specifically, the specific process of the neighborhood search optimization is as follows:

[0026] With a feasible parameter sample in the feasible sample data as the center, sample points are generated at a fixed interval in the dimension of each control parameter for parameter feasible region expansion search;

[0027] The sampling of the impedance frequency response in a specific frequency band is performed for each expanded sample point, and the obtained impedance plane feasible region / forbidden region is used to determine whether the expanded sample point meets the multi-objective constraint condition;

[0028] When the parameter sample point used for search in a certain dimension direction is a non-feasible point, or there is a feasible or non-feasible parameter sample near the sample point used for search, the expansion search in the dimension direction is stopped, and the feasible search point expanded to the farthest distance is taken as the boundary of the dimension;

[0029] The farthest feasible sample points used for search in each dimension are connected to each other to form a closed parameter feasible region;

[0030] The above process is repeated until the field search of each feasible parameter sample in the feasible sample data is completed.

[0031] In order to realize the second object of the application, the technical scheme is provided as follows: a grid-connected inverter instability prevention and positive design device, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the processor executes the computer program to execute the steps of the grid-connected inverter instability prevention and positive design method.

[0032] Compared with the prior art, the application has the following beneficial effects:

[0033] (1) A grid-connected inverter parameter design method considering multiple target constraints such as cascade stability, dynamic performance of multiple control loops, inertia and damping requirements is proposed;

[0034] (2) A small number of parameter sample points are used for rapid training to generate impedance-based visual feasible region and forbidden region, and then a data-driven method is used to guide subsequent design;

[0035] (3) The traditional theoretical analysis design is avoided, and the design process is more simple. DETAILED DESCRIPTION

[0036] Figure 1 A schematic diagram of a grid-connected inverter instability prevention and positive design method provided by the embodiment;

[0037] Figure 2This is a schematic diagram of the control strategy for the virtual synchronous grid-connected inverter provided in this embodiment;

[0038] Figure 3 The flowchart of the improved K-nearest neighbor algorithm provided in this embodiment;

[0039] Figure 4 The flowchart of the support vector machine and Gaussian regression process provided in this embodiment. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0041] like Figure 1 The diagram shown is a schematic of a forward design method for preventing instability in a grid-connected inverter provided in this embodiment. The specific process is as follows:

[0042] Establish the impedance model and control loop models of the grid-connected inverter;

[0043] More specifically, the impedance model includes the inverter output impedance. Z o and grid impedance Z g ;

[0044] like Figure 2 As shown, this embodiment uses a virtual synchronous grid-connected inverter control as an example. The control loop model includes an active-frequency loop, a reactive-voltage loop, a voltage control loop, and a current control loop. The active-frequency loop and the reactive-voltage loop generate reference values ​​for voltage and phase angle, and can also simulate the inertia characteristics of a synchronous generator. Then, the voltage, current, and power output by the grid-connected inverter are controlled by the voltage and current dual loops.

[0045] in v abc and i abc These are the three-phase grid-connected voltage and current, respectively. P 0 and Q0 represents the active power and reactive power of the inverter, respectively. In the active-frequency control loop, P ref This is a reference value for active power. ω 0 represents the theoretical power grid angular frequency. J and D p These are the virtual inertia coefficient and the virtual damping coefficient, respectively. s For the Laplace operator, ω and θ These are the actual grid angular frequency and grid phase angle, respectively; in the reactive power-voltage control loop, Q ref This is a reference value for reactive power. K q This is the primary voltage regulation coefficient. E m This is the rated voltage amplitude of the power grid. E dref The no-load electromotive force amplitude is output to the virtual synchronous machine inverter; in the voltage-current dual loop, v d and v q These are the grid connection point voltages in the dq coordinate system, respectively. i Labc Let be the filter inductor current in the abc coordinate system. i Ld and i Lq These are the filter inductor currents in the dq coordinate system, respectively. c d and c q These represent the duty cycles in the dq coordinate system. c abc The modulated wave is in the abc coordinate system. G v ( s )and G i ( s These are voltage regulators and current regulators, respectively.

[0046] Based on the impedance model and control loop model, multi-objective constraints are defined, and the impedance to be designed is confirmed. A small number of parameter sample points are generated by meshing within the domain of control parameters. Based on the theoretical model analysis, it is determined whether each sample satisfies the multi-objective constraints, and preliminary feasible / infeasible parameter samples are obtained.

[0047] More specifically, the impedance to be designed includes the elements of the 2×2 matrix of the dq domain impedance model. Z dd , Z dq ,Z qd 、 Z qq 。

[0048] Draw the impedance curve Bode diagram of each sample and find out the main difference frequency band, sample the impedance curve in the frequency band to obtain the labeled impedance frequency response data point set, and the obtained frequency response data point set is as follows: ; ; Among them, M and P respectively represent the sampled amplitude response point data set and the phase response point data set, i represents that the point belongs to the i th impedance curve, flag represents the feasible / infeasible label.

[0049] The improved K-neighbor (K-NN) algorithm is used for data screening processing of the impedance frequency response points in the above impedance frequency response data point set, and the problem of serious mixing of two types of label data is improved. In this embodiment, the flow as shown in Figure 3 is referred to, wherein, K is the number of selected adjacent data, since the progressive data classification method with iterative decreasing K value is used, there is no need to strictly design K , and a larger K can be taken at the beginning to ensure the feasibility of the algorithm; K err is the number of allowed errors, and the initial K err should be as small as possible to achieve more stringent classification, but can be increased slightly as the iteration enters the later stage; K is the iterative decreasing K value, K min is the iteration end value, and the two can be designed according to the accuracy requirement and screening effect. For example, if the accuracy needs to be improved, a smaller K (such as 1~3) can be selected, otherwise K can be larger (such as 4 or higher), if there are still significant data samples that cannot be classified after the iteration is completed, the K min should be further reduced to further classify; n_judge is the threshold for curve data rejection, which is related to the data point sampling density of each curve, and the specific steps are as follows:

[0050] Determine the algorithm initialization parameters, and normalize the impedance amplitude and phase response points respectively to unify the scale of each dimension coordinate;

[0051] respectively dd axis, dq axis, qd axis and qq axis impedance, taking each frequency response point of the amplitude plane and the phase plane as a new added data point, finding K the nearest sample point and counting the number of feasible labels as K 1, the number of non-feasible labels is K 2, and the label attribution of the new added data point is determined according to the following formula: ; In the formula, the labels 1, 0 and -1 respectively represent feasible, non-feasible and pending labels, K err , wherein if the point is defined as a feasible or non-feasible label, the point will not be classified again in the subsequent iteration process; if the point is defined as a pending label, the point will continue to be classified in the subsequent iteration process; the number of amplitude response points defined as non-feasible labels in each parameter sample curve is counted n_mag j and the number of frequency response points is counted n_ pha j When n_mag j or n_pha j at least one reaches the threshold value n_judge , the points defined as pending labels in the parameter sample impedance curve are removed, and the total amplitude and phase data set is updated;

[0052] is reduced K and K err is slightly increased. dd axis, dq axis, qd axis and qq axis impedance is screened to construct an independent data set corresponding to each impedance.

[0053] The deep neural network (DNN) algorithm is used to train each independent data set to obtain the feasible region and forbidden zone of the impedance plane and the decision boundary thereof;

[0054] As shown in Figure 4 , the flowchart of the deep neural network (DNN) algorithm provided in the embodiment is combined with the support vector machine (SVM) and Gaussian process regression (GPR) algorithm, and the implementation steps are as follows:

[0055] The amplitude and phase data of the impedance frequency response point are mapped into a feature vector xi and label them with binary tags. y i ∈{0,1} (0 indicates infeasibility, 1 indicates feasibility), and normalize the result;

[0056] The Z-score method is used for normalization, and its normalization function is: ,in μ The mean, σ Standard deviation;

[0057] An SVM model is constructed using a radial basis function kernel (RBF), whose kernel function is: ;

[0058] Optimize hyperparameters using grid search and cross-validation. γ and penalty coefficient C Generate preliminary decision boundaries;

[0059] Based on the SVM classification results, sample points near the decision boundary are... x k A Gaussian process regression model is constructed, using the Marting kernel as the kernel function, and its expression is as follows:

[0060] ; in, K ( x i , x j ) is the input point x i and x j The covariance between them, Γ( v ) is the gamma function. K v Is the order of v The second kind of modified Bessel function, || x i - x j ||Yes x i and x j The Euclidean distance between them l It is a positive length scaling parameter used to control the rate at which correlation decays with distance. v It is a positive smoothness parameter that controls the smoothness of the function.

[0061] Calculate the feasibility probability of unlabeled areas p ( y =1| x ),likep y x if the condition of |f (x) - f (x)| ≥ 0.5 is satisfied, the region is determined as a feasible region;

[0062] According to the probability distribution output by the GPR, the decision boundary of the SVM is locally smoothed to generate a final impedance feasible region.

[0063] Based on the obtained impedance feasible region and forbidden region, the neighborhood of the preliminary feasible sample is searched and optimized to expand the parameter combination satisfying the multi-objective constraints, and finally a parameter feasible region is formed;

[0064] More specifically, the neighborhood search step is as follows:

[0065] Taking a certain preliminary feasible sample point as the center, sample points are generated at regular intervals around the center (in each control parameter dimension) for parameter feasible region expansion search;

[0066] For each sample point, the impedance frequency response in a specific frequency band is sampled, and the obtained impedance plane feasible region / forbidden region is used to determine whether the sample point satisfies the proposed steady-state and dynamic constraints;

[0067] When the sample point used for search in a certain direction is a non-feasible point, or there are other preliminary feasible / non-feasible parameter samples near the sample point used for search, the expansion search in that direction is stopped, and the farthest feasible search point is taken as the boundary in that dimension;

[0068] The farthest feasible sample points used for search in each direction are connected to each other to form a closed parameter feasible region;

[0069] The above steps are repeated until the field of each preliminary feasible sample point is searched, and a final parameter feasible region is formed.

[0070] The embodiment also provides a grid-connected inverter instability prevention and positive design device, which comprises a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, and the processor executes the computer program to execute the steps of the grid-connected inverter instability prevention and positive design method provided in the above embodiment.

[0071] In addition, the terms "upper", "lower", "inner", "outer", "front", "back" are only used for description purposes, and cannot be understood as indicating or implying relative importance. Unless otherwise specified, the relative steps, numerical expressions and numerical values of the components and steps set forth in the embodiments do not limit the scope of the present application.

[0072] ​​Of course, the above only is the specific embodiment of the present application, and does not limit the scope of the present application, and equivalent changes or modifications made according to the structure, features and principles described in the patent application scope of the present application should be included in the patent application scope of the present application.

[0073] Finally, it should be noted that: the above-described embodiments, only for the specific embodiments of the present application, to illustrate the technical solutions of the present application, and not limited, the protection scope of the present application is not limited to this, although the foregoing detailed description of the present application, those skilled in the art should understand: any familiar with the technical field of the technical personnel in the technical range disclosed by the present application, it still can be modified or easily thought of changes to the technical solutions recorded in the foregoing examples, or equivalent replacement of some technical features; and these modifications, changes or replacement, and the corresponding technical solutions of the spirit and scope of the present application embodiment technical solutions, all should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be said to the scope of the claims for the protection of the present application.

[0074] The present application does not depend on the experience of experts, greatly saves manpower, and the idea of machine learning and digital-analog hybrid driving helps to simplify the design process of grid-connected inverter system, provides an effective design idea, and has great engineering application value and popularization prospect.

Claims

1. A forward design method for instability prevention of grid-connected inverters, characterized in that, Includes the following steps: Based on the grid-connected inverter, a corresponding impedance model and control loop model are constructed. The impedance model includes the inverter output impedance and the grid impedance. Obtain sample data, including the control parameters and corresponding target impedance of the grid-connected inverter; Multi-objective constraints are constructed and incorporated into the control loop model and impedance model. The sample data is then filtered to distinguish between feasible and infeasible sample data that satisfy the multi-objective constraints. Based on the feasible sample data and the infeasible sample data, the corresponding impedance curve Bode plot is drawn, and the difference frequency bands in the impedance curve Bode plot are collected and labeled with the sample source. The collected difference frequency bands and labels are combined to form an impedance frequency response data point set. The impedance amplitude and phase response points in the differential frequency bands of the impedance frequency response data set are normalized, and then respectively... dd axis, dq axis, qd shaft and qq The axial impedance is filtered to construct an independent dataset corresponding to each impedance. A deep neural network algorithm is then used to train the independent datasets corresponding to each impedance to output the feasible region and forbidden region of each impedance plane, as well as their decision boundaries. The deep neural network includes a support vector machine for generating binary classification boundaries and a Gaussian process regression joint model for predicting the feasibility probability distribution of unlabeled regions. The implementation steps are as follows: Mapping the amplitude and phase data of the impedance frequency response point into an eigenvector x i and label them with binary tags. y i ∈{0,1}; Construct an SVM model using radial basis function kernels; Optimize hyperparameters using grid search and cross-validation. γ and penalty coefficient C Generate preliminary decision boundaries; Based on the SVM classification results, sample points near the decision boundary are... x k Construct a Gaussian process regression model, using the Marton kernel as the kernel function; calculate the feasibility probability of the unlabeled region. p ( y =1| x ),like p ( y =1| x If the value is ≥0.5, it is considered a feasible region; Based on the probability distribution of the GPR output, the decision boundary of the SVM is locally smoothed to generate the final impedance feasible region. Based on the feasible and forbidden regions and decision boundaries of all impedance planes, a neighborhood search optimization is performed on the surrounding neighborhood of the feasible sample data to output a parameter combination that satisfies the multi-objective constraints. The specific process of the neighborhood search optimization is as follows: Taking a feasible parameter sample in the feasible sample data as the center, expand sample points are generated at fixed intervals in each dimension of the control parameter for parameter feasible domain expansion search. For each extended sample point, the impedance frequency response within a specific frequency band is sampled, and the obtained feasible / no-go region of the impedance plane is used to determine whether the extended sample point satisfies the multi-objective constraint conditions. When the parameter sample point used for searching in a certain dimension is an infeasible point, or when there are feasible or infeasible parameter samples near the sample point used for searching, the expansion search in that dimension is stopped, and the feasible search point that has been expanded to the farthest distance is taken as the boundary of that dimension. Connect the furthest feasible sample points used for the search in each dimension to form a closed parameter feasible region; Repeat the above process until a neighborhood search is completed for each feasible parameter sample in the feasible sample data.

2. The forward design method for instability prevention of grid-connected inverters according to claim 1, characterized in that, The control loop model includes an active-frequency loop, a reactive-voltage loop, a voltage control loop, and a current control loop.

3. The forward design method for instability prevention of grid-connected inverters according to claim 1, characterized in that, The impedance frequency response data point set is expressed as follows: ； ； in, M This represents a dataset of sampled amplitude response points. P This represents the sampled phase response point dataset. i Indicates that the collection point belongs to the first... i impedance curves, flag This indicates data labels, including feasible and infeasible.

4. The forward design method for instability prevention of grid-connected inverters according to claim 1, characterized in that, The multi-objective constraints include one or more of the following: amplitude and phase margin of the grid-connected inverter-weak grid cascade system; amplitude and phase margin of the active-frequency loop; amplitude and phase margin of the reactive-voltage loop; amplitude and phase margin of the voltage loop; amplitude and phase margin of the current loop; voltage loop bandwidth; current loop bandwidth; virtual inertia; and virtual damping.

5. The forward design method for instability prevention of grid-connected inverters according to claim 1, characterized in that, The process of constructing the independent dataset is as follows: Determine the algorithm initialization parameters and normalize the impedance amplitude and phase response points respectively to unify the scale of the coordinates in each dimension; Design for the target's axial impedance: Treat each frequency response point in the amplitude and phase planes as a newly added data point, and find... K Find the nearest sample points and count the number of feasible labels for each of them. K 1. The number of non-feasible tags is K 2. Determine the label assignment for newly added data points using the following formula: In the formula, labels 1, 0, and -1 represent feasible, infeasible, and pending labels, respectively. K err This indicates the allowable error; the number of amplitude response points and frequency response points defined as non-feasible labels in each parameter sample curve is counted. When the number of amplitude response points or frequency response points meets the threshold, the points defined as undetermined labels in the corresponding parameter sample impedance curve are removed, and the total amplitude response point dataset and phase dataset are updated to reduce... K and increase K err until K Reduce the size until the iteration termination condition is met to obtain the processed independent dataset.

6. A forward design device for instability prevention of a grid-connected inverter, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that, When the processor executes a computer program, it performs the steps of the forward design method for preventing instability in grid-connected inverters as described in any one of claims 1 to 5.

Images