High-efficiency data trajectory driven control method for grid-connected inverter oscillation suppression

Abstract

The application discloses a kind of high-efficiency solution data trajectory driving control methods of grid-connected inverter oscillation suppression.The active power and reactive power of inverter output are used as the input of controller, and the additional current reference value is used as the output of controller, to construct a model-free predictive controller; the input and output sequence of system in historical period is obtained to construct offline data matrix, and the low-dimensional offline data matrix is obtained by order reduction to construct a predictive control optimization function; the input in current period is obtained, and the additional current reference value is obtained by solving the optimization function by controller; the additional current reference value and the current reference value are added and input to the current inner loop for processing, and then the inverter is controlled.The application extracts the dominant dynamic characteristics of the system, solves the problem of low efficiency in solving optimization problems caused by using high-dimensional data matrix; without accurate physical model, it can quickly respond and effectively suppress the system oscillation caused by phase-locked loop interaction or model mismatch in weak grid environment.

Description

Technical Field

[0001] This invention relates to the field of grid-connected control technology for power electronic converters, and in particular to an efficient solution-driven control method for oscillation suppression in grid-connected inverters. Background Technology

[0002] In recent years, the installed capacity of new energy sources, represented by wind power and photovoltaics, has continued to rise, and the power system is undergoing a fundamental transformation from being dominated by synchronous generators to being dominated by power electronic devices. This transformation has significantly changed the physical characteristics of the power grid, gradually exhibiting typical "weak grid" characteristics such as high impedance, low inertia, and low short circuit ratio (SCR).

[0003] Under the current technological framework for integrating new energy sources into the grid, grid-connected power electronic equipment such as photovoltaic inverters and wind power converters typically employ a grid-following (GFL) control method relying on phase-locked loops (PLLs). However, in scenarios with low grid strength, the voltage at the point of common coupling is easily affected by grid-connected current disturbances, resulting in significant changes. Since the PLL needs to achieve phase synchronization based on the port voltage, voltage disturbances caused by current control will further enter the synchronization and control stages. When there is poor coupling between the equivalent output impedance of the converter and the larger equivalent impedance of the weak grid, the system may generate an enhanced feedback effect, thereby inducing insufficient damping or even oscillation instability. Such small-disturbance stability problems, primarily caused by PLL dynamics, can, in mild cases, cause the device to shut down due to current or voltage exceeding limits, and in severe cases, may develop into multi-frequency oscillations, threatening the stability of the local grid and even a larger power system.

[0004] To address the aforementioned issues, many grid-connected inverters currently employ traditional PI controllers with dual-loop voltage and current control structures. Their control parameters are often based on simplified models such as idealized power grid assumptions and single-machine infinite loops, or designed for a specific rated operating point. However, real power systems exhibit significant time-varying, nonlinear characteristics and complex coupling relationships. Factors such as line structure, load distribution, and short-circuit capacity can continuously change with operating conditions. Traditional fixed-parameter control methods lack online identification and self-adjustment capabilities to changes in external power grid conditions. When there is a significant deviation between the actual system and the design model, not only is it difficult to guarantee stability margins, but it may also further amplify system oscillation risks. Although adaptive control and robust control methods based on accurate models can improve this problem to some extent, these methods typically require the establishment of relatively accurate mathematical models and do not adequately consider factors such as power device dead-zone effects, digital sampling and computation delays, and unmodeled dynamics on the grid side. Therefore, their application in complex real-world conditions remains limited.

[0005] Model-free predictive control (MDR) is generally considered a class of advanced control methods that weakens mechanistic model dependence. Its basic idea is to use input and output data collected during system operation to describe the dynamic characteristics of the system and generate control commands within a predictive optimization framework. Compared to traditional model predictive control, this method does not require pre-establishing a complete and accurate state-space model or transfer function model, and can reduce the impact of parameter identification errors, line impedance uncertainties, load fluctuations, and changes in grid strength on control performance to a certain extent. Therefore, model-free predictive control has good application potential in the operation of grid-connected inverters in weak grids, oscillation suppression, and adaptation to complex operating conditions.

[0006] However, existing model-free predictive control methods still have certain limitations in the engineering application of oscillation suppression in grid-connected inverters. On the one hand, although some methods do not directly rely on the complete analytical model of the inverter, they still need to introduce some system parameters, such as filter inductance, line impedance, grid equivalent impedance, grid connection point short-circuit ratio, operating point information, or local linearization coefficients, during the construction of predictive relationships, control gain tuning, constraint boundary determination, or dynamic response correction. Therefore, these methods have not completely eliminated their dependence on system models and parameters. When grid impedance, grid connection point strength, topology, or operating conditions change, these parameters are prone to deviation, leading to inconsistencies between the controller's prediction results and the actual dynamic response, thereby weakening the oscillation suppression effect and even affecting system stability. On the other hand, while model-free predictive control methods based on input-output data can avoid dependence on the analytical model of the inverter and system parameters, their predictive models usually need to be constructed from a large amount of historical data. As the data window length, input / output dimensions, and prediction time domain increase, the size of the data matrix grows rapidly, making the online rolling optimization process face high computational complexity and solution burden, making it difficult to meet the control requirements of grid-connected inverters for rapid dynamic adjustment and real-time oscillation suppression. Summary of the Invention

[0007] To address the challenges posed by dynamic coupling of phase-locked loops (PLLs) in weak grid conditions, which can easily induce oscillations in grid-connected inverters, the insufficient adaptability of traditional fixed-parameter control to changes in grid operating conditions, and the large scale and significant real-time solution pressure of conventional model-free predictive control in online optimization, this invention proposes an efficient solution data trajectory-driven control method for suppressing grid-connected inverter oscillations.

[0008] The technical solution adopted in this invention is: The method of the present invention includes the following steps: S1. Obtain the three-phase voltage and three-phase current at the inverter's grid connection point, and then calculate the active power and reactive power output by the inverter. S2. The active power and reactive power output by the inverter are used as the controller input, and the components of the additional current reference value on the d-axis and q-axis are used as the controller output, thereby constructing a model-free predictive controller. The model-free predictive controller includes an offline preparation stage and an online operation stage. S3. In the offline preparation stage, the system input and output sequence of a preset historical period is obtained to construct an offline data matrix. The offline data matrix is ​​reduced to a low-dimensional offline data matrix, and then a predictive control optimization function is constructed based on the low-dimensional offline data matrix. S4. During the online operation phase, the active power and reactive power output of the inverter in the current time period are obtained and input into the modelless predictive controller. The modelless predictive controller solves the predictive control optimization function to obtain the additional current reference value. S5. Add the additional current reference value to the current reference value to obtain the reference value of the inner current loop, input it to the inner current loop for processing to obtain the modulation wave voltage amplitude signal, generate a drive signal based on the modulation wave voltage amplitude signal and control the inverter.

[0009] Specifically, S3 is: S3.1. The active power and reactive power output by the inverter are used as the system output, and the components of the additional current reference value on the d-axis and q-axis are used as the system input. S3.2 Apply white noise excitation to the system, and then collect the active power, reactive power, and additional current reference values ​​of the inverter output in the d-axis and q-axis for a first preset time length, and then construct the system input-output sequence in time order; S3.3 Construct an offline data matrix based on the system input and output sequence, perform singular value decomposition on the offline data matrix to determine the order of the low-dimensional offline data matrix, and then construct the low-dimensional offline data matrix based on the order of the low-dimensional offline data matrix. S3.4 Construct a predictive control optimization function based on the low-dimensional offline data matrix.

[0010] The order of the low-dimensional offline data matrix is ​​determined according to the following steps: 1) Determine if measurement noise exists in the offline data matrix: If it exists, then the full-rank dimension of the offline data matrix is ​​used as the order of the low-dimensional offline data matrix; If it does not exist, the singular value decomposition threshold is determined based on whether the noise level of the data in the offline data matrix is ​​known, and then the number of singular values ​​that are greater than the singular value decomposition threshold after the offline data matrix is ​​decomposed is taken as the first candidate order. 2) Calculate the cumulative energy ratio based on the singular values ​​obtained from the singular decomposition of the offline data matrix, and take the minimum number of singular values ​​that satisfy the cumulative energy ratio being greater than or equal to the preset energy retention coefficient as the second candidate order; 3) Compare the first candidate order with the second candidate order, and take the smaller value of the two as the comprehensive candidate order. Then compare the comprehensive candidate order with the preset minimum order for maintaining closed-loop controllability, and take the larger value of the two as the order of the low-dimensional offline data matrix.

[0011] When the standard deviation of offline data noise is known, the singular value decomposition threshold is determined according to the following formula:

[0012]

[0013] When the noise standard deviation of the offline data is unknown, the singular value decomposition threshold is determined according to the following formula:

[0014]

[0015] in, This represents the singular value decomposition threshold. The shape factor represents the shape of the offline data matrix. , and These represent the lengths of the input and output sequences, the prediction length of the model-free predictive controller, and the observation length, respectively. This represents the standard deviation of noise in offline data. This represents the optimal hard threshold coefficient when the noise level is known. This represents the optimal hard threshold coefficient when the noise level is unknown. This represents the median of the singular values ​​obtained from the singular decomposition of the offline data matrix.

[0016] The cumulative energy ratio is set according to the following formula:

[0017] in, Let represent the proportion of the cumulative energy of the first r singular values ​​arranged in descending order to the total energy; let i represent the index number of the singular values ​​after arranging them in descending order; let m represent the total number of singular values; and let r represent the number of singular values ​​currently retained. This represents the i-th singular value obtained after performing singular value decomposition on the offline data matrix.

[0018] The predictive control optimization function is set according to the following formula: ;

[0019] ; in, Represents the predictive control optimization function. and Let represent the lengths of the input-output sequences and the prediction length of the model-free predictive controller, respectively. Let Q, R, and S represent the input, output, and control smoothing weight matrices, respectively. Represents the first term in the prediction time domain. One prediction step, This indicates that at time k, the future... The predicted value output by the step system. This represents the output reference value for the corresponding prediction step. This indicates the future time predicted at time k. Additional current reference value, Represents the regularization coefficient. Represents the regularization term, The coefficients of the slack variable. Represents slack variables. , , and These represent the submatrices representing historical input, historical output, future input, and future output within the low-dimensional data matrix, respectively. Represents the most recent time before time k. An additional current reference increment Represents the most recent time before time k. One measurement output, and These represent the lower and upper limits of the additional current reference value, respectively. and These represent the lower and upper limits of the system output, respectively. This represents the column vector concatenation operator. This represents the square of the weighted L2 norm.

[0020] The model-free predictive controller employs a rolling optimization method. After the low-dimensional offline data matrix is ​​initially constructed, it remains fixed. During the online operation phase, the active and reactive power output of the inverter at the current moment is collected in each control cycle. The predictive control optimization problem is solved to obtain the additional current reference value for the current control cycle.

[0021] The beneficial effects of this invention are: This invention characterizes the input-output dynamics of a grid-connected inverter within its current operating range using an offline constructed data matrix, eliminating the need for detailed converter parameter models and reducing the impact of model mismatch on oscillation suppression. Through a reduction-order process combining singular value thresholding and energy verification, the main dynamic components in the high-dimensional data matrix are preserved while significantly reducing the dimensionality of online optimization variables, thus improving real-time solution speed. Under weak grid conditions or varying operating points, grid-connected inverters are prone to persistent oscillations. The model-free predictive controller can continuously adjust the additional current reference value based on the system's real-time input-output response, providing coordinated support for grid connection voltage, reactive power, active power, and DC bus voltage. This enhances the controller's ability to suppress persistent oscillations and improves dynamic adjustment performance under complex operating conditions.

[0022] This invention is applicable to various types of grid-connected inverters. The method introduces singular value decomposition to reduce the dimensionality of high-dimensional data matrices, extracting the dominant dynamic features of the system and eliminating redundancy, thereby solving the problem of low efficiency in solving optimization problems caused by using high-dimensional data matrices. While significantly improving computational efficiency, this technology does not require an accurate physical model, is robust to parameter changes, and can quickly respond to and effectively suppress system oscillations caused by phase-locked loop interactions or model mismatches in weak grid environments. It achieves adaptive and stable operation of grid-connected inverters under complex and variable operating conditions, providing key technical support for building high-proportion renewable energy power systems that combines high performance and engineering practicality. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method of this embodiment.

[0024] Figure 2 This is a flowchart illustrating the method for selecting the dimensionality reduction threshold in this embodiment.

[0025] Figure 3 This is a control block diagram for this embodiment.

[0026] Figure 4 This is a schematic diagram of the simulation model of the grid-connected converter operating system according to an embodiment of the present invention.

[0027] Figure 5 The waveform diagram shows the reactive power output of the converter in this embodiment.

[0028] Figure 6 This is a waveform diagram of the active power output of the converter in this embodiment.

[0029] Figure 7 This is a comparison chart of the optimization problem solution time before and after dimensionality reduction in this embodiment. Detailed Implementation

[0030] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0032] The control section of this embodiment employs a dual closed-loop structure consisting of a power outer loop and a current inner loop. The outer loop uses a PI regulator to track and control active and reactive power, thereby generating the current setpoint required by the inner loop. The inner loop also uses a PI algorithm to regulate the current. Based on this, a model-free predictive controller is introduced. This controller calculates the optimal control increment using historical input / output data and injects it into the control loop to improve dynamic performance. After setting the power reference value and the parameters of each controller, the voltage command output from the current inner loop is synthesized into a modulated wave through inverse Park transform. Finally, a drive pulse is generated through space vector pulse width modulation (SVPWM) to achieve data trajectory drive control of the grid-connected inverter. A detailed control logic block diagram is shown below. Figure 3 As shown.

[0033] like Figure 1 As shown, this embodiment includes the following steps: S1. Obtain the three-phase voltage and three-phase current at the inverter's grid connection point, and then calculate the active power and reactive power output by the inverter. Before using a model-free predictive controller, it is necessary to first acquire the port voltage signal of the grid-connected inverter using a voltage transformer and then lock the system frequency using a phase-locked loop (PLL). The phase angle can be obtained by integrating the frequency. This angle will serve as the reference phase for the Park transformation and subsequent modulation stages. Simultaneously, the inverter output current is monitored in real time via a current transformer, and decomposed into components in the dq rotating coordinate system through coordinate transformation. The system then uses the acquired voltage... With current Data, real-time calculation of current output active power With reactive power .

[0034] The present invention describes the process of acquiring inverter electrical quantities and calculating its output active and reactive power based on the acquired voltage and current information. The specific steps include: In the appendix Figure 3 In this context, the electrical parameters are defined as follows: Represents the filter inductor on the converter side; This is the sum of the network test filter inductance and the transmission line inductance; and The real-time values ​​of the output voltage and output current of the strain gauge were measured respectively; This represents the line current value on the power grid side. This indicates the modulation wave voltage command used to drive SVPWM, and , These represent the active and reactive power outputs of the converter, respectively.

[0035] A voltage sensor is installed at the grid connection point of the converter to collect the three-phase voltage. A current sensor is installed on the output side of the filter inductor to collect the three-phase current. Simultaneously, the DC-side capacitor voltage is collected. The phase-locked loop uses the grid connection point voltage as its input and output synchronization phase. and angular frequency .

[0036] For any three-phase quantity Convert to synchronous rotating coordinate components using the following formula:

[0037] Thus, the grid connection point voltage is obtained. , and output current , PLL adjusts Phase locking is completed by approaching zero, and with As a phase reference for subsequent modulation waves The calculation formula is:

[0038] The real-time power output of the inverter to the grid in a three-phase three-wire system can be calculated using the following formula:

[0039] S2. The active power and reactive power output by the inverter are used as the controller input, and the components of the additional current reference value on the d-axis and q-axis are used as the controller output, thereby constructing a model-free predictive controller. The model-free predictive controller includes an offline preparation stage and an online operation stage. That is, the input and output quantities of the model-free predictive controller are set, and then the controller data matrix is ​​constructed offline using a noisy excitation system, specifically: Define controller input and output variables: Define the additional current reference value as the controller output. The above. and They are respectively in k Additional current reference value output by the time controller.

[0040] Define the inverter's output active power and reactive power as the controller's input: The above. , These represent the active and reactive power outputs of the inverter at time k, respectively.

[0041] S3. In the offline preparation stage, the system input and output sequence of a preset historical period is obtained to construct an offline data matrix. The offline data matrix is ​​reduced to a low-dimensional offline data matrix, and then a predictive control optimization function is constructed based on the low-dimensional offline data matrix. Specifically, S3 is: S3.1. The active power and reactive power output by the inverter are used as the system output, and the components of the additional current reference value on the d-axis and q-axis are used as the system input. S3.2 Apply white noise excitation to the system, and then collect data for a first preset time length. The active power, reactive power, and additional current reference values ​​of the inverter output are calculated in terms of their components on the d-axis and q-axis, and then the system input-output sequence is constructed in time order. S3.3 Construct an offline data matrix based on the system input and output sequence, perform singular value decomposition on the offline data matrix to determine the order of the low-dimensional offline data matrix, and then construct the low-dimensional offline data matrix based on the order of the low-dimensional offline data matrix. Let the observation length of the model-free predictive controller be... The predicted length is Before the model-free predictive controller is put into closed-loop operation, a white noise excitation signal is injected into the additional current reference channel. The sampling period is consistent with the control period, and the recording length is [missing information]. Input / output sequence: , , ..., . Need to meet .

[0042] The offline data is truncated in chronological order to obtain several data segments containing past system information and future information. For the first... Given a data segment, define its past input, past output, predicted input, and predicted output as follows:

[0043] Combine the above four types of data into a data column in sequence:

[0044] Then, all data columns are arranged in order to form an offline data matrix:

[0045] like Figure 2 As shown, the order of the low-dimensional offline data matrix is ​​determined according to the following steps: 1) Determine if measurement noise exists in the offline data matrix: If it exists, then the full-rank dimension of the offline data matrix is ​​used as the order of the low-dimensional offline data matrix; If it does not exist, the singular value decomposition threshold is determined based on whether the noise level of the data in the offline data matrix is ​​known. Then, the number of singular values ​​in the offline data matrix that are greater than the singular value decomposition threshold is used as the first candidate order. ; In order to ensure that the data matrix can cover as many of the system's dynamic characteristics as possible, the number of columns in the constructed data matrix is... Much larger than its number of rows Data matrix The rank simultaneously satisfies Therefore, the data matrix rank .

[0046] If the data acquired by the sensor contains measurement noise, the data matrix constructed using this noisy data will often be full rank. Therefore, in this case, the dimensionality of the data matrix can be reduced. .

[0047] If the data acquired by the sensor does not contain measurement noise, then the dimension to be reduced needs to be selected by setting a singular value threshold. Define the data matrix. shape factor for:

[0048] When the offline data noise standard deviation Given the information, set a threshold for singular value decomposition. for:

[0049] When the noise level of offline data is unknown, the median of singular values ​​can be used. Estimated threshold :

[0050] in, This represents the singular value decomposition threshold. The shape factor represents the shape of the offline data matrix. , and These represent the lengths of the input and output sequences, the prediction length of the model-free predictive controller, and the observation length, respectively. This represents the standard deviation of noise in offline data. This represents the optimal hard threshold coefficient when the noise level is known. This represents the optimal hard threshold coefficient when the noise level is unknown. This represents the median of the singular values ​​obtained from the singular decomposition of the offline data matrix.

[0051] 2) Calculate the cumulative energy ratio based on the singular values ​​obtained from the singular decomposition of the offline data matrix, and select the minimum number of singular values ​​that satisfy the cumulative energy ratio being greater than or equal to the preset energy retention coefficient as the second candidate order. ; Further, an energy preservation check is introduced. Let the matrix be... Total A singular value The cumulative energy ratio is defined as:

[0052] in, Let represent the proportion of the cumulative energy of the first r singular values ​​arranged in descending order to the total energy; let i represent the index number of the singular values ​​after arranging them in descending order; let m represent the total number of singular values; and let r represent the number of singular values ​​currently retained. This represents the i-th singular value obtained after performing singular value decomposition on the offline data matrix.

[0053] 3) Compare the first candidate order with the second candidate order, and take the smaller value of the two as the comprehensive candidate order. Then compare the comprehensive candidate order with the preset minimum order for maintaining closed-loop controllability, and take the larger value of the two as the order of the low-dimensional offline data matrix.

[0054] make To meet The number of singular values, To meet The smallest integer. The energy retention factor can take values ​​between 0.95 and 0.995. The final retention order is determined as follows:

[0055] in, To maintain the minimum order of controllability of the closed loop, if the mean square error of the one-step prediction calculated using the validation samples exceeds the set threshold, then the order should be appropriately increased. If the online optimization solution time exceeds the allowable value of the control cycle, then appropriately reduce the time. or improve This reduces the scale of online computing.

[0056] S3.4 Construct a predictive control optimization function based on the low-dimensional offline data matrix.

[0057] The specific steps for reconstructing a high-dimensional data matrix into a low-dimensional form and constructing a predictive control optimization problem include: 1) Low-dimensional data matrix reconstruction To reduce the dimensionality of online optimization and mitigate the impact of sampling noise, the offline data matrix was optimized. Perform singular value decomposition:

[0058] Where s are the singular values ​​of the data matrix, satisfying The retention order is determined according to the singular value threshold rule in step 3). By selecting the main singular values ​​and their corresponding eigenvectors, a compressed data matrix is ​​constructed from the original data matrix:

[0059] in, express The former Columns. Finally, based on the original data matrix. The low-dimensional data matrix is ​​divided according to the row intervals corresponding to past inputs, past outputs, predicted inputs, and predicted outputs. Divided into:

[0060] The above four sub-blocks serve as fixed data parameters in online rolling optimization, used to establish a mapping relationship between the currently acquired system data and the controller's predicted output.

[0061] 2) Constructing the predictive control optimization problem After the controller is put into closed-loop operation, at time Read the most recent An additional current reference increment and measurement output form a historical observation window:

[0062] Decision variables for online optimization include coefficients Future control input sequence Future output prediction sequence and relaxation amount Based on low-dimensional data matrices The following quadratic programming problem is established:

[0063]

[0064] in, Represents the predictive control optimization function. and Let represent the lengths of the input-output sequences and the prediction length of the model-free predictive controller, respectively. Let Q, R, and S represent the input, output, and control smoothing weight matrices, respectively. Represents the first term in the prediction time domain. One prediction step, This indicates that at time k, the future... The predicted value output by the step system. This represents the output reference value for the corresponding prediction step. This indicates the future time predicted at time k. Additional current reference value, Represents the regularization coefficient. Represents the regularization term, The coefficients of the slack variable. Represents slack variables. , , and These represent the submatrices representing historical input, historical output, future input, and future output within the low-dimensional data matrix, respectively. and These represent the lower and upper limits of the additional current reference value, respectively. and These represent the lower and upper limits of the system output, respectively. This represents the column vector concatenation operator. This represents the square of the weighted L2 norm. It is by The scaling quadratic regularization term, with Slack variables of punishment Used to relax data consistency. This means that the input and output variables of the controller must be within a certain set of constraints.

[0065] S4. During the online operation phase, the active power and reactive power output of the inverter in the current time period are obtained and input into the modelless predictive controller. The modelless predictive controller solves the predictive control optimization function to obtain the additional current reference value. After the controller is put into operation, it collects the latest input and output data in real time, solves the optimization problem online to obtain the optimal control input, and obtains the modulated wave voltage amplitude signal through current inner loop control. The specific steps include: Within each control cycle, the controller solves for the optimal future additional current reference sequence: Each component in the above formula is According to the rolling optimization control strategy, only the first optimal additional current reference value is used. It operates on the current control cycle, and in the next control cycle, the system output is re-acquired and the optimization problem is solved again.

[0066] The model-free predictive controller employs a rolling optimization method. After the low-dimensional offline data matrix is ​​initially constructed, it remains fixed. During the online operation phase, the active and reactive power output of the inverter at the current moment is collected in each control cycle. The predictive control optimization problem is solved to obtain the additional current reference value for the current control cycle.

[0067] S5. Add the additional current reference value to the current reference value to obtain the reference value of the inner current loop, input it to the inner current loop for processing to obtain the modulation wave voltage amplitude signal, generate a drive signal based on the modulation wave voltage amplitude signal and control the inverter.

[0068] Let the reference current given by the power outer loop be... and The final reference current value entering the inner current loop is:

[0069] The above reference values ​​were compared with the measured current. , The comparison is performed, and the result is generated after PI regulation and cross-coupling compensation. Shaft voltage command:

[0070] in, For filtering inductors, , , , These are the parameters for the inner loop current control. For the grid synchronization angular velocity, , For the grid connection point voltage at Components in the coordinate system. To prevent overmodulation, the generated voltage command can be limited according to the DC bus voltage, and then sent to the modulation stage to generate the converter switching signal.

[0071] Reference Figure 4 This is an embodiment of the present invention, specifically a three-phase grid-connected converter system. To fully verify the effectiveness and superiority of the proposed method, this study designed three different control strategies for comparative experiments: ① a traditional dual-closed-loop control strategy with an outer power loop and an inner current loop; ② model-free predictive control using an undimension-reduced data matrix; and ③ model-free predictive control using a dimension-reduced data matrix. The simulation operating condition was set as the grid-connected converter operating in grid-connected mode, and the specific simulation parameter configurations for each key component are shown in Table 1.

[0072] Table 1. Relevant parameters of the converter grid-connected system in the simulation verification of the embodiment.

[0073] For the specific configuration of the simulation experiment, the steady-state operation target of the grid-connected inverter is set as follows: reference value of converter output active power. Output reactive power reference value Rated frequency value of power grid The proportional and integral gains of the PI controller in the phase-locked loop are respectively... and The proportional and integral gains of the current loop PI controller are respectively... and .

[0074] The core parameters of the model-free predictive controller are configured as follows: , , Because of measurement noise in the sensor, the dimension reduction is determined according to the method in step 3. Reference for system output trajectory Cost matrix and , Regularization weights and .in, Indicates the order is The identity matrix.

[0075] The converter is operated in grid-connected mode. At t = 0.2 s, the converter is subjected to a disturbance. The simulation is run, and the active and reactive power outputs of the system are recorded as follows: Figure 5 and Figure 6 As shown.

[0076] Combination Figure 5 , Figure 6 Simulation waveform analysis shows that under the traditional grid-connected control mode, due to the influence of the high impedance characteristics of the weak grid, poor dynamic coupling occurs between the converter phase-locked loop and the grid, resulting in oscillations in the active and reactive power of the grid-connected inverter. This indicates that the system has a small disturbance instability problem, and the existing constant parameter control loop is difficult to effectively dampen such oscillations.

[0077] When the simulation time reaches After implementing the model-free predictive controller (SVD), power oscillations were rapidly and effectively suppressed. A comparison clearly shows that the control effects of the controllers using the dimensionality-reduced data matrix (orange curve) and those without (red curve) highly overlapped throughout the dynamic adjustment process; their response speeds and final steady-state accuracy were similar. This result objectively confirms that using SVD to reduce the dimensionality of the data matrix significantly reduces computational dimensionality without sacrificing control performance, fully inheriting the ability of a controller using a full-order data matrix to adaptively suppress grid oscillations and maintain stable system operation without requiring a precise model.

[0078] Further evaluation of the real-time performance of the control algorithm was conducted. Figure 6 This demonstrates a dynamic comparison of the controller's time consumption in solving the optimization problem before and after data matrix dimensionality reduction. The blue curve in the figure clearly shows that before dimensionality reduction, due to the involvement of high-dimensional data matrix operations, the single-step optimization solution time is not only longer, fluctuating between 8ms and 20ms, but also occurs at the moment of controller activation (…). The initialization computation delay is significant, which negatively impacts the real-time performance of the controller. In contrast, the controller (red curve) exhibits computational acceleration after introducing singular value decomposition for dimensionality reduction. Due to the significant reduction in the dimensionality of the optimization variables, the single-step solution time of this algorithm is generally within 5ms, and the entire process maintains high stability with almost no large-scale time fluctuations. Figure 7 As shown.

[0079] This method does not rely on a precise physical model of the inverter. Instead, it constructs a data matrix reflecting the dynamic characteristics of the inverter system based on input and output data collected by sensors during operation. Singular Value Decomposition (SVD) is then used to extract and compress the main dynamic information from the data matrix, obtaining a low-dimensional data representation that accelerates the solution of the optimization problem. Based on this, a model-free predictive controller is established using the low-dimensional data matrix and embedded into the grid-connected inverter control loop to achieve rapid suppression of system oscillations.

[0080] The proposed dimensionality-reduction accelerated model-free predictive control strategy reduces reliance on precise grid parameters and inverter mathematical models, effectively mitigating performance degradation caused by model uncertainty, line impedance variations, and operating condition fluctuations. Furthermore, by compressing the dimension of the data matrix, it significantly reduces the number of decision variables and computational burden during online optimization, improving the real-time execution capability of the control algorithm in high-switching-frequency power electronic systems. This method can rapidly suppress grid-connected oscillations under complex conditions such as weak grids, multiple disturbances, and parameter variations, enhancing the inverter's dynamic response performance, disturbance rejection capability, and operational robustness. It provides a model-free predictive control scheme with both computational efficiency and engineering application value for improving grid stability under scenarios with high proportions of renewable energy integration.

[0081] In summary, this invention does not rely on inverter analytical models, grid equivalent impedance, line parameters, or locally linearized models. Instead, it directly utilizes input-output data obtained from offline acquisition and online measurement to establish predictive relationships. Simultaneously, through data matrix dimensionality reduction, key dynamic information extraction, and rolling optimization control, the online computation load is significantly reduced while preserving the system's main dynamic characteristics. Based on this method, the controller can achieve coordinated adjustment of the grid-connected inverter's power, voltage, and DC bus status under complex conditions such as unknown grid parameters, weak grid disturbances, topology changes, and operating point switching. This improves oscillation suppression capability, real-time control performance, and stable operation under complex conditions.

[0082] This invention enables the controller to achieve fast and stable oscillation suppression control without relying on precise mathematical models of converters, equivalent impedance parameters of the power grid, and line parameters. Addressing the problems of high-dimensional data matrices in online optimization, such as high variable dimensionality, long solution time, and difficulty in meeting rapid response requirements, this invention uses singular value decomposition to reduce the dimensionality of the high-dimensional data matrix and provides corresponding principles for selecting the dimensionality reduction dimension. This extracts a low-dimensional data matrix that reflects the dominant dynamic characteristics of the system, thereby constructing a predictive control optimization problem with smaller computational scale and faster solution speed. The proposed method not only avoids the advantage of relying on precise analytical models but also significantly reduces the computational burden and hardware requirements for online optimization. This allows the controller to establish predictive control relationships based solely on collected input and output data, and to quickly sense and continuously optimize and adjust the system's oscillation trends under changing power grid operating conditions, improving its dynamic performance, real-time control capabilities, and stable operation level in complex power grid environments.

[0083] The above detailed embodiments illustrate the technical solution and beneficial effects of the present invention. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A highly efficient solution-based data trajectory-driven control method for oscillation suppression in grid-connected inverters, characterized in that, The method includes the following steps: S1. Obtain the three-phase voltage and three-phase current at the inverter's grid connection point, and then calculate the active power and reactive power output by the inverter. S2. The active power and reactive power output by the inverter are used as the controller input, and the components of the additional current reference value on the d-axis and q-axis are used as the controller output, thereby constructing a model-free predictive controller. The model-free predictive controller includes an offline preparation stage and an online operation stage. S3. In the offline preparation stage, the system input and output sequences of a preset historical period are obtained to construct an offline data matrix. The offline data matrix is ​​reduced to a low-dimensional offline data matrix, and then a predictive control optimization function is constructed based on the low-dimensional offline data matrix. S4. During the online operation phase, the active power and reactive power output of the inverter for the current period are obtained and input into the modelless predictive controller. The modelless predictive controller solves the predictive control optimization function to obtain the additional current reference value. S5. Add the additional current reference value to the current reference value to obtain the reference value of the inner current loop, input it to the inner current loop for processing to obtain the modulation wave voltage amplitude signal, generate a drive signal based on the modulation wave voltage amplitude signal and control the inverter.

2. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 1, characterized in that: Specifically, S3 is: S3.

1. The active power and reactive power output by the inverter are used as the system output, and the components of the additional current reference value on the d-axis and q-axis are used as the system input. S3.2 Apply white noise excitation to the system, and then collect the active power, reactive power, and additional current reference values ​​of the inverter output in the d-axis and q-axis for a first preset time length, and then construct the system input-output sequence in time order; S3.3 Construct an offline data matrix based on the system input and output sequence, perform singular value decomposition on the offline data matrix to determine the order of the low-dimensional offline data matrix, and then construct the low-dimensional offline data matrix based on the order of the low-dimensional offline data matrix. S3.4 Construct a predictive control optimization function based on the low-dimensional offline data matrix.

3. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 2, characterized in that: The order of the low-dimensional offline data matrix is ​​determined according to the following steps: 1) Determine if measurement noise exists in the offline data matrix: If it exists, then the full-rank dimension of the offline data matrix is ​​used as the order of the low-dimensional offline data matrix; If it does not exist, the singular value decomposition threshold is determined based on whether the noise level of the data in the offline data matrix is ​​known, and then the number of singular values ​​that are greater than the singular value decomposition threshold after the offline data matrix is ​​decomposed is taken as the first candidate order. 2) Calculate the cumulative energy ratio based on the singular values ​​obtained from the singular decomposition of the offline data matrix, and take the minimum number of singular values ​​that satisfy the cumulative energy ratio being greater than or equal to the preset energy retention coefficient as the second candidate order; 3) Compare the first candidate order with the second candidate order, and take the smaller value of the two as the comprehensive candidate order. Then compare the comprehensive candidate order with the preset minimum order for maintaining closed-loop controllability, and take the larger value of the two as the order of the low-dimensional offline data matrix.

4. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 3, characterized in that: When the standard deviation of offline data noise is known, the singular value decomposition threshold is determined according to the following formula: When the noise standard deviation of the offline data is unknown, the singular value decomposition threshold is determined according to the following formula: in, This represents the singular value decomposition threshold. The shape factor represents the shape of the offline data matrix. , and These represent the lengths of the input and output sequences, the prediction length of the model-free predictive controller, and the observation length, respectively. This represents the standard deviation of noise in offline data. This represents the optimal hard threshold coefficient when the noise level is known. This represents the optimal hard threshold coefficient when the noise level is unknown. This represents the median of the singular values ​​obtained from the singular decomposition of the offline data matrix.

5. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 1, characterized in that: The cumulative energy ratio is set according to the following formula: in, Let represent the proportion of the cumulative energy of the first r singular values ​​arranged in descending order to the total energy; let i represent the index number of the singular values ​​after arranging them in descending order; let m represent the total number of singular values; and let r represent the number of singular values ​​currently retained. This represents the i-th singular value obtained after performing singular value decomposition on the offline data matrix.

6. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 1, characterized in that: The predictive control optimization function is set according to the following formula: ； ； in, Represents the predictive control optimization function. and Let represent the lengths of the input-output sequences and the prediction length of the model-free predictive controller, respectively. Let Q, R, and S represent the input, output, and control smoothing weight matrices, respectively. Represents the first term in the prediction time domain. One prediction step, This indicates that at time k, the future... The predicted value output by the step system. This represents the output reference value for the corresponding prediction step. This indicates the future time predicted at time k. Additional current reference value, Represents the regularization coefficient. Represents the regularization term, The coefficients of the slack variable. Represents slack variables. , , and These represent the submatrices representing historical input, historical output, future input, and future output within the low-dimensional data matrix, respectively. Represents the most recent time before time k. An additional current reference increment Represents the most recent time before time k. One measurement output, and These represent the lower and upper limits of the additional current reference value, respectively. and These represent the lower and upper limits of the system output, respectively. This represents the column vector concatenation operator. This represents the square of the weighted L2 norm.

7. The efficient solution data trajectory-driven control method for oscillation suppression of grid-connected inverters according to claim 1, characterized in that: The model-free predictive controller employs a rolling optimization method. After the low-dimensional offline data matrix is ​​initially constructed, it remains fixed. During the online operation phase, the active and reactive power output of the inverter at the current moment is collected in each control cycle. The predictive control optimization problem is solved to obtain the additional current reference value for the current control cycle.

Images