A three-phase inverter parallel droop control method based on reinforcement learning

Abstract

This invention discloses a parallel droop control method for three-phase inverters based on reinforcement learning, belonging to the field of power electronics control technology. The method includes the following steps: Step 1: Collecting operational sampling data, capacity identifiers, and current droop control parameters to form parallel operation data; Step 2: Performing power conversion and deviation aggregation to form power state and parallel deviation state; Step 3: Forming parameter tuning state quantities and parameter tuning evaluation values; Step 4: Establishing an improved Kiefer-Wolfowitz parameter tuning process; Step 5: Forming a mirror disturbance vector, a positive disturbance parameter set, and a negative disturbance parameter set; Step 6: Forming gradient estimates and candidate update parameters; Step 7: Forming target droop control parameters through safety projection and evaluation hysteresis determination; Step 8: Executing parallel droop control to generate parallel droop control results. This invention improves power balance, circulating current suppression capability, and operational stability.

Description

Technical Field

[0001] This invention relates to the field of power electronic control technology, and in particular to a method for parallel droop control of a three-phase inverter based on reinforcement learning. Background Technology

[0002] With the continuous development of distributed power sources, microgrids, and new energy power supply scenarios, droop control and power coordination technologies for parallel operation of three-phase inverters have received widespread attention. Existing parallel control methods for three-phase inverters mainly rely on fixed droop coefficients, preset virtual impedance parameters, and empirical adjustment strategies for power distribution. However, these methods commonly suffer from the following problems in practical applications: During parallel operation, the rated capacity, line impedance, and load sharing conditions of different inverters vary. Fixed droop parameters cannot simultaneously take into account active power distribution, reactive power distribution, and circulating current suppression, which can easily lead to power state deviation and continuous accumulation of parallel deviation. Virtual impedance parameters usually rely on manual setting and lack online adaptive tuning capabilities when facing load changes, voltage and frequency fluctuations, and sampling disturbances. This makes it difficult for the parameter tuning evaluation value to decrease steadily and causes repeated oscillations in the adjustment process. Traditional parameter optimization methods can easily introduce common bus disturbances when adjusting the droop coefficient and virtual impedance parameters. Candidate update parameters lack operating boundary constraints and evaluation hysteresis judgment, which may cause output current to exceed limits, voltage and frequency deviations to increase, and the parallel droop control results to be unstable.

[0003] Therefore, how to provide a method for parallel droop control of three-phase inverters based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a droop control method for three-phase inverters in parallel operation based on reinforcement learning. This invention utilizes an improved Kiefer-Wolfowitz parameter tuning process to adaptively tune the droop coefficient and virtual impedance parameters of the three-phase inverters in parallel operation. It achieves stable parameter updates through a weighted zero-sum mirror disturbance vector, safe projection, and evaluation hysteresis determination. This method has the advantages of more balanced power distribution, stronger circulating current suppression capability, smaller voltage and frequency deviation, and higher stability in parallel operation.

[0005] A method for parallel droop control of a three-phase inverter based on reinforcement learning according to an embodiment of the present invention includes the following steps: Step 1: Collect the operating sampling data, capacity identification, and current droop control parameters of each three-phase inverter to form parallel operation data; Step 2: Perform power conversion and deviation collection on the parallel operation data to form the power status and the parallel deviation status characterizing the power distribution, circulating current and voltage frequency deviation; Step 3: Generate parameter tuning state quantities based on parallel deviation states, and classify the parallel deviation states into parameter tuning evaluation values ​​according to the deviation amplitude; Step 4: Using the droop coefficient and virtual impedance parameter as the parameters to be tuned, establish the improved Kiefer-Wolfowitz parameter tuning process; Step 5: Determine the inverter capacity weight based on the capacity identifier, form a mirror disturbance vector with a weighted sum of zero according to the inverter capacity weight, and form a positive disturbance parameter set and a reverse disturbance parameter set by the mirror disturbance vector and the parameters to be tuned; Step 6: Use the improved Kiefer-Wolfowitz parameter tuning process to process the parameter tuning evaluation values ​​corresponding to the forward and reverse perturbation parameter sets respectively, and form gradient estimates and candidate update parameters; Step 7: Perform safe projection on the candidate update parameters according to the running boundary constraints, and perform evaluation hysteresis judgment on the projection results to form target sagging control parameters; Step 8: Execute parallel droop control of the three-phase inverter according to the target droop control parameters, and generate parallel droop control results.

[0006] Optionally, the three-phase inverter is a power electronic conversion device connected to the same AC bus and participating in parallel droop control. Each three-phase inverter has an independent inverter identifier, capacity identifier, and current droop control parameters.

[0007] Optionally, step one specifically includes: Collect the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value of each three-phase inverter according to a unified sampling period, and collect the sampling results according to the inverter identification to form operation sampling data; Read the capacity identifier of each three-phase inverter, pair the capacity identifier with the inverter identifier, and use the capacity identifier to represent the rated capacity of the corresponding three-phase inverter. Obtain the current droop control parameters executed by each three-phase inverter at the sampling time. The current droop control parameters include the droop coefficient and the virtual impedance parameter. Using the sampling time and inverter identifier as indexes, the operating sampling data, capacity identifier, and current droop control parameters are associated and arranged to form parallel operation data.

[0008] Optionally, step two specifically includes: Extract the operating sampling data of each three-phase inverter at the same sampling moment from the parallel operation data, and perform phase alignment, amplitude calibration and time position correction on the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value in the operating sampling data to form a voltage and current synchronous record; A three-phase coordinate transformation is performed on the synchronous voltage and current records to convert the AC side voltage and AC side current sampling values ​​into voltage and current components in a rotating coordinate system. Power conversion is then performed based on the voltage and current components to generate the power state of each three-phase inverter. Capacity identifiers are extracted from parallel operation data. Based on the capacity identifiers, the power status is normalized to the capacity benchmark. The power distribution deviation of each three-phase inverter relative to the capacity benchmark is calculated. The circulating current deviation and voltage frequency deviation are calculated based on the voltage and current synchronization records to form a deviation collection record. The deviations are collected and recorded according to the sampling time. The power distribution deviation, circulating current deviation and voltage frequency deviation are correlated and organized to form a parallel deviation state that characterizes the deviations of power distribution, circulating current and voltage frequency.

[0009] Optionally, step three specifically includes: Extract the power distribution deviation, circulating current deviation, and voltage frequency deviation corresponding to the same sampling time from the parallel deviation state, arrange them according to the inverter identifier and sampling time, and form a deviation state record; The deviation state record is normalized in amplitude and calibrated in direction. The power distribution deviation, circulating current deviation and voltage frequency deviation are converted into deviation components under the same dimension. The parameter tuning state quantity is formed according to the arrangement of the deviation components. The parameter tuning state variables are used as the learning state of the improved Kiefer-Wolfowitz parameter tuning process, so that the parameter tuning state variables can be used to characterize the adjustment basis of the parameter to be tuned at the current sampling time. The evaluation weights are assigned to power distribution deviation, circulating current deviation, and voltage frequency deviation based on the amplitude of the deviation components. The deviation components are then merged and calculated according to the evaluation weights to form parameter tuning evaluation values ​​that serve as feedback for reinforcement learning gains.

[0010] Optionally, step four specifically includes: Extract the droop coefficient and virtual impedance parameters of each three-phase inverter from the existing droop control parameters. Divide the droop coefficient according to the active power regulation direction and the reactive power regulation direction, and divide the virtual impedance parameters according to the resistance regulation direction and the reactance regulation direction. Arrange them according to the inverter identification to form the parameters to be tuned. The parameter tuning state variables are defined as the learning states of the improved Kiefer-Wolfowitz parameter tuning process, the parameter tuning evaluation values ​​are defined as the reinforcement learning benefit feedback of the improved Kiefer-Wolfowitz parameter tuning process, and the parameter tuning correlation between the learning states, the parameters to be tuned, and the reinforcement learning benefit feedback is established. Two-sided perturbation rules are configured around the parameter to be tuned. The two-sided perturbation rules are used to generate parameter perturbations with the same amplitude and opposite directions on the same parameter to be tuned, and provide generation rules for the positive perturbation parameter group and the negative perturbation parameter group. A stochastic approximation update rule is configured around the parameter to be tuned. The stochastic approximation update rule is used to calculate the gradient estimate based on the parameter tuning evaluation value corresponding to the positive perturbation parameter group and the parameter tuning evaluation value corresponding to the negative perturbation parameter group, and to generate candidate update parameters based on the gradient estimate. By merging parameter tuning correlations, two-sided perturbation rules, and random approximation update rules, an improved Kiefer-Wolfowitz parameter tuning process is established.

[0011] Optionally, step five specifically includes: Capacity identifiers are extracted from parallel operation data, and the rated capacity of each three-phase inverter is determined based on the capacity identifiers. The ratio of the rated capacity to the total rated capacity of the parallel inverters is determined as the inverter capacity weight. Arrange the parameters to be tuned according to the inverter identification, group the droop coefficient and virtual impedance parameters corresponding to the same three-phase inverter into the same parameter unit, and determine the parameter positions corresponding to the active power adjustment direction, reactive power adjustment direction, resistance adjustment direction and reactance adjustment direction within the parameter unit. Based on the disturbance scale, an initial disturbance component is configured for each parameter position. The initial disturbance component is then subjected to capacity-weighted correction in the same adjustment direction, so that the sum of the product of the disturbance component corresponding to each three-phase inverter and the inverter capacity weight is zero, thus forming a mirrored disturbance vector. The mirror perturbation vector is superimposed on the parameter to be tuned along the positive direction to form a positive perturbation parameter set, and the mirror perturbation vector is superimposed on the parameter to be tuned along the negative direction to form a negative perturbation parameter set. Establish a correspondence between the forward and reverse disturbance parameter sets and the parameter tuning state variables, so that the forward and reverse disturbance parameter sets enter the improved Kiefer-Wolfowitz parameter tuning process.

[0012] Optionally, step six specifically includes: An improved Kiefer-Wolfowitz parameter tuning process is adopted. Under the same parameter tuning state variable, the parallel droop control of the three-phase inverter is performed with the forward disturbance parameter group and the reverse disturbance parameter group respectively, to obtain the parameter tuning evaluation value corresponding to the forward disturbance parameter group and the parameter tuning evaluation value corresponding to the reverse disturbance parameter group. The difference between the parameter tuning evaluation values ​​corresponding to the positive disturbance parameter group and the parameter tuning evaluation values ​​corresponding to the negative disturbance parameter group is calculated to form an evaluation difference. The evaluation difference is then associated with the disturbance components in the mirror disturbance vector according to the parameter position. Based on the evaluation difference, disturbance scale, and disturbance component, the gradient estimates of the parameter to be tuned in the active power regulation direction, reactive power regulation direction, resistance regulation direction, and reactance regulation direction are calculated according to the Kiefer-Wolfowitz stochastic approximation rule. Based on the gradient estimate and update step size, the parameters to be tuned are adjusted to form candidate update parameters, which are then used as the objects of safe projection.

[0013] Optionally, step seven specifically includes: Voltage frequency boundary, output current boundary, circulating current boundary and parameter boundary are extracted from parallel operation data and parallel deviation state, and operation boundary constraints are formed based on voltage frequency boundary, output current boundary, circulating current boundary and parameter boundary; The candidate update parameters are compared against the boundary constraints according to the operational boundary constraints. The parameter components that fail the boundary comparison are mapped to the range limited by the operational boundary constraints, while the parameter components that pass the boundary comparison are retained as their original values, thus forming the projection result. Based on the projection results, a projection evaluation value is generated, and the projection evaluation value is compared with the parameter tuning evaluation value to form an evaluation change record; The evaluation hysteresis determination is performed on the evaluation change record according to the continuous control window. The evaluation hysteresis determination is used to determine whether the projection result maintains the state of decreasing parameter tuning evaluation value within the continuous control window. If the evaluation hysteresis judgment is passed, the projection result is determined as the target sagging control parameter; if the evaluation hysteresis judgment is not passed, the current sagging control parameter is determined as the target sagging control parameter.

[0014] Optionally, step eight specifically includes: The target droop control parameters are assigned according to the inverter identification, and the droop coefficient and virtual impedance parameters corresponding to each three-phase inverter are compiled into a target parameter allocation record. The power state is converted by using the droop coefficient in the target parameter allocation record, and the frequency reference and voltage reference of each three-phase inverter are calculated to form a droop control reference record. The virtual impedance parameters in the target parameter allocation record are used to perform impedance compensation conversion on the running sample data, calculate the virtual impedance compensation amount for each three-phase inverter, and synthesize the virtual impedance compensation amount with the droop control reference record to form a parallel control reference record. Based on the parallel control reference record, the modulation control quantity of each three-phase inverter is generated, and the parallel droop control of the three-phase inverter is executed according to the modulation control quantity to form a parallel operation control record. The power state and parallel deviation state are updated on the parallel operation control record. The target droop control parameters, the updated power state, and the updated parallel deviation state are then correlated and organized to generate the parallel droop control result.

[0015] The beneficial effects of this invention are: This invention collects operational sampling data, capacity identification, and current droop control parameters from each three-phase inverter to form parallel operation data that simultaneously reflects the operating status, capacity differences, and control parameter status. Then, through power conversion and deviation aggregation, it generates power status and parallel deviation status, enabling power distribution deviation, circulating current deviation, and voltage / frequency deviation in parallel operation of three-phase inverters to be uniformly incorporated into the parameter tuning chain. Compared to traditional control methods that rely on fixed droop coefficients and manually set virtual impedance parameters, this invention can generate parameter tuning status variables based on parallel deviation status and aggregate deviation amplitudes into parameter tuning evaluation values. This eliminates reliance on manual experience for adjusting droop coefficients and virtual impedance parameters, instead establishing a closed-loop evaluation basis around power distribution balance, circulating current suppression, and voltage / frequency stability. By using the parameter tuning state variables as the learning state of the improved Kiefer-Wolfowitz parameter tuning process and the parameter tuning evaluation values ​​as reinforcement learning benefit feedback, this invention can continuously identify the deviation change trend at different sampling times during the parallel operation of three-phase inverters, so as to establish a stable correspondence between the adjustment direction of the parameter to be tuned and the actual operating deviation, thereby improving the adaptive control capability of parallel inverters under conditions of inconsistent rated capacity, load changes and line impedance differences.

[0016] This invention further utilizes inverter capacity weighting to construct a weighted, zero-sum mirrored disturbance vector, causing the forward and reverse disturbance parameter sets to form pairs around the parameter to be tuned. This avoids the introduction of common-mode disturbances in the parallel bus during ordinary disturbance optimization, reducing the impact on the common voltage and common frequency during parameter tuning. By using an improved Kiefer-Wolfowitz parameter tuning process to process the parameter tuning evaluation values ​​corresponding to the forward and reverse disturbance parameter sets respectively, and calculating the gradient estimate based on the evaluation difference, disturbance scale, and disturbance components, this invention can obtain the effective update direction of the parameter to be tuned without relying on an accurate line impedance model, improving the accuracy of the coordinated tuning of virtual impedance parameters and droop coefficients. Furthermore, this invention performs a safe projection of candidate update parameters according to operating boundary constraints and performs evaluation hysteresis determination on the projection results. This ensures that the target droop control parameters are only determined when the operating boundary is satisfied and the parameter tuning evaluation value continuously decreases. This suppresses erroneous parameter updates caused by sampling noise, transient load disturbances, and single evaluation fluctuations, reducing the risks of output current exceeding limits, voltage frequency deviation expansion, and circulating current increase. Therefore, this invention can improve the stability of parallel droop control results while ensuring safety boundaries, enabling three-phase inverter parallel systems to achieve more balanced power distribution, smoother voltage and frequency response, and more reliable parallel operation. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a three-phase inverter parallel droop control method based on reinforcement learning proposed in this invention; Figure 2 This is a schematic diagram of the improved Kiefer-Wolfowitz parameter tuning process for a three-phase inverter parallel droop control method based on reinforcement learning proposed in this invention. Figure 3 This is a schematic diagram of the parallel droop control structure of a three-phase inverter, which is based on a reinforcement learning-based method for controlling parallel droop in a three-phase inverter, as proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for parallel droop control of a three-phase inverter based on reinforcement learning includes the following steps: Step 1: Collect the operating sampling data, capacity identification, and current droop control parameters of each three-phase inverter to form parallel operation data; Step 2: Perform power conversion and deviation collection on the parallel operation data to form the power status and the parallel deviation status characterizing the power distribution, circulating current and voltage frequency deviation; Step 3: Generate parameter tuning state quantities based on parallel deviation states, and classify the parallel deviation states into parameter tuning evaluation values ​​according to the deviation amplitude; Step 4: Using the droop coefficient and virtual impedance parameter as the parameters to be tuned, establish the improved Kiefer-Wolfowitz parameter tuning process; Step 5: Determine the inverter capacity weight based on the capacity identifier, form a mirror disturbance vector with a weighted sum of zero according to the inverter capacity weight, and form a positive disturbance parameter set and a reverse disturbance parameter set by the mirror disturbance vector and the parameters to be tuned; Step 6: Use the improved Kiefer-Wolfowitz parameter tuning process to process the parameter tuning evaluation values ​​corresponding to the forward and reverse perturbation parameter sets respectively, and form gradient estimates and candidate update parameters; Step 7: Perform safe projection on the candidate update parameters according to the running boundary constraints, and perform evaluation hysteresis judgment on the projection results to form target sagging control parameters; Step 8: Execute parallel droop control of the three-phase inverter according to the target droop control parameters, and generate parallel droop control results.

[0020] In this embodiment, the three-phase inverter is a power electronic conversion device connected to the same AC bus and participating in parallel droop control. Each three-phase inverter has an independent inverter identifier, capacity identifier, and current droop control parameters.

[0021] In this embodiment, step one specifically includes: Each three-phase inverter is synchronously sampled according to a unified sampling period. At the sampling time, the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value are collected. The sampling values ​​from the same three-phase inverter at the same sampling time are grouped into the same sampling record. The inverter identifier is written into the sampling record. The sampling records are collected according to the inverter identifier to form the operation sampling data. Read the pre-configured capacity identifier of each three-phase inverter, pair the capacity identifier with the inverter identifier, so that the capacity identifier represents the rated capacity of the corresponding three-phase inverter, and the capacity identifier can be used as the basis for determining the inverter capacity weight during the parameter tuning process. Obtain the current droop control parameters executed by each three-phase inverter at the sampling time, extract the droop coefficient and virtual impedance parameters from the current droop control parameters, match the droop coefficient and virtual impedance parameters with the inverter identifier, and arrange them according to the sampling time to form a parameter record. Using the sampling time and inverter identifier as a joint index, the sampling records, capacity identifiers, and current droop control parameters in the running sampling data are time-position aligned and matched with the inverter object to form parallel operation data containing the correspondence between the running sampling data, capacity identifiers, and current droop control parameters.

[0022] In this embodiment, step two specifically includes: Extract the operating sampling data of each three-phase inverter at the same sampling time from the parallel operation data. Align the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value according to the sampling time. Convert the AC side voltage sampling value and AC side current sampling value to a unified amplitude scale. Correct the sampling sequence offset within the same sampling time to form a voltage and current synchronous record. A three-phase coordinate transformation is performed on the AC side voltage sampling value and AC side current sampling value in the voltage and current synchronous record. The sampling value in the three-phase stationary coordinate system is converted into voltage and current components in the rotating coordinate system. Active power and reactive power are converted according to the voltage and current components corresponding to the same sampling time. The conversion results are collected according to the inverter identifier to generate the power status of each three-phase inverter. Capacity identifiers are extracted from parallel operation data. Capacity benchmarks for each three-phase inverter are determined based on these identifiers. Power status is converted into power sharing under the capacity benchmark. Power distribution deviation is calculated based on the deviation relationship between power sharing. Circulating current deviation is calculated based on the current difference at the same parallel node in the voltage and current synchronization record. Voltage frequency deviation is calculated based on the deviation relationship between voltage amplitude and output frequency relative to the control benchmark, forming a deviation collection record. The deviation collection records are arranged in chronological order according to the sampling time. Power distribution deviation, circulating current deviation and voltage frequency deviation are included in the parallel deviation record at the same sampling time. The correspondence between the parallel deviation record and the power state is maintained according to the inverter identifier, forming a parallel deviation state that characterizes the deviation of power distribution, circulating current and voltage frequency.

[0023] In this embodiment, step three specifically includes: Extract the power distribution deviation, circulating current deviation, and voltage frequency deviation corresponding to the same sampling time from the parallel deviation state, bind the power distribution deviation, circulating current deviation, and voltage frequency deviation with the corresponding inverter identifier, and arrange them in the order of sampling time so that the deviation content at each sampling time has the corresponding inverter source and time position, forming a deviation state record. When performing amplitude normalization on the deviation state record, the value ranges of power distribution deviation, circulating current deviation and voltage frequency deviation are used as normalization benchmarks to convert deviations of different dimensions to a unified numerical range. When performing direction calibration on the deviation state record, the direction of deviation increase and the direction of deviation decrease are marked according to the positive and negative directions of the deviation. The normalized deviation value is combined with the direction calibration result to form the deviation component. The parameter tuning state quantity is formed according to the arrangement result of power distribution deviation, circulating current deviation and voltage frequency deviation. The parameter tuning state variables are used as the learning state of the improved Kiefer-Wolfowitz parameter tuning process, so that the parameter tuning state variables characterize the degree of deviation of the three-phase inverter parallel operation at the sampling time, and enable the improved Kiefer-Wolfowitz parameter tuning process to determine the disturbance handling basis of the parameter to be tuned based on the parameter tuning state variables. The proportions of power distribution deviation, circulating current deviation, and voltage frequency deviation in parameter tuning evaluation are determined according to the amplitude of the deviation components, forming evaluation weights. The deviation components are then weighted and merged based on the evaluation weights, and the weighted merging result is used as reinforcement learning benefit feedback to form parameter tuning evaluation values ​​used to evaluate the adjustment effect of the parameters to be tuned.

[0024] In this embodiment, step four specifically includes: Extract the droop coefficient and virtual impedance parameters of each three-phase inverter from the existing droop control parameters. In the droop coefficient, the parameters used for active power and output frequency corresponding adjustment are assigned to the active power adjustment direction, the parameters used for reactive power and output voltage corresponding adjustment are assigned to the reactive power adjustment direction, the resistance parameters used for voltage drop compensation in the virtual impedance parameters are assigned to the resistance adjustment direction, and the reactance parameters used for phase compensation in the virtual impedance parameters are assigned to the reactance adjustment direction. Then arrange them according to the inverter identification to form the parameters to be set. The parameter tuning state variable is defined as the learning state of the improved Kiefer-Wolfowitz parameter tuning process, so that the learning state characterizes the degree of deviation of power distribution, circulating current and voltage frequency at the sampling time. The parameter tuning evaluation value is defined as the reinforcement learning benefit feedback of the improved Kiefer-Wolfowitz parameter tuning process, so that the reinforcement learning benefit feedback characterizes the adjustment effect of the parameter to be tuned at the sampling time. The parameter tuning correlation between the learning state, the parameter to be tuned and the reinforcement learning benefit feedback is established. Two-sided perturbation rules are configured around the parameter to be tuned. The two-sided perturbation rules determine the perturbation amplitude according to the parameter category of the parameter to be tuned, and form parameter perturbations with opposite directions and the same amplitude on the same parameter to be tuned, so that the parameter to be tuned can form a positive perturbation parameter group and a negative perturbation parameter group based on the parameter perturbation. A stochastic approximation update rule is configured around the parameter to be tuned. The stochastic approximation update rule calculates the difference between the parameter tuning evaluation value corresponding to the positive perturbation parameter group and the parameter tuning evaluation value corresponding to the negative perturbation parameter group, and determines the adjustment direction of the parameter to be tuned based on the difference calculation result, forming a gradient estimate, and then forming candidate update parameters based on the gradient estimate. By merging parameter tuning correlations, two-sided perturbation rules, and stochastic approximation update rules, the learning state, parameters to be tuned, positive perturbation parameter set, negative perturbation parameter set, reinforcement learning reward feedback, gradient estimate, and candidate update parameters are linked and run in the parameter tuning order, thus establishing an improved Kiefer-Wolfowitz parameter tuning process.

[0025] This invention transforms the traditional Kiefer-Wolfowitz stochastic approximation algorithm into a parameter tuning process for parallel droop control of three-phase inverters. First, the droop coefficient is divided into active power regulation and reactive power regulation directions, and the virtual impedance parameter is divided into resistance regulation and reactance regulation directions. This transforms the parameters to be tuned from a single numerical optimization to a set of directional parameters directly corresponding to power distribution, circulating current suppression, and voltage / frequency regulation. Simultaneously, the parameter tuning state variables are defined as learning states, and the parameter tuning evaluation values ​​are defined as reinforcement learning benefit feedback. This allows the algorithm to move beyond relying solely on fixed disturbances and static objective functions, and instead evaluate the effectiveness of each parameter adjustment based on the deviation of power distribution, circulating current, and voltage / frequency at the sampling time. Furthermore, a dual-sided disturbance rule is configured around the parameters to be tuned, pairing forward and reverse disturbance parameter sets in opposite directions with the same amplitude. This process generates a gradient estimate that retains the advantage of the Kiefer-Wolfowitz algorithm in estimating gradients based on the difference between two-sided evaluations, while reducing the judgment bias caused by unidirectional disturbances. Then, a stochastic approximation update rule is used to calculate the difference between the parameter tuning evaluation values ​​corresponding to the forward disturbance parameter group and the parameter tuning evaluation values ​​corresponding to the reverse disturbance parameter group, forming a gradient estimate that characterizes the adjustment direction of the parameter to be tuned. Candidate update parameters are then generated based on the gradient estimate, enabling the droop coefficient and virtual impedance parameters to have online adaptive correction capabilities. Finally, the parameter tuning correlation, the two-sided disturbance rule, and the stochastic approximation update rule are merged, so that the learning state, the parameter to be tuned, the reinforcement learning benefit feedback, the gradient estimate, and the candidate update parameters operate in a unified parameter tuning order, improving the accuracy of the parameter update direction, the timeliness of the parallel operation deviation response, and the stability of the three-phase inverter parallel droop control.

[0026] In this embodiment, step five specifically includes: Extract the capacity identifier corresponding to each three-phase inverter from the parallel operation data, determine the rated capacity of each three-phase inverter according to the capacity identifier, and divide the rated capacity of each three-phase inverter by the sum of the rated capacities of the parallel inverters to obtain the inverter capacity weight of the corresponding three-phase inverter. Arrange the parameters to be set according to the inverter identification. Group the droop coefficient and virtual impedance parameters corresponding to the same three-phase inverter into the same parameter unit. In the parameter unit, map the active power adjustment direction to the active power adjustment parameter in the droop coefficient, map the reactive power adjustment direction to the reactive power adjustment parameter in the droop coefficient, map the resistance adjustment direction to the resistance parameter in the virtual impedance parameter, map the reactance adjustment direction to the reactance parameter in the virtual impedance parameter, and record the parameter position in the parameter unit. An initial disturbance component is configured for each parameter position according to the disturbance scale. The initial disturbance component is then corrected by capacity weighting in the same adjustment direction. The capacity weighting correction multiplies the disturbance component corresponding to each three-phase inverter with the inverter capacity weight, and makes the sum of the products under the same adjustment direction zero. The disturbance components after capacity weighting correction are arranged according to the inverter identifier and parameter position to form a mirror disturbance vector. Each perturbation component in the mirror perturbation vector is superimposed on the parameter to be tuned according to its corresponding parameter position to form a positive perturbation parameter set. Each perturbation component in the mirror perturbation vector is subtracted from the parameter to be tuned according to its corresponding parameter position to form a negative perturbation parameter set. The positive and negative perturbation parameter sets are distributed in pairs around the parameter to be tuned. By establishing a correspondence between the forward and reverse perturbation parameter sets and the parameter tuning state variables at the same sampling time, the improved Kiefer-Wolfowitz parameter tuning process can process the forward and reverse perturbation parameter sets separately based on the parameter tuning state variables, and generate parameter tuning evaluation values ​​corresponding to the forward and reverse perturbation parameter sets.

[0027] In this embodiment, step six specifically includes: When using the improved Kiefer-Wolfowitz parameter tuning process, under the same parameter tuning state variables, the forward disturbance parameter set is substituted into the parallel droop control of the three-phase inverter to obtain the forward operation result, and the parameter tuning evaluation value corresponding to the forward disturbance parameter set is calculated based on the forward operation result. Then, the reverse disturbance parameter set is substituted into the parallel droop control of the three-phase inverter to obtain the reverse operation result, and the parameter tuning evaluation value corresponding to the reverse disturbance parameter set is calculated based on the reverse operation result. The evaluation difference is formed by subtracting the evaluation value of the parameter tuning corresponding to the positive disturbance parameter group from the evaluation value of the parameter tuning corresponding to the negative disturbance parameter group. According to the position of the parameter in the parameter to be tuned, the evaluation difference is associated with the disturbance component located at the same parameter position in the mirror disturbance vector, so that the evaluation difference can characterize the direction of the influence of the disturbance component on the parameter tuning evaluation value. Based on the evaluation difference, disturbance scale and disturbance component, the Kiefer-Wolfowitz stochastic approximation rule is used to perform proportional conversion to obtain the changing trend of each parameter position. Then, the changing trends are aggregated according to the active power regulation direction, reactive power regulation direction, resistance regulation direction and reactance regulation direction to form the gradient estimate corresponding to the parameter to be tuned. Based on the gradient estimate and the update step size, parameter correction is performed on the parameter to be tuned. The parameter correction is directed to reduce the parameter tuning evaluation value. The magnitude of the parameter to be tuned is adjusted along the direction corresponding to the gradient estimate to form candidate update parameters, which are then used as the processing objects for safe projection.

[0028] In this embodiment, step seven specifically includes: The system extracts operating sampling data and current droop control parameters from the parallel operation data. It also extracts voltage frequency deviation and circulating current deviation from the parallel deviation state. The voltage sample value, output frequency sample value, and output current sample value in the operating sampling data are compared by amplitude to obtain the voltage frequency boundary and output current boundary. The circulating current deviation is compared by amplitude to obtain the circulating current boundary. The droop coefficient and virtual impedance parameter in the current droop control parameters are compared by value range to obtain the parameter boundary. The voltage frequency boundary, output current boundary, circulating current boundary, and parameter boundary are merged to form the operating boundary constraint. The parameter components in the candidate update parameters are checked one by one according to the operational boundary constraints. Boundary comparisons are performed on the voltage frequency deviation, output current amplitude, circulating current deviation and parameter values ​​caused by the parameter components. Parameter components that fail the boundary comparison are truncated according to the corresponding upper and lower boundary values. Parameter components that pass the boundary comparison retain their original values. The truncated parameter components and the parameter components that retain their original values ​​are rearranged to form the projection result. Based on the projection results, power conversion and deviation aggregation are re-executed to form power distribution deviation, circulating current deviation and voltage frequency deviation. Amplitude normalization and weighted merging are performed on the power distribution deviation, circulating current deviation and voltage frequency deviation to form the projection evaluation value. The difference between the projection evaluation value and the parameter setting evaluation value before safe projection is calculated to form the evaluation change record. The evaluation change records of adjacent sampling times are read according to the continuous control window. Each evaluation change record is marked with a decreasing state and a decreasing magnitude. Then, the satisfaction status of voltage frequency boundary, output current boundary, circulating current boundary and parameter boundary within the same continuous control window is checked to form an evaluation hysteresis judgment. When the evaluation hysteresis determination shows that the projected evaluation value remains in a decreasing state within the continuous control window and the operating boundary constraints remain satisfied, the projection result is determined as the target droop control parameter. When the evaluation hysteresis determination shows that the projected evaluation value does not remain in a decreasing state or the operating boundary constraints do not remain satisfied, the current droop control parameter is determined as the target droop control parameter.

[0029] In this embodiment, step eight specifically includes: The target droop control parameters are grouped according to the inverter identifier. The droop coefficient and virtual impedance parameter corresponding to the same inverter identifier are grouped into the same target parameter unit. The target parameter units are arranged according to the inverter identifier to form a target parameter allocation record that can characterize the correspondence between the target droop control parameters and each three-phase inverter. The droop coefficient in the target parameter allocation record is used to perform droop conversion on the power state. First, the active power change and reactive power change of each three-phase inverter are extracted from the power state. Then, the active power change is converted into frequency regulation and the reactive power change is converted into voltage regulation using the droop coefficient. The frequency regulation is applied to the frequency reference to form a frequency reference value, and the voltage regulation is applied to the voltage reference to form a voltage reference value. The frequency reference value and voltage reference value are organized according to the inverter identification to form a droop control reference record. The virtual impedance parameters in the target parameter allocation record are used to perform impedance compensation conversion on the running sample data. First, the AC side current sampling value of each three-phase inverter is extracted from the running sample data. Then, the AC side current sampling value is multiplied with the resistance adjustment direction parameter in the virtual impedance parameters to form a resistance compensation component. The AC side current sampling value is phase converted with the reactance adjustment direction parameter in the virtual impedance parameters to form a reactance compensation component. The resistance compensation component and the reactance compensation component are combined to form the virtual impedance compensation amount. The virtual impedance compensation amount is then superimposed on the droop control reference record to form the parallel control reference record. Based on the parallel control reference record, the modulation control quantity of each three-phase inverter is generated. First, the frequency reference quantity in the parallel control reference record is determined as the modulation wave frequency. The voltage reference quantity and the virtual impedance compensation quantity are combined into the modulation wave amplitude value. Then, the corresponding modulation control quantity is generated based on the modulation wave frequency and the modulation wave amplitude value. The parallel droop control of the three-phase inverter is executed according to the modulation control quantity, forming a parallel operation control record that records the correspondence between the modulation control quantity and the inverter identifier. The parallel operation control records are sampled and fed back. Based on the sampled feedback, the power conversion and deviation collection are re-executed to obtain the updated power state and the updated parallel deviation state. The target droop control parameters, the updated power state and the updated parallel deviation state are then associated and organized according to the sampling time and inverter identifier to generate the parallel droop control result.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to an experimental operation scenario of parallel power supply of three-phase inverters. The experimental subjects were three three-phase inverters with different rated capacities, which supplied power to a combination of resistive and inductive loads via parallel AC bus connections. During load operation, issues arose such as load step fluctuations, capacity inconsistencies, and line impedance inconsistencies. Traditional fixed droop control is prone to uneven active power distribution, increased reactive power distribution deviation, continuous increase in circulating current, and aggravated voltage-frequency deviation after load changes. To demonstrate the application effect of this invention, a comparative verification was conducted using both the traditional fixed droop control method and the method of this invention under the same parallel operation conditions. The rated capacity ratio of the three three-phase inverters was set to 1.00:0.75:0.50, the DC bus voltage was maintained within the set range, the AC side frequency reference was 50Hz, and the sampling period was kept consistent.

[0031] In this embodiment, the AC side voltage, AC side current, DC bus voltage, output frequency, capacity identifier, and current droop control parameters of each three-phase inverter are first collected to form parallel operation data. Then, phase alignment, amplitude calibration, and power conversion are performed on the parallel operation data to obtain the power state of each three-phase inverter. Furthermore, power distribution deviation, circulating current deviation, and voltage-frequency deviation are calculated to form the parallel deviation state. Traditional fixed droop control methods maintain the droop coefficient and virtual impedance parameters unchanged throughout the entire operation. The method of this invention converts the parallel deviation state into parameter tuning state variables and aggregates the parallel deviation states according to the deviation amplitude into parameter tuning evaluation values, allowing uneven power distribution, excessive circulating current, and voltage-frequency deviation to jointly participate in parameter tuning.

[0032] In the parameter tuning process, this invention uses the droop coefficient and virtual impedance parameter as the parameters to be tuned, and employs an improved Kiefer-Wolfowitz parameter tuning process. Specifically, the inverter capacity weight is determined based on the capacity identifier, and a mirrored disturbance vector with a weighted sum of zero is constructed under the same adjustment direction. This ensures that the disturbance components experienced by the three three-phase inverters satisfy the capacity weighted balance relationship, avoiding direct changes to the AC bus common reference by ordinary disturbance methods. The mirrored disturbance vector is superimposed along the forward direction to the parameters to be tuned to form a forward disturbance parameter set, and along the reverse direction to form a reverse disturbance parameter set. Then, parallel droop control of the three-phase inverters is executed separately, and a gradient estimate is formed based on the difference between the two sets of parameter tuning evaluation values. After the candidate update parameters are formed, a safe projection is performed through operating boundary constraints, and the target droop control parameters are screened through evaluation hysteresis judgment within the continuous control window to avoid incorrect parameter updates caused by single sampling fluctuations. The comparison results are shown in Table 1: Table 1 Comparison of Parallel Sag Control Effects

[0033] As shown in Table 1, under the combined effects of inconsistent rated capacities of three-phase inverters and load step fluctuations, traditional fixed droop control methods are prone to significant deviations in reactive power distribution and circulating current suppression due to the lack of online tuning mechanisms for the droop coefficient and virtual impedance parameters. This invention establishes a reinforcement learning feedback link through parameter tuning state variables and parameter tuning evaluation values, and utilizes a weighted, zero-sum mirror disturbance vector to reduce common bus disturbances, enabling the gradient estimate to more accurately reflect the impact of the parameters to be tuned on the parallel deviation state. Safety projection and evaluation hysteresis further restrict the application conditions of candidate update parameters, allowing the target droop control parameters to gradually converge within the operating boundary constraints. Experimental data show that this invention can significantly reduce active power distribution deviation, reactive power distribution deviation, and circulating current RMS value, while also reducing voltage frequency deviation and improving the recovery speed after load step fluctuations. This demonstrates that this invention can effectively solve the problems of fixed parameters, unstable disturbances, and accumulated parallel deviations in traditional three-phase inverter parallel droop control.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A parallel droop control method for a three-phase inverter based on reinforcement learning, characterized in that, Includes the following steps: Step 1: Collect the operating sampling data, capacity identification, and current droop control parameters of each three-phase inverter to form parallel operation data; Step 2: Perform power conversion and deviation collection on the parallel operation data to form the power status and the parallel deviation status characterizing the power distribution, circulating current and voltage frequency deviation; Step 3: Generate parameter tuning state quantities based on parallel deviation states, and classify the parallel deviation states into parameter tuning evaluation values ​​according to the deviation amplitude; Step 4: Using the droop coefficient and virtual impedance parameter as the parameters to be tuned, establish the improved Kiefer-Wolfowitz parameter tuning process; Step 5: Determine the inverter capacity weight based on the capacity identifier, form a mirror disturbance vector with a weighted sum of zero according to the inverter capacity weight, and form a positive disturbance parameter set and a reverse disturbance parameter set by the mirror disturbance vector and the parameters to be tuned; Step 6: Use the improved Kiefer-Wolfowitz parameter tuning process to process the parameter tuning evaluation values ​​corresponding to the forward and reverse perturbation parameter sets respectively, and form gradient estimates and candidate update parameters; Step 7: Perform safe projection on the candidate update parameters according to the running boundary constraints, and perform evaluation hysteresis judgment on the projection results to form target sagging control parameters; Step 8: Execute parallel droop control of the three-phase inverter according to the target droop control parameters, and generate parallel droop control results.

2. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, The three-phase inverters are power electronic conversion devices connected to the same AC bus and participating in parallel droop control. Each three-phase inverter has an independent inverter identifier, capacity identifier, and current droop control parameters.

3. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step one specifically involves: Collect the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value of each three-phase inverter according to a unified sampling period, and collect the sampling results according to the inverter identification to form operation sampling data; Read the capacity identifier of each three-phase inverter, pair the capacity identifier with the inverter identifier, and use the capacity identifier to represent the rated capacity of the corresponding three-phase inverter. Obtain the current droop control parameters executed by each three-phase inverter at the sampling time. The current droop control parameters include the droop coefficient and the virtual impedance parameter. Using the sampling time and inverter identifier as indexes, the operating sampling data, capacity identifier, and current droop control parameters are associated and arranged to form parallel operation data.

4. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step two specifically involves: Extract the operating sampling data of each three-phase inverter at the same sampling moment from the parallel operation data, and perform phase alignment, amplitude calibration and time position correction on the AC side voltage sampling value, AC side current sampling value, DC bus voltage sampling value and output frequency sampling value in the operating sampling data to form a voltage and current synchronous record; A three-phase coordinate transformation is performed on the synchronous voltage and current records to convert the AC side voltage and AC side current sampling values ​​into voltage and current components in a rotating coordinate system. Power conversion is then performed based on the voltage and current components to generate the power state of each three-phase inverter. Capacity identifiers are extracted from parallel operation data. Based on the capacity identifiers, the power status is normalized to the capacity benchmark. The power distribution deviation of each three-phase inverter relative to the capacity benchmark is calculated. The circulating current deviation and voltage frequency deviation are calculated based on the voltage and current synchronization records to form a deviation collection record. The deviations are collected and recorded according to the sampling time. The power distribution deviation, circulating current deviation and voltage frequency deviation are correlated and organized to form a parallel deviation state that characterizes the deviations of power distribution, circulating current and voltage frequency.

5. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step three specifically involves: Extract the power distribution deviation, circulating current deviation, and voltage frequency deviation corresponding to the same sampling time from the parallel deviation state, arrange them according to the inverter identifier and sampling time, and form a deviation state record; The deviation state record is normalized in amplitude and calibrated in direction. The power distribution deviation, circulating current deviation and voltage frequency deviation are converted into deviation components under the same dimension. The parameter tuning state quantity is formed according to the arrangement of the deviation components. The parameter tuning state variables are used as the learning state of the improved Kiefer-Wolfowitz parameter tuning process, so that the parameter tuning state variables can be used to characterize the adjustment basis of the parameter to be tuned at the current sampling time. The evaluation weights are assigned to power distribution deviation, circulating current deviation, and voltage frequency deviation based on the amplitude of the deviation components. The deviation components are then merged and calculated according to the evaluation weights to form parameter tuning evaluation values ​​that serve as feedback for reinforcement learning gains.

6. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step four specifically involves: Extract the droop coefficient and virtual impedance parameters of each three-phase inverter from the existing droop control parameters. Divide the droop coefficient according to the active power regulation direction and the reactive power regulation direction, and divide the virtual impedance parameters according to the resistance regulation direction and the reactance regulation direction. Arrange them according to the inverter identification to form the parameters to be tuned. The parameter tuning state variables are defined as the learning states of the improved Kiefer-Wolfowitz parameter tuning process, the parameter tuning evaluation values ​​are defined as the reinforcement learning benefit feedback of the improved Kiefer-Wolfowitz parameter tuning process, and the parameter tuning correlation between the learning states, the parameters to be tuned, and the reinforcement learning benefit feedback is established. Two-sided perturbation rules are configured around the parameter to be tuned. The two-sided perturbation rules are used to generate parameter perturbations with the same amplitude and opposite directions on the same parameter to be tuned, and provide generation rules for the positive perturbation parameter group and the negative perturbation parameter group. A stochastic approximation update rule is configured around the parameter to be tuned. The stochastic approximation update rule is used to calculate the gradient estimate based on the parameter tuning evaluation value corresponding to the positive perturbation parameter group and the parameter tuning evaluation value corresponding to the negative perturbation parameter group, and to generate candidate update parameters based on the gradient estimate. By merging parameter tuning correlations, two-sided perturbation rules, and random approximation update rules, an improved Kiefer-Wolfowitz parameter tuning process is established.

7. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step five specifically involves: Capacity identifiers are extracted from parallel operation data, and the rated capacity of each three-phase inverter is determined based on the capacity identifiers. The ratio of the rated capacity to the total rated capacity of the parallel inverters is determined as the inverter capacity weight. Arrange the parameters to be tuned according to the inverter identification, group the droop coefficient and virtual impedance parameters corresponding to the same three-phase inverter into the same parameter unit, and determine the parameter positions corresponding to the active power adjustment direction, reactive power adjustment direction, resistance adjustment direction and reactance adjustment direction within the parameter unit. Based on the disturbance scale, an initial disturbance component is configured for each parameter position. The initial disturbance component is then subjected to capacity-weighted correction in the same adjustment direction, so that the sum of the product of the disturbance component corresponding to each three-phase inverter and the inverter capacity weight is zero, thus forming a mirrored disturbance vector. The mirror perturbation vector is superimposed on the parameter to be tuned along the positive direction to form a positive perturbation parameter set, and the mirror perturbation vector is superimposed on the parameter to be tuned along the negative direction to form a negative perturbation parameter set. Establish a correspondence between the forward and reverse disturbance parameter sets and the parameter tuning state variables, so that the forward and reverse disturbance parameter sets enter the improved Kiefer-Wolfowitz parameter tuning process.

8. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step six specifically involves: An improved Kiefer-Wolfowitz parameter tuning process is adopted. Under the same parameter tuning state variable, the parallel droop control of the three-phase inverter is performed with the forward disturbance parameter group and the reverse disturbance parameter group respectively, to obtain the parameter tuning evaluation value corresponding to the forward disturbance parameter group and the parameter tuning evaluation value corresponding to the reverse disturbance parameter group. The difference between the parameter tuning evaluation values ​​corresponding to the positive disturbance parameter group and the parameter tuning evaluation values ​​corresponding to the negative disturbance parameter group is calculated to form an evaluation difference. The evaluation difference is then associated with the disturbance components in the mirror disturbance vector according to the parameter position. Based on the evaluation difference, disturbance scale, and disturbance component, the gradient estimates of the parameter to be tuned in the active power regulation direction, reactive power regulation direction, resistance regulation direction, and reactance regulation direction are calculated according to the Kiefer-Wolfowitz stochastic approximation rule. Based on the gradient estimate and update step size, the parameters to be tuned are adjusted to form candidate update parameters, which are then used as the objects of safe projection.

9. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step seven specifically involves: Voltage frequency boundary, output current boundary, circulating current boundary and parameter boundary are extracted from parallel operation data and parallel deviation state, and operation boundary constraints are formed based on voltage frequency boundary, output current boundary, circulating current boundary and parameter boundary; The candidate update parameters are compared against the boundary constraints according to the operational boundary constraints. The parameter components that fail the boundary comparison are mapped to the range limited by the operational boundary constraints, while the parameter components that pass the boundary comparison are retained as their original values, thus forming the projection result. Based on the projection results, a projection evaluation value is generated, and the projection evaluation value is compared with the parameter tuning evaluation value to form an evaluation change record; The evaluation hysteresis determination is performed on the evaluation change record according to the continuous control window. The evaluation hysteresis determination is used to determine whether the projection result maintains the state of decreasing parameter tuning evaluation value within the continuous control window. If the evaluation hysteresis judgment is passed, the projection result is determined as the target sagging control parameter; if the evaluation hysteresis judgment is not passed, the current sagging control parameter is determined as the target sagging control parameter.

10. The method for parallel droop control of a three-phase inverter based on reinforcement learning according to claim 1, characterized in that, Step eight specifically involves: The target droop control parameters are assigned according to the inverter identification, and the droop coefficient and virtual impedance parameters corresponding to each three-phase inverter are compiled into a target parameter allocation record. The power state is converted by using the droop coefficient in the target parameter allocation record, and the frequency reference and voltage reference of each three-phase inverter are calculated to form a droop control reference record. The virtual impedance parameters in the target parameter allocation record are used to perform impedance compensation conversion on the running sample data, calculate the virtual impedance compensation amount for each three-phase inverter, and synthesize the virtual impedance compensation amount with the droop control reference record to form a parallel control reference record. Based on the parallel control reference record, the modulation control quantity of each three-phase inverter is generated, and the parallel droop control of the three-phase inverter is executed according to the modulation control quantity to form a parallel operation control record. The power state and parallel deviation state are updated on the parallel operation control record. The target droop control parameters, the updated power state, and the updated parallel deviation state are then correlated and organized to generate the parallel droop control result.

Images