Virtual synchronous power compensation control method and system for grid-forming new energy grid-connected system

Abstract

The application relates to the technical field of new energy power systems, and discloses a virtual synchronous power compensation control method and system for a grid-connection type new energy grid-connection system. The method collects operation data of multiple parallel grid-connection type new energy converters in real time, and constructs a global state space representing power distribution, voltage and frequency operation conditions of the system. A trained soft actor-critic (SAC) deep reinforcement learning strategy network is used to output active and reactive compensation signals of each converter according to the global state space, and amplitude limiting constraints are implemented. The amplitude-limited compensation signals are embedded into active-frequency and reactive-voltage control loops of virtual synchronous generators of each converter, and a virtual synchronous control equation is updated. Converter modulation signals are generated based on the updated control equation to drive the operation of a device. The application can adaptively compensate for active and reactive power distribution deviations caused by line impedance mismatch, and improve the voltage and frequency stability level of the grid-connection type new energy grid-connection system.

Description

Technical Field

[0001] This invention relates to the field of new energy power system technology, specifically to a virtual synchronous power compensation control method and system for grid-connected new energy systems. Background Technology

[0002] With the increasing integration of new energy power generation units such as wind power, photovoltaics, and energy storage into the power system, the power system is gradually exhibiting a high proportion of power electronics characteristics. New energy power generation units are typically connected to the AC grid through power electronic converters, and their operating characteristics differ significantly from those of traditional synchronous generators. To improve the active support capability of new energy grid-connected systems in scenarios such as weak grids, islanded operation, and multi-converter parallel operation, grid-connected converter control technology has received widespread attention. Grid-connected converters can actively establish AC voltage amplitude, frequency, and phase angle, possessing voltage source characteristics, and are an important technical direction for supporting the stable operation of high-proportion new energy systems. Among them, virtual synchronous generator control is a typical grid-connected control method. This method simulates the rotor motion equations, virtual inertia, and damping characteristics of a synchronous generator, enabling the new energy converter to possess inertia support and frequency regulation capabilities similar to a synchronous generator.

[0003] In AC systems composed of multiple grid-connected renewable energy converters connected in parallel, each converter is typically connected to a common AC bus via filters and line impedance. Ideally, the multiple converters should share active and reactive loads according to their respective rated capacities. When the converter capacities are the same, the output active and reactive power of each converter should be as consistent as possible. However, in actual engineering projects, the line lengths, line resistances, and line reactances between each renewable energy converter and the common coupling point are usually not completely consistent. Line impedance mismatch alters the power transfer relationship between the converter output voltage and the common bus voltage, making it difficult to achieve accurate active and reactive power distribution even when multiple converters use the same virtual synchronization control parameters.

[0004] For AC systems primarily composed of inductive circuits, the active power output of the converter is mainly affected by the power angle and line reactance, while the reactive power output is mainly affected by the voltage amplitude difference and line reactance. Therefore, when the line reactance differs between different converters, the line reactance alters the ratio between active power and the power angle, and changes the equivalent droop characteristic of the reactive-voltage control loop, leading to deviations in both active and reactive power distribution. The reactive power distribution deviation is typically more pronounced.

[0005] Currently, existing methods for addressing power distribution problems caused by line impedance mismatch mainly include virtual impedance control, secondary control, distributed control, and adaptive control. Virtual impedance control improves power distribution by changing the equivalent output impedance of the converter; secondary control and distributed control achieve power coordination and voltage-frequency restoration through communication networks; and adaptive control improves system adaptability by adjusting control parameters.

[0006] In recent years, deep reinforcement learning methods have been increasingly applied to the control of microgrids and power electronic systems. The soft actor-critic deep reinforcement learning algorithm is a type of reinforcement learning algorithm suitable for continuous action spaces. It can learn the mapping relationship between states and actions through interaction with the environment, making it suitable for handling control problems with nonlinear, uncertain, and multi-condition characteristics.

[0007] The existing technology has at least the following shortcomings: 1. Although traditional virtual synchronous generator control can improve the grid connection capability and inertia support capability of converters, it cannot automatically eliminate the active and reactive power distribution deviations between converters under line impedance mismatch conditions.

[0008] 2. Virtual impedance methods typically require designing virtual impedance parameters based on line impedance, load level, and control objectives. When line impedance parameters are unknown, measurements are inaccurate, or system operating conditions change, fixed-parameter virtual impedance methods struggle to maintain good compensation performance.

[0009] 3. Excessive virtual impedance may cause additional voltage drop, affecting system voltage quality and potentially having an adverse impact on system dynamic performance and stability.

[0010] 4. Secondary control and distributed control usually rely on communication networks, which are easily affected by factors such as communication delays, data packet loss, network topology changes and single points of failure, resulting in high engineering implementation complexity.

[0011] 5. Existing power compensation methods based on deep reinforcement learning are mostly designed for traditional droop control. However, virtual synchronous generator control has virtual inertia, damping, and dynamic frequency regulation characteristics. Directly applying the compensation strategies in traditional droop control to virtual synchronous control systems is not entirely applicable.

[0012] 6. Existing methods lack a power compensation control method that can autonomously generate active power compensation signals and reactive power compensation signals based on the real-time operating status of multiple grid-type renewable energy converters without explicitly relying on line impedance parameters, and embed them into the virtual synchronous active-frequency loop and reactive-voltage loop respectively.

[0013] Therefore, there is an urgent need for a virtual synchronous power compensation control method applicable to grid-connected renewable energy systems to improve the accuracy of active and reactive power distribution under impedance mismatch conditions when multiple grid-connected renewable energy converters are operating in parallel. Summary of the Invention

[0014] In order to overcome the defects of the existing technology, the purpose of this invention is to provide a virtual synchronous power compensation control method and system for grid-connected new energy systems, so as to solve the technical problem of how to improve the active and reactive power distribution accuracy under impedance mismatch conditions when multiple grid-connected new energy converters are connected in parallel.

[0015] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a virtual synchronous power compensation control method for a grid-connected renewable energy system, comprising: Real-time acquisition of operating data from multiple parallel grid-connected new energy converters to construct a global state space reflecting the system's power distribution and voltage and frequency operating status; The training of the policy network is completed based on the preset soft actor-critic SAC deep reinforcement learning algorithm. According to the global state space, the active power compensation signal and reactive power compensation signal corresponding to each converter are output, and the amplitude of the compensation signal is constrained. The limited active power compensation signal and reactive power compensation signal are respectively embedded into the active power-frequency control loop and reactive power-voltage control loop of the virtual synchronous generator of each converter, and the virtual synchronous control equation is updated. The converter modulation signal is generated based on the compensated virtual synchronous control equation to drive the converter to operate, compensate for the active and reactive power distribution deviation caused by line impedance mismatch, and complete the virtual synchronous power compensation control work of the grid-connected new energy system.

[0016] Preferably, the grid-type new energy converter includes a wind power converter, a photovoltaic inverter, and an energy storage converter, and each of the converters is connected to a common AC bus through a filter and line impedance. The formula for calculating the line impedance of the i-th converter is defined as follows:

[0017] in, Let be the line resistance between the i-th converter and the common AC bus. Let be the line reactance between the i-th converter and the common AC bus, i=1,2,...,N.

[0018] Preferably, the global state space is composed of spliced ​​local states of each converter; The local state of the i-th converter at time k is as follows:

[0019] in, Let be the output active power of the i-th converter at time k. Let be the output reactive power of the i-th converter at time k; Voltage deviation is defined as:

[0020] Frequency deviation is defined as:

[0021] in, For voltage deviation, This is for frequency deviation; and These are the rated voltage amplitude and the rated angular frequency, respectively. The global state space is:

[0022] Right now: .

[0023] Preferably, the action space of the preset soft actor-critic SAC deep reinforcement learning algorithm is the set of power compensation signals for each converter, defined as:

[0024] in, Let be the active power compensation signal of the i-th converter at time k. Let be the reactive power compensation signal of the i-th converter at time k; The compensation signal amplitude limiting constraint condition is as follows:

[0025]

[0026] in, , , and It can be determined based on the rated capacity and allowable adjustment range of the converter.

[0027] The preferred, compensated active-frequency control equation is:

[0028] The active-frequency control equation under steady state simplifies to:

[0029] in, For virtual rotational inertia, The damping coefficient is... =1 / The equivalent active power-frequency droop factor. This is a reference value for active power.

[0030] The preferred reactive power-voltage control equation after compensation is:

[0031] in, This is the reactive power-voltage droop factor. This is a reference value for reactive power. Let be the voltage amplitude at the terminal of the i-th converter.

[0032] Preferably, the preset soft actor-critic SAC deep reinforcement learning algorithm is optimized through a preset reward function, wherein the reward function is:

[0033] in, , , , and These are the weighting coefficients; and Used to constrain voltage and frequency deviations; Used to limit the amplitude of compensation actions; For active power distribution error, For reactive power distribution error, For voltage deviation error, For frequency deviation error, This is a penalty item for actions.

[0034] Furthermore, the expressions for the active power allocation error, reactive power allocation error, voltage deviation, frequency deviation, and action penalty term are as follows: Active power distribution error is defined as:

[0035] Reactive power distribution error is defined as:

[0036] Voltage deviation is defined as:

[0037] Frequency deviation is defined as:

[0038] The action penalty term is used to limit the amplitude of the compensation signal. The action penalty term is defined as follows:

[0039] in, and Let be the rated active capacity and rated reactive capacity of the i-th converter, respectively. and Let K represent the total active load and total reactive load of the system at time k, respectively.

[0040] Preferably, the SAC algorithm adopts an offline training and online inference mode; During the offline training phase, a simulation environment was built that included line impedance mismatch, load fluctuation, and load change conditions. The policy network and value network were iteratively updated through state, action, and reward value to learn the optimal compensation strategy and save the network parameters. During the online inference phase, the network parameters are fixed after training, and the compensation signal is output only through forward computation of the network, without a complex online optimization solution process.

[0041] Secondly, the present invention also provides a virtual synchronous power compensation control system for a grid-connected renewable energy system, used to implement the aforementioned virtual synchronous power compensation control method for a grid-connected renewable energy system, comprising: The state construction module is used to collect the operating data of multiple parallel grid-connected new energy converters in real time and construct a global state space that reflects the system's power distribution and voltage and frequency operating status. The compensation output module is used to complete the training of the policy network based on the preset soft actor-critic SAC deep reinforcement learning algorithm, output the active power compensation signal and reactive power compensation signal corresponding to each converter according to the global state space, and perform amplitude limiting constraint on the compensation signal. The control correction module is used to embed the limited active power compensation signal and reactive power compensation signal into the active power-frequency control loop and reactive power-voltage control loop of each converter's virtual synchronous generator, respectively, and update the virtual synchronous control equation. The drive module is used to generate converter modulation signals based on the compensated virtual synchronous control equations, drive the converter to operate, compensate for the active and reactive power distribution deviations caused by line impedance mismatch, and complete the virtual synchronous power compensation control work of the grid-connected new energy system.

[0042] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a virtual synchronous power compensation control method for grid-connected renewable energy systems. It constructs a global state space for the system by collecting real-time operating data from multiple parallel grid-connected renewable energy converters. Using a trained SAC deep reinforcement learning strategy network, it adaptively generates active and reactive power compensation signals and performs amplitude limiting. These compensation signals are then embedded into a virtual synchronous generator dual-control loop to update the control equations. Finally, the updated control logic drives the converter operation, effectively solving the core technical problem of poor power distribution accuracy caused by line impedance mismatch in multi-converter parallel operation. Compared to traditional virtual synchronous control schemes, this invention can adaptively correct the power reference values ​​of each converter, significantly reducing the active and reactive power distribution deviations caused by line impedance differences, and significantly improving the power distribution accuracy of multi-converter parallel systems.

[0043] Furthermore, this invention eliminates the need for online precise measurement and identification of line impedance parameters, thus avoiding the dependence of traditional virtual impedance compensation methods on accurate line impedance parameters. It also avoids compensation failure issues caused by parameter measurement errors and operating condition fluctuations, greatly improving the engineering applicability of the solution. Unlike existing reinforcement learning compensation strategies that only adapt to traditional droop control, this invention deeply integrates the compensation mechanism into the virtual synchronous generator control framework, perfectly adapting to the unique virtual inertia, damping characteristics, and dynamic frequency and voltage regulation characteristics of virtual synchronous control, resulting in stronger adaptability and compatibility.

[0044] Furthermore, this invention simultaneously sets active and reactive power compensation action dimensions, enabling it to simultaneously correct the active power angle offset problem caused by impedance mismatch and the reactive power voltage equivalent droop characteristic offset problem, thus achieving synchronous optimization of the two types of power distribution deviations. The composite reward function designed in this invention takes into account power distribution error, system voltage and frequency deviation, and compensation action amplitude constraints, allowing the agent to adaptively optimize the power distribution effect while strictly ensuring the voltage and frequency stability of the grid-connected system and avoiding system fluctuations caused by excessive compensation signal amplitude.

[0045] Furthermore, the system adopts an offline training and online inference operation mode. During the online phase, only state acquisition and network forward computation are required, eliminating the need for complex model solving and iterative optimization. This results in low computational complexity and fast response speed. Simultaneously, because the offline training phase covers various line impedance parameters, load steady-state conditions, and sudden changes, the agent learns optimal compensation strategies for multiple scenarios. This allows the system to adapt to dynamic load changes and quickly converge to a precise power-sharing state after load fluctuations, effectively improving the operational stability and adaptability of the new energy grid-connected system. Attached Figure Description

[0046] Figure 1 This is a flowchart of the virtual synchronous power compensation control method for a grid-connected new energy system in an embodiment of the present invention; Figure 2 This is a diagram of a parallel AC system of multi-grid converters in an embodiment of the present invention; Figure 3 This is a flowchart of the SAC virtual synchronous power compensation control in an embodiment of the present invention; Figure 4 A comparison graph showing the power, frequency, and voltage before and after using the method of the present invention; Figure 5 This is a schematic diagram illustrating the power distribution when the proposed method is used under load variation conditions in an embodiment of the present invention. Figure 6 This is a schematic diagram of the virtual synchronous power compensation control system for a grid-connected new energy system in an embodiment of the present invention. In the diagram: 1. State construction module; 2. Compensation output module; 3. Control correction module; 4. Run-drive module. Detailed Implementation

[0047] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] The purpose of this invention is to provide a virtual synchronous power compensation control method and system for grid-connected renewable energy systems. This method addresses scenarios where multiple grid-connected renewable energy converters operate in parallel. It generates active and reactive power compensation signals for each converter using a soft actor-commentator deep reinforcement learning agent. The active power compensation signal is added to the active-frequency control loop of the virtual synchronous generator, and the reactive power compensation signal is added to the reactive-voltage control loop of the virtual synchronous generator, thereby compensating for power distribution deviations caused by line impedance mismatch. This addresses the technical problem of improving the accuracy of active and reactive power distribution under impedance mismatch conditions when multiple grid-connected renewable energy converters operate in parallel.

[0049] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 See Figure 1In one embodiment of the present invention, a virtual synchronous power compensation control method for a grid-connected renewable energy system is provided. This method is applicable to scenarios where multiple grid-connected renewable energy converters are connected in parallel. It effectively solves the technical problem of low active and reactive power allocation accuracy in traditional virtual synchronous generator control under complex operating conditions such as line impedance mismatch, load fluctuations, and sudden load changes. Furthermore, it eliminates the need for precise identification of line impedance parameters, adaptation to the inherent inertia and damping characteristics of virtual synchronous control, reduces online computation, and improves system stability. The method employs an offline training and online inference working mode. Its core utilizes a soft actor-commentator (SAC) deep reinforcement learning algorithm to adaptively generate active and reactive power compensation signals and embeds a virtual synchronous dual control loop to achieve accurate power deviation compensation. Specifically, the core steps include: Step 1: Collect real-time operating data from multiple parallel grid-connected new energy converters to construct a global state space that reflects the system's power distribution and voltage and frequency operating status. Specifically, the grid-connected new energy converters adapted to this invention include various types of new energy grid-connected converters such as wind power converters, photovoltaic inverters, and energy storage converters. Each converter is connected to a common AC bus via a filter and line impedance, such as... Figure 2 As shown. For the i-th converter in the system, the line impedance expression from it to the common AC bus is:

[0050] in, For line resistance, Let i be the line reactance, i = 1, 2, ..., N, and N be the total number of parallel converters.

[0051] In a multi-converter parallel grid-connected system, the output voltage of the i-th converter is:

[0052] The common AC bus voltage is:

[0053] in, This represents the amplitude of the converter terminal voltage. This refers to the bus voltage amplitude. The power angle of the converter output voltage relative to the bus voltage; The complex power output of the converter can be expressed as:

[0054] in, , These represent the output active and reactive power, respectively.

[0055] The converter output current satisfies

[0056] The power system lines are mainly inductive lines, which meet the requirements of Much larger Under steady-state operating conditions, the converter's power angle is relatively small, which can be approximated as...

[0057]

[0058] At this point, the active and reactive power can be simplified as follows:

[0059]

[0060] As can be seen from the above power characteristics, active power is mainly determined by the power angle and line reactance, while reactive power is mainly determined by the voltage amplitude difference and line reactance.

[0061] In actual engineering, the line reactance of each converter differs, which can be expressed as:

[0062] in, For the reference line reactance, This is the reactance deviation.

[0063] This deviation will cause an additional power angle deviation:

[0064] Therefore, line reactance mismatch will change the ratio between active power and power angle, resulting in active power distribution deviation.

[0065] For reactive power, from:

[0066] We can obtain:

[0067] Substituting this into the reactive-voltage control equation of the virtual synchronous generator, it can be seen that the line reactance changes the equivalent droop coefficient of the reactive-voltage control loop. The equivalent droop coefficient can be expressed as:

[0068] in, Let be the reactive power-voltage droop coefficient of the i-th converter. When the line reactances of each converter are inconsistent, the equivalent reactive power-voltage regulation characteristics of each converter are different, making it difficult to accurately allocate reactive power according to the rated capacity.

[0069] To accurately characterize the real-time operating status of the system, this invention designs a hierarchical state space, which is obtained by splicing together the local states of multiple converters to obtain the global state space.

[0070] The local state of the i-th converter at time k is

[0071] in These represent the real-time active and reactive power outputs of the converter, respectively. Voltage deviation

[0072] Frequency deviation

[0073] in, , These are the system's rated voltage amplitude and rated angular frequency, respectively.

[0074] The system's global state space is a set of local states.

[0075] Right now: It provides complete coverage of power, voltage, and frequency operating information for all converters.

[0076] Step 2: The policy network is trained based on the preset soft actor-critic SAC deep reinforcement learning algorithm. The active power compensation signal and reactive power compensation signal corresponding to each converter are output according to the global state space, and the amplitude of the compensation signal is constrained. Specifically, the action space of the SAC algorithm of this invention corresponds to the power compensation amount of all converters, defined as follows:

[0077] in , These are the active and reactive power compensation signals for the i-th converter, respectively.

[0078] After the SAC agent outputs an action, the controller will Input the active-frequency control loop of the i-th converter, and... Input the reactive power-voltage control loop of the i-th converter to achieve online compensation for power distribution deviation.

[0079] To prevent excessive amplitude of the compensation signal from causing system voltage and frequency oscillations and to ensure operational stability, a hard amplitude limit is set for the compensation signal:

[0080]

[0081] in, , , and It can be determined based on the rated capacity and allowable adjustment range of the converter.

[0082] Step 3: Embed the limited active power compensation signal and reactive power compensation signal into the active power-frequency control loop and reactive power-voltage control loop of each converter's virtual synchronous generator, respectively, and update the virtual synchronous control equations. Specifically, this invention is based on an optimized design of the traditional virtual synchronous generator control framework. The active power-frequency control equation of the traditional virtual synchronous control is as follows:

[0083] Under steady-state conditions, we have:

[0084] Therefore, the active-frequency control equation can be written as:

[0085] in:

[0086] It is the equivalent active power-frequency droop coefficient.

[0087] The traditional reactive power-voltage control equation for the i-th converter is:

[0088] in, This is the rated voltage amplitude. This is the reactive power-voltage droop factor. This is a reference value for reactive power. For virtual rotational inertia, The damping coefficient is... =1 / The active power-frequency droop factor is... This is the reactive power droop factor. , These are the reference values ​​for active and reactive power, respectively. Traditional control systems lack a compensation mechanism and cannot offset power deviations caused by line impedance mismatch.

[0089] according to Figure 3 As shown, this embodiment introduces a SAC agent to generate active power compensation signals and reactive power compensation signals. For the i-th grid-connected renewable energy converter, the SAC agent outputs:

[0090] in, This is the active power compensation signal for the i-th converter. This is the reactive power compensation signal for the i-th converter.

[0091] Will The active-frequency control loop of the virtual synchronous generator added to the i-th converter is as follows:

[0092] Steady state simplifies to

[0093] The updated reactive power-voltage control equation is as follows:

[0094] By using compensation signals to equivalently correct the real-time power reference value of the converter, the power angle deviation and equivalent droop coefficient deviation caused by line impedance mismatch are offset from the control source, and the active and reactive power distribution deviations are compensated and corrected simultaneously.

[0095] Step 4: Generate converter modulation signal based on the compensated virtual synchronous control equation, drive converter operation, compensate for active and reactive power distribution deviation caused by line impedance mismatch, and complete the virtual synchronous power compensation control of grid-connected new energy system.

[0096] Specifically, to ensure the convergence of SAC agent training and optimal control performance, this invention designs a multi-constraint composite reward function that balances power allocation accuracy, system voltage and frequency stability, and the rationality of compensation actions. The overall expression of the reward function is as follows:

[0097] The active power allocation error is as follows:

[0098] Reactive power distribution error is used to quantify power distribution deviation, and the formula is:

[0099] The voltage deviation error is:

[0100] Frequency deviation error is used to constrain the steady-state performance of the system, and the formula is:

[0101] The action penalty term is used to limit the amplitude of the compensation signal from being too large. The formula is:

[0102] in, , , , and These are weighting coefficients. Since this invention focuses on power allocation accuracy, therefore... and It can be set to a larger value; and Used to constrain voltage and frequency deviations; Used to limit the amplitude of compensation actions. and Let be the rated active capacity and rated reactive capacity of the i-th converter, respectively. and Let K be the total active load and total reactive load of the system at time k, respectively. When the rated capacity of each converter is the same, the power distribution error can be simplified to the deviation between the output power of each converter and the average load power of the system.

[0103] When the rated capacity of each grid-type new energy converter is the same, the active power distribution error and reactive power distribution error can be simplified to the deviation between the output power of each converter and the average load power.

[0104] This embodiment employs a hierarchical working mechanism of offline training and online inference. During the offline training phase, a simulation environment covering multiple operating conditions is built, including various scenarios such as different line resistances, different line reactances, steady-state loads, and load surges, simulating the impedance mismatch and dynamic fluctuations in actual engineering conditions. During training, the SAC agent outputs compensation actions based on the real-time state s_k. ,Right now:

[0105] Compensation action After being applied to the virtual synchronization control loop, the system update status is collected and the real-time reward value is calculated. And obtain the state at the next moment. To form an empirical sample:

[0106] By iteratively updating the policy network and value network through experience replay and mini-batch gradient descent, the optimal power compensation strategy under different operating conditions is continuously learned, and the policy network parameters are solidified after training.

[0107] During the online inference phase, the parameters of the SAC policy network remain unchanged after training. The controller collects the output active power, output reactive power, terminal voltage, and output frequency of each grid-type renewable energy converter in real time to construct the current state s_k.

[0108] State Input the trained SAC policy network to obtain the compensation action:

[0109] Right now:

[0110] in, This represents the SAC policy network after training is complete.

[0111] Then, the compensation signal is subjected to amplitude limiting, and the amplitude-limited signal is then... and The active-frequency control loop and reactive-voltage control loop are respectively added to the i-th converter. Thus, the converter controller generates modulation signals based on the compensated frequency and voltage commands, driving the power switching devices to operate and achieving power distribution compensation under line impedance mismatch conditions.

[0112] The online phase only requires forward computation of the policy network, without the need to solve complex optimization problems online or explicitly obtain line impedance parameters. Therefore, it has low online computational complexity and good engineering applicability.

[0113] This embodiment includes three grid-connected renewable energy converters. All three converters have the same rated capacity, are controlled by a virtual synchronous generator, and are connected to a common AC bus through different line impedances. The system's rated frequency is 50Hz, and the rated line voltage RMS value is 400V.

[0114] The line impedances of the three converters are set to be inconsistent to simulate the line impedance mismatch situation in actual engineering. Under the condition of line impedance mismatch, traditional virtual synchronous control will cause a significant difference in the output active power and reactive power of the three converters, with the reactive power distribution deviation being more obvious.

[0115] After adopting the method of this invention, the SAC agent generates corresponding active power compensation signals and reactive power compensation signals based on the active power, reactive power, voltage deviation, and frequency deviation of the three converters. The active power compensation signal is used to correct the active power-frequency control loop, and the reactive power compensation signal is used to correct the reactive power-voltage control loop.

[0116] like Figure 4 As shown in the simulation results, under traditional virtual synchronous control, the active and reactive power of the three converters cannot be completely evenly distributed; after implementing the SAC virtual synchronous power compensation control of this invention, the active and reactive power of the three converters gradually become consistent. Meanwhile, the system frequency and terminal voltage remain near their rated values, without significant oscillations or exceeding limits.

[0117] like Figure 5As shown, under the condition of sudden load change, the load change will cause transient power fluctuations. However, after the system enters a new steady state, the output power of the three converters can still maintain good consistency, indicating that the method of the present invention has a certain adaptability to operating conditions.

[0118] This embodiment can improve the accuracy of active and reactive power distribution when multiple grid-type new energy converters are operating in parallel.

[0119] Traditional virtual synchronous control can lead to power distribution deviations under line impedance mismatch conditions. This invention addresses this by outputting active and reactive power compensation signals through the SAC intelligent agent, embedding virtual synchronous active-frequency and reactive-voltage loops respectively, enabling adaptive correction of the power reference values ​​for each converter, thereby reducing power distribution deviations.

[0120] In this embodiment, it is not necessary to obtain the line impedance parameters accurately online.

[0121] Traditional compensation methods, such as virtual impedance, typically rely on line impedance parameter design, but in actual engineering projects, line impedance may be difficult to measure accurately. This invention utilizes a SAC intelligent agent to generate a power compensation signal based on real-time operating status, eliminating the need for explicit measurement or identification of line impedance parameters and improving engineering applicability.

[0122] This embodiment applies to the virtual synchronous generator control framework.

[0123] Existing deep reinforcement learning power compensation methods are mainly geared towards traditional droop control. This invention directly embeds the compensation signal into the active-frequency control loop and reactive-voltage control loop of the virtual synchronous generator, which can adapt to the inertia, damping, and dynamic power regulation characteristics in virtual synchronous control.

[0124] This embodiment can improve both active power distribution and reactive power distribution simultaneously.

[0125] Line impedance mismatch not only affects the relationship between active power and power angle, but also alters the equivalent droop characteristics of the reactive power-voltage control loop. The operating space of this invention includes both active power compensation signals and reactive power compensation signals, thus enabling simultaneous compensation for active and reactive power distribution deviations.

[0126] This embodiment can maintain voltage and frequency stability while improving power distribution accuracy.

[0127] The reward function in this embodiment includes not only active power allocation error and reactive power allocation error, but also voltage deviation, frequency deviation, and action penalty terms. Therefore, when learning the power compensation strategy, the SAC agent will simultaneously consider power equalization, voltage stability, frequency stability, and compensation amplitude limitations.

[0128] This embodiment has low online computational complexity.

[0129] This embodiment employs an offline training and online inference approach. During online operation, the controller only needs to collect system status data and perform forward computation of the policy network to obtain the compensation signal; it does not require online execution of complex model solving or optimization calculations.

[0130] This embodiment is adaptable to load changes.

[0131] During the offline training phase, by setting different load levels and load abrupt changes, the SAC agent can learn compensation strategies under different operating conditions. Therefore, after a load change, the system can reconverge to a new power-sharing state.

[0132] In this embodiment, the SAC agent constructs the input state based on the active power, reactive power, voltage deviation, and frequency deviation of multiple grid-connected renewable energy converters. The SAC agent outputs active power compensation signals and reactive power compensation signals corresponding to each grid-connected renewable energy converter. The active power compensation signal is added to the active-frequency control loop of the virtual synchronous generator to equivalently correct the active power reference value. The reactive power compensation signal is added to the reactive-voltage control loop of the virtual synchronous generator to equivalently correct the reactive power reference value. The SAC reward function is constructed based on active power allocation error, reactive power allocation error, voltage deviation, frequency deviation, and compensation action penalty term. An offline training and online inference control method is adopted, enabling the generation of power compensation signals without explicitly obtaining line impedance parameters during the online phase. Amplitude constraints on the compensation signal prevent excessively large compensation amplitudes, thus balancing power allocation accuracy and system stability. This method is applicable to scenarios where multiple grid-connected renewable energy converters such as wind power, photovoltaic, and energy storage are operated in parallel.

[0133] Example 2 according to Figure 6 As shown, this embodiment also provides a virtual synchronous power compensation control system for a grid-connected new energy system, used to fully execute the control method described in Embodiment 1, solving technical problems such as poor power distribution accuracy, weak system adaptability, reliance on precise parameters in traditional control, and inability to adapt to virtual synchronous control characteristics caused by impedance mismatch in parallel lines of multiple converters. The system includes a state construction module, a compensation output module, a control correction module, and a running drive module. These modules work collaboratively without redundancy, completely covering the entire technical process of the method. The specific module functions are as follows: State construction module 1 is used to collect real-time operating data of multiple parallel grid-connected new energy converters, including the output active power, reactive power, terminal voltage, and output angular frequency of each converter. It calculates the voltage deviation and frequency deviation of each converter and constructs a global state space that can completely reflect the system's power distribution, voltage, and frequency operating status based on the splicing of the local operating status of each converter, providing accurate state input basis for subsequent compensation decisions.

[0134] The compensation output module 2 is used to load the trained SAC deep reinforcement learning policy network, receive the global state space output by the state construction module, and output the active power compensation signal and reactive power compensation signal corresponding to all converters through forward calculation of the network. At the same time, it performs amplitude limiting constraints on all compensation signals according to preset limits to avoid system instability caused by excessive compensation amplitude and output compliant power compensation signals.

[0135] The control correction module 3 is used to receive the limited active and reactive power compensation signals output by the compensation output module, embed the active power compensation signal into the active power-frequency control loop of the virtual synchronous generator of each converter, embed the reactive power compensation signal into the reactive power-voltage control loop, update and optimize the dynamic control equation and steady-state control equation of the virtual synchronous generator, and equivalently correct the power reference value of each converter to offset the power transmission characteristic deviation caused by line impedance mismatch.

[0136] The running drive module 4 is used to generate converter power modulation signals based on the updated virtual synchronous control equations of the control correction module, drive the converter power switching devices to operate stably, compensate for the active and reactive power distribution deviations caused by line impedance mismatch in real time, realize accurate power distribution in the multi-converter parallel system, and complete the entire process of virtual synchronous power compensation control for grid-connected new energy systems.

[0137] This system relies on the autonomous learning capability of the SAC deep reinforcement learning algorithm, without relying on precise measurement parameters of line impedance. It is compatible with various grid-connected converter parallel scenarios such as wind power, photovoltaic, and energy storage, and is compatible with the virtual inertia, damping, and dynamic frequency and voltage regulation characteristics of virtual synchronous generators. At the same time, through multi-dimensional reward constraints and compensation limiting mechanisms, it improves the accuracy of power distribution while fully ensuring the voltage and frequency stability of the grid-connected system. Moreover, it has low online operation computation and strong adaptability to operating conditions, and has extremely high engineering implementation value.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A virtual synchronous power compensation control method for a grid-connected new energy system, characterized in that, include: Real-time acquisition of operating data from multiple parallel grid-connected new energy converters to construct a global state space reflecting the system's power distribution and voltage and frequency operating status; The training of the policy network is completed based on the preset soft actor-critic SAC deep reinforcement learning algorithm. According to the global state space, the active power compensation signal and reactive power compensation signal corresponding to each converter are output, and the amplitude of the compensation signal is constrained. The limited active power compensation signal and reactive power compensation signal are embedded into the active power-frequency control loop and reactive power-voltage control loop of the virtual synchronous generator of each converter, respectively, and the virtual synchronous control equation is updated. The converter modulation signal is generated based on the compensated virtual synchronous control equation to drive the converter to operate, compensate for the active and reactive power distribution deviation caused by line impedance mismatch, and complete the virtual synchronous power compensation control work of the grid-connected new energy system.

2. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The grid-type new energy converter includes a wind power converter, a photovoltaic inverter, and an energy storage converter. Each of the converters is connected to a common AC bus through a filter and line impedance. The formula for calculating the line impedance of the i-th converter is defined as follows: in, Let be the line resistance between the i-th converter and the common AC bus. Let be the line reactance between the i-th converter and the common AC bus, i=1,2,...,N.

3. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The global state space is composed of the splicing of the local states of each converter; The local state of the i-th converter at time k is as follows: in, Let be the output active power of the i-th converter at time k. Let be the output reactive power of the i-th converter at time k; Voltage deviation is defined as: Frequency deviation is defined as: in, For voltage deviation, This is for frequency deviation; and These are the rated voltage amplitude and the rated angular frequency, respectively. The global state space is: Right now: 。 4. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The action space of the preset soft actor-critic SAC deep reinforcement learning algorithm is the set of power compensation signals for each converter, defined as: in, Let be the active power compensation signal of the i-th converter at time k. Let be the reactive power compensation signal of the i-th converter at time k; The compensation signal amplitude limiting constraint condition is as follows: in, , , and It can be determined based on the rated capacity and allowable adjustment range of the converter.

5. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The compensated active-frequency control equation is as follows: The active-frequency control equation under steady state simplifies to: in, For virtual rotational inertia, The damping coefficient is... =1 / The equivalent active power-frequency droop factor. This is a reference value for active power.

6. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The compensated reactive power-voltage control equation is as follows: in, This is the reactive power-voltage droop factor. This is a reference value for reactive power. Let be the voltage amplitude at the terminal of the i-th converter.

7. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The pre-defined soft actor-critic SAC deep reinforcement learning algorithm is trained using a pre-defined reward function, which is: in, , , , and These are the weighting coefficients; and Used to constrain voltage and frequency deviations; Used to limit the amplitude of compensation actions; For active power distribution error, For reactive power distribution error, For voltage deviation error, For frequency deviation error, This is a penalty item for actions.

8. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 7, characterized in that, The expressions for the active power distribution error, reactive power distribution error, voltage deviation, frequency deviation, and action penalty term are as follows: Active power distribution error is defined as: Reactive power distribution error is defined as: Voltage deviation is defined as: Frequency deviation is defined as: The action penalty term is used to limit the amplitude of the compensation signal. The action penalty term is defined as follows: in, and Let be the rated active capacity and rated reactive capacity of the i-th converter, respectively. and Let K represent the total active load and total reactive load of the system at time k, respectively.

9. The virtual synchronous power compensation control method for a grid-connected new energy system according to claim 1, characterized in that, The SAC algorithm employs an offline training and online inference mode. During the offline training phase, a simulation environment was built that included line impedance mismatch, load fluctuation, and load change conditions. The policy network and value network were iteratively updated through state, action, and reward value to learn the optimal compensation strategy and save the network parameters. During the online inference phase, the network parameters are fixed after training, and the compensation signal is output only through forward computation of the network, without a complex online optimization solution process.

10. A virtual synchronous power compensation control system for a grid-connected new energy system, characterized in that, A method for implementing the virtual synchronous power compensation control of a grid-connected renewable energy system as described in any one of claims 1-9 includes: The state construction module is used to collect the operating data of multiple parallel grid-connected new energy converters in real time and construct a global state space that reflects the system's power distribution and voltage and frequency operating status. The compensation output module is used to complete the training of the policy network based on the preset soft actor-critic SAC deep reinforcement learning algorithm, output the active power compensation signal and reactive power compensation signal corresponding to each converter according to the global state space, and perform amplitude limiting constraint on the compensation signal. The control correction module is used to embed the limited active power compensation signal and reactive power compensation signal into the active power-frequency control loop and reactive power-voltage control loop of each converter's virtual synchronous generator, respectively, and update the virtual synchronous control equation. The drive module is used to generate converter modulation signals based on the compensated virtual synchronous control equations, drive the converter to operate, compensate for the active and reactive power distribution deviations caused by line impedance mismatch, and complete the virtual synchronous power compensation control work of the grid-connected new energy system.

Images