Power grid transient voltage support SVG dynamic response system based on ai reinforcement learning

Abstract

The present application relates to the field of power grid voltage stabilization technology, and particularly relates to a power grid transient voltage support SVG dynamic response system based on AI reinforcement learning, comprising: a virtual synchronous machine control module, used for controlling a static var generator to present an external characteristic of a voltage source; a signal acquisition module, used for acquiring original data from a target power grid and the static var generator; a reinforcement learning online decision module, used for receiving a system state vector, outputting a virtual inertia adjustment amount and a virtual damping coefficient adjustment amount, so as to online adjust parameters of the virtual synchronous machine control module; a power grid state detection module, used for identifying power grid strength by real-time estimation of a short-circuit ratio of a grid connection point of the target power grid, and real-time detection of a wind power access state and a wind power output fluctuation strength of the grid connection point; a mode switching and management module, used for determining a current operation mode, issuing a basic virtual inertia and a basic virtual damping coefficient corresponding to the current operation mode, and an adjustment amount allowable range corresponding to the current operation mode.

Description

Technical Field

[0001] This invention relates to the field of power grid voltage stabilization technology, and in particular to a power grid transient voltage support SVG dynamic response system based on AI reinforcement learning. Background Technology

[0002] Static var generators (SVGs) are widely used dynamic reactive power compensation devices in power systems, playing a crucial role in supporting transient voltage in the power grid. Traditional SVGs typically employ a grid-tracking control architecture, using a phase-locked loop (PLL) to track the grid voltage phase and inject reactive current into the grid as a controlled current source. However, grid-tracking control has the following inherent drawbacks: under weak grid conditions (low short-circuit ratio), the PLL is prone to dynamic interaction with grid impedance, leading to low-frequency oscillations or even instability; during severe voltage dips, the reactive power output of the SVG decreases linearly with the voltage, significantly weakening its support capability.

[0003] To address these issues, grid-based control technology has emerged. Grid-based SVG simulates the external characteristics of a synchronous generator, autonomously establishing the amplitude and phase of its internal electromotive force through virtual synchronous machine control, thus functioning as a controllable voltage source. Grid-based control possesses inherent stability under weak grid conditions and can provide stronger reactive power support at low voltages. However, grid-based control introduces new problems: fixed virtual inertia and virtual damping parameters are difficult to adapt to the demands of full-condition operation, potentially leading to transient instability of the power angle under large disturbances, and low-frequency oscillations may occur after the disturbance due to insufficient damping; under strong grid conditions, the voltage source characteristics of grid-based control may result in excessive fault current or high-frequency oscillations due to interaction with the grid.

[0004] Furthermore, with the large-scale integration of new energy sources such as wind power, the dynamic characteristics of SVG grid connection points have become more complex. The random fluctuations in wind power output are concentrated in the 0.1-5Hz frequency band, which, combined with transient grid faults, makes it difficult for fixed-parameter control strategies to simultaneously address the dual needs of smoothing normal fluctuations and supporting transient faults. The optimal control parameters and optimal control modes for SVG differ significantly under different grid strengths and wind power integration conditions. Under strong grid conditions, grid-connected current source control is more efficient and safer; under weak grid conditions, grid-connected voltage source control is more stable and powerful. However, grid strength is not a constant parameter; line switching, turbine start-up and shutdown, and changes in output from adjacent wind farms can all cause significant changes in the short-circuit ratio at the grid connection point within seconds to minutes, making it impossible for a single control mode to maintain optimal performance throughout the entire timeframe.

[0005] Therefore, there is an urgent need for an SVG dynamic response system that can automatically match the grid strength and wind power access status and smoothly switch control modes between strong and weak grids. Summary of the Invention

[0006] This invention solves the problem of insufficient adaptability of fixed parameter control under all operating conditions in the prior art by dynamically adjusting the virtual inertia parameter and virtual damping coefficient parameter through a reinforcement learning online decision-making module.

[0007] The technical solution proposed in this invention is: a dynamic response system for grid transient voltage support SVG based on AI reinforcement learning, comprising:

[0008] The virtual synchronous machine control module is used to control the static var generator to exhibit voltage source characteristics.

[0009] The signal acquisition module is used to acquire raw data from the target power grid and static var generator in each control cycle and construct the system state vector.

[0010] The reinforcement learning online decision-making module includes a pre-trained policy network, which receives the system state vector and outputs virtual inertia adjustment and virtual damping coefficient adjustment to adjust the parameters of the virtual synchronous machine control module online.

[0011] The power grid status detection module is used to identify the power grid strength by estimating the short-circuit ratio of the target power grid connection point in real time, and to detect the wind power access status and wind power output fluctuation intensity of the connection point in real time.

[0012] The mode switching and management module is used to determine the current operating mode from a variety of preset operating modes based on the short-circuit ratio and wind power access status, and to send the basic virtual inertia and basic virtual damping coefficient corresponding to the current operating mode to the virtual synchronous machine control module, and to send the allowable range of adjustment corresponding to the current operating mode to the reinforcement learning online decision-making module.

[0013] Preferably, the raw data includes the output angular frequency, output electromagnetic power, output current amplitude, grid connection point voltage amplitude, and grid connection point wind power active power of the static var generator (SVG).

[0014] Based on the original data, derived characteristics are calculated, including the frequency deviation between the output angular frequency and the grid rated frequency, the rate of change of the frequency deviation, the power deviation between the active power reference value and the output electromagnetic power of the static var generator, and the active power fluctuation amplitude of the wind farm.

[0015] After normalizing the output current amplitude, grid connection point voltage amplitude, grid connection point wind power active power and derived characteristics, the system state vector is obtained.

[0016] The pre-trained policy network is a pre-trained soft actor critic SAC algorithm network, which includes an input layer, a hidden layer, and an output layer.

[0017] The input layer comprises 7 neurons, corresponding to each element of the system state vector; the hidden layer comprises 2-3 fully connected layers with ReLU activation function; the output layer comprises 4 neurons, corresponding to the action mean and the action mean standard deviation, respectively, wherein the action mean is the mean of the virtual inertia adjustment and the virtual damping coefficient adjustment.

[0018] The virtual synchronous machine control module includes an active-frequency control model and a reactive-voltage control model;

[0019] The active-frequency control model is constructed based on the rotor equations of a simulated synchronous generator, including virtual inertia parameters. and virtual damping coefficient parameters The equation is:

[0020] ;in, Output angular frequency for SVG. The rated angular frequency of the power grid. This is a reference value for active power. δ represents the electromagnetic power output of the SVG, and δ is the virtual power angle. From the primary frequency modulation characteristics Given, where As a meritorious work point, This is the frequency droop coefficient.

[0021] The reactive-voltage control model is used to generate the internal potential amplitude. The formula is:

[0022] ; To control the cycle length;

[0023] in, The set value for the no-load internal potential. This is the reactive power droop factor. This is a reactive power command. Output reactive power to SVG This is the reference value for the grid connection point voltage. This is the measured value of the voltage amplitude at the grid connection point. and These are the proportional coefficient and integral coefficient for voltage deviation adjustment, respectively.

[0024] Preferably, the mode switching and management module is further configured to: determine multiple operating modes based on the short-circuit ratio and wind power access status, the operating modes including: a first mode, a second mode, a third mode, a fourth mode, a fifth mode, and a sixth mode;

[0025] The first mode is characterized by a weak power grid and no wind power access; the second mode is characterized by a strong power grid and no wind power access; the third mode is characterized by a weak power grid and wind power access; and the fourth mode is characterized by a strong power grid and wind power access.

[0026] The fifth mode is the occurrence of a special wind power event; the sixth mode is the duration of a low-level confidence level in grid strength identification exceeding the maximum permissible time.

[0027] Preferably, the mode switching and management module is further configured to: switch from the current mode among the first to fourth modes to the fifth mode when a special wind power event is detected;

[0028] When the duration of the power grid strength identification confidence level being low exceeds the maximum allowable time, the current mode from the first to the fourth modes is switched to the sixth mode;

[0029] In the fifth mode, the upper limit of the allowable range of adjustment of the reinforcement learning online decision-making module is increased;

[0030] In the sixth mode, the current operating mode is frozen, mode switching is refused, and the allowable range of adjustment of the parameters of the reinforcement learning online decision-making module is narrowed.

[0031] Preferably, the mode switching and management module includes an adaptive confirmation time generator;

[0032] The adaptive confirmation time generator is used to dynamically calculate the confirmation time based on the rate of change of the short-circuit ratio, the confidence level of the grid strength identification, and the fluctuation range of wind power processing.

[0033] When the current operating mode is inconsistent with the target operating mode and the duration reaches the confirmation time of the dynamic calculation, the mode switching and management module triggers mode switching.

[0034] Preferably, the dynamic calculation of the confirmation time includes:

[0035] The basic confirmation time is calculated based on the rate of change of the short-circuit ratio. The basic confirmation time decreases monotonically as the rate of change increases, and has a preset minimum confirmation time and a preset maximum confirmation time.

[0036] An acknowledgment time correction coefficient is generated based on the identification confidence level of the power grid strength, and the acknowledgment time correction coefficient increases as the confidence level decreases;

[0037] A wind power fluctuation correction coefficient is generated based on the wind power output fluctuation intensity. The wind power fluctuation correction coefficient is greater than 1 when the fluctuation intensity is lower than a first threshold and less than 1 when the fluctuation intensity is higher than a second threshold.

[0038] The dynamic confirmation time is obtained by multiplying the basic confirmation time by the confirmation time correction factor and the wind power fluctuation correction factor.

[0039] Preferably, the mode switching and management module is further configured to: in the fifth mode, adaptively adjust the confirmation time for exiting the fifth mode according to the intensity of active power fluctuation in the wind farm;

[0040] Within the medium power grid intensity range, discrete mode switching is not triggered. Instead, the basic virtual inertia and the basic virtual damping coefficient are generated by continuous interpolation between the parameter sets of adjacent modes based on the short-circuit ratio.

[0041] Preferably, it also includes a motion post-processing and safety limiting module;

[0042] The post-processing and safety limiting module is used to: receive the original adjustment amount output by the policy network, limit the original adjustment amount according to the allowable range of adjustment amount corresponding to the current operating mode issued by the mode switching and management module, and then output it to the virtual synchronizer control module.

[0043] Preferably, the power grid state detection module uses a recursive least squares method with a forgetting factor to estimate the equivalent impedance of the power grid, and performs recursive updates only when a valid excitation condition is detected.

[0044] The wind power access status detection module quantifies the output fluctuation intensity by analyzing the mean and peak power change of the wind farm's active power, and determines the wind power access status based on the mean, standard deviation, and peak power change of the wind farm's active power. The wind power access status is used by the mode switching and management module to adjust the operation mode judgment logic and generate pre-biases related to the operating conditions.

[0045] A computer-readable storage medium storing a computer program that is executed by a processor to implement the AI ​​reinforcement learning-based grid transient voltage support SVG dynamic response system.

[0046] The beneficial effects of this invention are:

[0047] 1. This invention, through a reinforcement learning online decision-making module, can dynamically adjust the virtual inertia parameters and virtual damping coefficient parameters according to the real-time power grid status, thus solving the problem of insufficient adaptability of fixed parameter control under complex operating conditions.

[0048] 2. This invention defines six operating modes and ensures reliable switching between multiple modes, such as strong / weak grid and wind power presence / absence, through dynamic confirmation time. Each time a new operating mode is switched, a parameter set corresponding to the current operating mode is sent to the virtual synchronous machine control module, enabling the SVG's response to automatically match the grid strength and wind power access status. Attached Figure Description

[0049] Figure 1 This is a block diagram of the SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning, as described in this invention. Detailed Implementation

[0050] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious modifications will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0051] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0052] refer to Figure 1 The technical solution provided by this invention is: a dynamic response system for transient voltage support of power grid based on AI reinforcement learning, comprising: a virtual synchronous machine control module, a signal acquisition module, a reinforcement learning online decision-making module, a power grid state detection module, and a mode switching and management module.

[0053] The virtual synchronous generator (SVG) control module is used to control the static var generator (SGR) to exhibit voltage source characteristics (grid-type SVG). Specifically, it includes an active-frequency control model and a reactive-voltage control model. The active-frequency control model is constructed based on the rotor equations of a simulated synchronous generator and includes virtual inertia parameters. and virtual damping coefficient parameters The equation is:

[0054] ;in, Output angular frequency for SVG. The rated angular frequency of the power grid. This is a reference value for active power. δ represents the electromagnetic power output of the SVG, and δ is the virtual power angle. From the primary frequency modulation characteristics Given, where As a meritorious work point, This is the frequency droop coefficient.

[0055] The reactive-voltage control model is used to generate the internal potential amplitude. The formula is:

[0056] ; To control the cycle length;

[0057] in, The set value for the no-load internal potential. This is the reactive power droop factor. This is a reactive power command. Output reactive power to SVG This is the reference value for the grid connection point voltage. This is the measured value of the voltage amplitude at the grid connection point. and These are the proportional coefficient and integral coefficient for voltage deviation adjustment, respectively.

[0058] In some preferred embodiments, the virtual synchronous generator control module also includes a virtual impedance model. The virtual impedance model is used to simulate the stator impedance of the synchronous generator, calculated by subtracting the voltage drop across the virtual impedance from the given voltage-current loop, as shown in the formula: ,in For virtual resistance, This is a virtual inductance. This virtual impedance not only improves the current sharing performance of multiple machines in parallel, but also provides passive fault current limiting, that is, when an external short circuit occurs, the SVG output voltage automatically decreases, limiting the output current amplitude.

[0059] The signal acquisition module is used to acquire raw data from the target power grid and the static var generator (SVG) in each control cycle to construct the system state vector. The raw data includes the output angular frequency of the SVG. Output electromagnetic power Output current amplitude Grid connection point voltage amplitude Active power of wind power at grid connection point .

[0060] Calculate derived characteristics, including: output angular frequency and grid rated frequency. frequency deviation Rate of change of frequency deviation Active power reference value Power deviation from the output electromagnetic power of the static var generator ; represents the amplitude of active power fluctuation in a wind farm. , .

[0061] After normalizing the above derived features and original data, the system state vector is obtained. .

[0062] The reinforcement learning online decision-making module includes a pre-trained policy network. This policy network receives the system state vector and outputs virtual inertia adjustment and virtual damping coefficient adjustment to adjust the parameters of the virtual synchronous machine control module online. Specifically:

[0063] The pre-trained policy network is a pre-trained Soft Actor-Critic (SAC) network. Its input layer consists of 7 neurons, corresponding to each element of the system state vector; its hidden layer consists of 2-3 fully connected layers with ReLU activation function; its output layer consists of 4 neurons, corresponding to the action mean and the action mean standard deviation, respectively. The action mean is the average of the virtual inertia adjustment and the virtual damping coefficient adjustment.

[0064] The policy network performs deterministic forward inference in each control cycle. The specific computation steps are as follows: the input layer receives a 7-dimensional system state vector; the first hidden layer calculates... ,in It is a 128×7 dimensional weight matrix. The bias vector is 128-dimensional; the second hidden layer is calculated. , ,in It is a 128×128 dimensional weight matrix. A 128-dimensional bias vector; output layer calculation ,in It is a 4×128 weight matrix. This is a 4-dimensional bias vector. Take... The first two components are used as virtual inertia adjustment quantities. and virtual damping coefficient adjustment amount Ignore the last two components. The weight matrix and bias vector are pre-stored in the SVG memory as float32 arrays.

[0065] In obtaining , Then, the output is scaled through the motion post-processing and safety limiting module to obtain the scaled virtual inertia adjustment amount. and virtual damping coefficient adjustment amount .in, ;Will , The data is sent to the virtual synchronization machine control module. and They are respectively Permissible range and The permissible range.

[0066] if Exceeding When, then , Set the value to zero, and then restore it after 3 to 5 control cycles, obtaining the corresponding adjustment value through policy network inference. This is the maximum permissible angle deviation threshold.

[0067] The power grid status detection module is used to identify the power grid strength by estimating the short-circuit ratio of the target power grid connection point in real time, and to detect the wind power access status and wind power output fluctuation intensity of the connection point in real time. This is achieved through the following steps:

[0068] First, the wind power connection status and wind power output fluctuation intensity at the grid connection point are monitored in real time, specifically:

[0069] Calculate the active power of the wind farm within the sliding window mean Standard deviation and peak power change Determine the current wind power connection status: And obtain the wind power output fluctuation intensity .

[0070] This indicates a working condition without wind power access, and the determination criteria are as follows: And it lasts for 0.5 seconds; The average power threshold is set to a value of [value to be filled in]. It is 0.05 times the historical average.

[0071] The condition indicating stable wind power output is determined by the following criteria: And it lasts for 1 second; The standard deviation threshold is 0.1 times the historical average standard deviation.

[0072] The conditions indicating fluctuations in wind power grid connection are determined by the following criteria: And it lasts for 0.2 seconds;

[0073] The condition indicating wind power recovery from the first low voltage ride-through is determined by the following criteria: And it lasts for 0.5 seconds.

[0074] For different wind power operating conditions, a condition-related pre-bias is superimposed on the original action output by the strategy network (before output scaling) to account for the impact of wind power access.

[0075] Under operating conditions, the motion bias is zero. Virtual inertia adjustment bias in motion bias under operating conditions .

[0076] The virtual inertia adjustment offset and the virtual inertia adjustment offset under W2 operating condition are: , The upper limit of the allowable range is increased by 1.3 times, and the adjusted virtual inertia adjustment amount is... Adjusted virtual damping coefficient adjustment amount .

[0077] Under working conditions , The upper limit is increased by 1.5 times, and the adjusted virtual damping coefficient is adjusted accordingly. . and Based on the virtual inertia and integrated virtual damping coefficient.

[0078] When the current wind power grid connection status is determined to have changed based on the judgment conditions, direct access from... Operating conditions switched to or You must switch to first During fault recovery, it is also not allowed to start from [the fault location]. Switch directly to For example, if a sudden, severe fluctuation or fault recovery signal occurs when there is no wind power connection, it must be processed through... Wind power connection has been confirmed. Furthermore, during the switching of operating conditions, a mandatory transition time limit is added, specifically... → The transition time is 0.5s. and The transition time between them is 0.2s. and The transition time between these states is 0.2 seconds. This ensures a smooth transition during wind power grid connection and a continuous fault recovery process.

[0079] Secondly, the grid strength is identified by estimating the short-circuit ratio at the target grid connection point in real time. Specifically:

[0080] The modulus of the Thevenin equivalent impedance of the power grid is estimated online using the Thevenin equivalent parameter estimation method with recursive least squares to calculate the short-circuit ratio (SCR). The method includes the following steps:

[0081] Step 1: Retrieve the current time and the previous recognition period (20ms ago). Calculate the SVG output current based on the measured values. Components in the grid voltage-oriented dq coordinate system ;calculate Components in the grid voltage-oriented dq coordinate system .in, Reflects amplitude, It tends to zero in steady state. Response amplitude, It tends to zero in steady state. The frequency for performing grid strength identification is 50Hz.

[0082] Calculate the change .

[0083] Step 2: Perform excitation detection, including: calculating the active power fluctuation amplitude of the wind farm. and reactive power fluctuation amplitude ,like And within two consecutive recognition periods and The direction of change remains consistent. , (The symbols must be consistent), and If the excitation is positive, it is considered a valid stimulus, and step 3 is executed; otherwise, the Kalman gain matrix is... Set to zero, parameter estimation error covariance matrix Keep.

[0084] in, The power change threshold, Current amplitude change threshold; , , This is the maximum output current of the SVG. This represents the maximum active power of the power grid.

[0085] Step 3, recursive least squares update, including: constructing Regression matrix at time points and measurement vector Calculate according to the recursive formula ,renew and Forgetting factor Specifically:

[0086] ; ; . Estimated vector of equivalent impedance of the power grid at time t ,in, for The estimated resistance value at time t. for The estimated reactance at time t.

[0087] Step 4: Calculate and obtain an estimated value for the short-circuit ratio. ,in, The rated capacity of the SVG (in MVA).

[0088] Step 5: Take the average value of the SCR estimates for 5 consecutive identification cycles (i.e., the SCR estimates within a 100ms time window).

[0089] Determine if the average of the estimated values ​​is within the range of [0.5, 100]. If it is, retain the average of the current estimated values; otherwise, determine that it is out of range, discard the average of the current estimated values, and use the average of the estimated values ​​from 5 identification periods ago.

[0090] If the current SCR estimate is less than 5, the power grid strength is considered weak.

[0091] If the current SCR estimate is greater than or equal to 5 and less than 10, the power grid strength is judged to be moderate.

[0092] If the current SCR estimate is greater than or equal to 10, the power grid strength is judged to be strong.

[0093] Step 6: Confidence assessment, including:

[0094] Based on the number of times the incentive condition has been met in the last second and The confidence level of the trace of a matrix includes:

[0095] If the proportion of times the incentive condition is met in the most recent second is greater than 80% of the total number of times, and If the trace is less than the preset threshold, the power grid strength identification confidence level is determined to be high confidence.

[0096] If the percentage of times the incentive condition is met within the last second is greater than or equal to 50% and less than or equal to 80% of the total number of times, or If the trace is greater than the preset threshold, the power grid strength identification confidence level is determined to be medium confidence.

[0097] If the proportion of times the excitation condition is met in the most recent second is less than 50% of the total number of times, then the confidence level of the power grid strength identification is judged to be low confidence level.

[0098] During SVG failure, i.e. Pause the recursive update and retain the pre-fault SCR estimate. The fault is cleared and the voltage recovers to [previous value]. Recognition will resume after a 0.2-second delay. This is the rated output voltage for the SVG.

[0099] The mode switching and management module is used to determine the current operating mode from a variety of preset operating modes based on the short-circuit ratio and wind power access status, and to send the basic virtual inertia and basic virtual damping coefficient corresponding to the current operating mode to the virtual synchronous machine control module, and to send the allowable range of adjustment corresponding to the current operating mode to the reinforcement learning online decision-making module. Specifically:

[0100] The preset multiple operating modes include: mode 1, mode 2, mode 3, mode 4, mode 5, and mode 6;

[0101] The first mode is used when the power grid is weak (i.e., the power grid strength is weak) and there is no wind power access;

[0102] The second mode is used when there is a strong power grid (i.e., a strong power grid) and no wind power is connected;

[0103] The third mode is used when the power grid is weak and wind power is connected.

[0104] The fourth mode is used when there is a strong power grid and wind power is connected.

[0105] The fifth mode is defined as the presence of wind power and the occurrence of a special wind power event.

[0106] The duration of a low-confidence level in power grid strength identification exceeding the maximum permissible time is classified as the sixth mode.

[0107] Upon detecting a wind power special event, the system switches from the current mode among the first to fourth modes to the fifth mode. In the fifth mode, the upper limit of the allowable range of the adjustment amount of the reinforcement learning online decision-making module is increased, for example, adjusted to 1.3 times the original upper limit, to allow a large virtual damping coefficient adjustment to suppress power oscillations caused by the wind power special event. Wind power special events include wind turbine tripping events and low voltage ride-through recovery events.

[0108] When the duration of the power grid strength identification confidence level being low exceeds the maximum allowable time, the current mode among the first to fourth modes is switched to the sixth mode;

[0109] In the sixth mode, the current operating mode is frozen, mode switching is refused, and the allowable range of adjustment of the parameters of the reinforcement learning online decision-making module is narrowed, that is, the upper limit of the allowable range of adjustment is adjusted to 0.8 times the original value, and the lower limit of the allowable range is adjusted to 0.8 times the original value.

[0110] Each time a new operating mode is switched, a parameter set corresponding to the current operating mode is sent to the virtual synchronization machine control module, specifically:

[0111] A parameter configuration table is preset within the mode switching and management module, storing the parameter set corresponding to each operating mode. The parameter set includes: basic virtual inertia, basic virtual damping coefficient, etc. Permissible range , Permissible range Frequency droop coefficient, reactive power droop coefficient;

[0112] The parameter set corresponding to each operating mode is sent to the virtual synchronous machine control module.

[0113] Furthermore, the switching between modes one and four requires a dynamic confirmation time. This is because, firstly, after a real change in the SCR (Stable Response Rate), the grid state detection module needs a certain amount of time to converge to the new steady-state value, primarily influenced by the convergence speed of the Recursive Least Squares (RLS) method. The confirmation time must be greater than the RLS convergence time. Secondly, changes in the wind power access status, i.e., the determination of whether or not wind power has been connected, also require a certain amount of time. This has already been mentioned in the determination.

[0114] Dynamic confirmation times are used instead of fixed confirmation times because in some scenarios, fixed confirmation times would make it impossible to determine the switching time. For example, in multi-event coupled scenarios, wind turbine tripping and grid strength changes occur simultaneously. With fixed confirmation times, the confirmation time for one event may be met, but the confirmation time for another event may not be met, making it impossible to determine the switching timing.

[0115] The calculation process for the dynamic confirmation time includes the following steps:

[0116] The basic confirmation time is calculated based on the rate of change of the short-circuit ratio. The basic confirmation time decreases monotonically as the rate of change increases, and has preset minimum and maximum confirmation times, specifically:

[0117] Short-circuit ratio change rate in, This is the window for estimating the rate of change. Second.

[0118] Basic confirmation time ,in, Indicates the earliest confirmation time. Seconds, corresponding to step change scenarios, are longer than the time it takes for a single device action, ensuring that the switch occurs only after the physical change has truly completed. Latest confirmation time. Seconds; reference rate of change This is the dividing line used to distinguish between "fast" and "slow" changes. This value is approximately the rate at which the SCR changes from 5 to 6 (crossing a weak network hysteresis loop) in 3 seconds. Steepness coefficient. To make the confirmation time in The nearby rapid transitions enhance the distinction between fast and slow changes.

[0119] An acknowledgment time correction coefficient is generated based on the power grid strength identification confidence level. This acknowledgment time correction coefficient increases as the confidence level decreases. Specifically:

[0120] When the confidence level of power grid strength identification is low, the confirmation time correction factor is used. ;

[0121] When the confidence level of power grid strength identification is medium confidence, the confirmation time correction factor is... ;

[0122] When the power grid strength identification confidence level is high, the confirmation time correction factor is used. .

[0123] A wind power fluctuation correction coefficient is generated based on the wind power output fluctuation intensity. This correction coefficient is greater than 1 when the fluctuation intensity is below a first threshold and less than 1 when the fluctuation intensity is above a second threshold. Specifically:

[0124] when At that time, wind power fluctuation correction coefficient ;

[0125] when At that time, wind power fluctuation correction coefficient ;

[0126] when At that time, wind power fluctuation correction coefficient .

[0127] Multiply the base confirmation time by the confirmation time correction factor and the wind power fluctuation correction factor to obtain the dynamic confirmation time. .

[0128] Example 2: Fifth Mode Special Mode. In this mode, the determination method for the confirmation time is different from the dynamic confirmation time. Adaptive adjustments are required. Because the exit confirmation time in the fifth mode addresses the issue of when to release enhanced protection, it prevents the enhanced protection from being prematurely revoked before the system has truly recovered. The specific solution is as follows:

[0129] In the fifth mode, entry does not require any confirmation time. The confirmation time for exiting the fifth mode is adaptively adjusted based on the intensity of active power fluctuations in the wind farm. Specifically:

[0130] Set the recovery speed coefficient, when At that time, the recovery rate coefficient ;

[0131] when At that time, the recovery rate coefficient ;

[0132] when At that time, the recovery rate coefficient .

[0133] Then, adaptive confirmation time .

[0134] The present invention also provides a computer-readable storage medium storing a computer program that is executed by a processor to implement the aforementioned AI reinforcement learning-based grid transient voltage support SVG dynamic response system.

[0135] The block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that a system according to various embodiments of the invention may implement. In this regard, each block in the block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0136] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the principles described, the implementation of the present invention may have any changes or modifications.

Claims

1. A dynamic response system for grid transient voltage support SVG based on AI reinforcement learning, characterized in that, include: The virtual synchronous machine control module is used to control the static var generator to exhibit voltage source characteristics. The signal acquisition module is used to acquire raw data from the target power grid and static var generator in each control cycle and construct the system state vector. The reinforcement learning online decision-making module includes a pre-trained policy network, which receives the system state vector and outputs virtual inertia adjustment and virtual damping coefficient adjustment to adjust the parameters of the virtual synchronous machine control module online. The power grid status detection module is used to identify the power grid strength by estimating the short-circuit ratio of the target power grid connection point in real time, and to detect the wind power access status and wind power output fluctuation intensity of the connection point in real time. The mode switching and management module is used to determine the current operating mode from a variety of preset operating modes based on the short-circuit ratio and wind power access status, and to send the basic virtual inertia and basic virtual damping coefficient corresponding to the current operating mode to the virtual synchronous machine control module, and to send the allowable range of adjustment corresponding to the current operating mode to the reinforcement learning online decision-making module.

2. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 1, characterized in that, The raw data includes the output angular frequency, output electromagnetic power, output current amplitude, grid connection point voltage amplitude, and grid connection point wind power active power of the static var generator (SVG). Based on the original data, derived characteristics are calculated, including the frequency deviation between the output angular frequency and the grid rated frequency, the rate of change of the frequency deviation, the power deviation between the active power reference value and the output electromagnetic power of the static var generator, and the active power fluctuation amplitude of the wind farm. After normalizing the output current amplitude, grid connection point voltage amplitude, grid connection point wind power active power and derived characteristics, the system state vector is obtained. The pre-trained policy network is a pre-trained soft actor critic SAC algorithm network, which includes an input layer, a hidden layer, and an output layer. The input layer comprises 7 neurons, corresponding to each element of the system state vector; the hidden layer comprises 2-3 fully connected layers with ReLU activation function; the output layer comprises 4 neurons, corresponding to the action mean and the action mean standard deviation, respectively, wherein the action mean is the mean of the virtual inertia adjustment and the virtual damping coefficient adjustment. The virtual synchronous machine control module includes an active-frequency control model and a reactive-voltage control model; The active-frequency control model is constructed based on the rotor equations of a simulated synchronous generator, including virtual inertia parameters. and virtual damping coefficient parameters The equation is: ;in, Output angular frequency for SVG. The rated angular frequency of the power grid. This is a reference value for active power. δ represents the electromagnetic power output of the SVG, and δ is the virtual power angle. From the primary frequency modulation characteristics Given, where As a meritorious work point, This is the frequency droop coefficient. The reactive-voltage control model is used to generate the internal potential amplitude. The formula is: ; To control the cycle length; in, The set value for the no-load internal potential. This is the reactive power droop factor. This is a reactive power command. Output reactive power for SVG This is the reference value for the grid connection point voltage. This is the measured value of the voltage amplitude at the grid connection point. and These are the proportional coefficient and integral coefficient for voltage deviation adjustment, respectively.

3. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 2, characterized in that, The mode switching and management module is also used to: determine multiple operating modes based on the short-circuit ratio and wind power access status, the operating modes including: a first mode, a second mode, a third mode, a fourth mode, a fifth mode and a sixth mode; The first mode is characterized by a weak power grid and no wind power access; the second mode is characterized by a strong power grid and no wind power access; the third mode is characterized by a weak power grid and wind power access; and the fourth mode is characterized by a strong power grid and wind power access. The fifth mode is the occurrence of a special wind power event; the sixth mode is the duration of a low-level confidence level in grid strength identification exceeding the maximum permissible time.

4. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 3, characterized in that, The mode switching and management module is also used to: switch from the current mode among the first to fourth modes to the fifth mode when a special wind power event is detected; When the duration of the power grid strength identification confidence level being low exceeds the maximum allowable time, the current mode from the first to the fourth modes is switched to the sixth mode; In the fifth mode, the upper limit of the allowable range of adjustment of the reinforcement learning online decision-making module is increased; In the sixth mode, the current operating mode is frozen, mode switching is refused, and the allowable range of adjustment of the parameters of the reinforcement learning online decision-making module is narrowed.

5. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 4, characterized in that, The mode switching and management module includes an adaptive confirmation time generator; The adaptive confirmation time generator is used to dynamically calculate the confirmation time based on the rate of change of the short-circuit ratio, the confidence level of the grid strength identification, and the fluctuation range of wind power processing. When the current operating mode is inconsistent with the target operating mode and the duration reaches the confirmation time of the dynamic calculation, the mode switching and management module triggers mode switching.

6. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 5, characterized in that, The dynamic calculation confirmation time includes: The basic confirmation time is calculated based on the rate of change of the short-circuit ratio. The basic confirmation time decreases monotonically as the rate of change increases, and has a preset minimum confirmation time and a preset maximum confirmation time. An acknowledgment time correction coefficient is generated based on the identification confidence level of the power grid strength, and the acknowledgment time correction coefficient increases as the confidence level decreases; A wind power fluctuation correction coefficient is generated based on the wind power output fluctuation intensity. The wind power fluctuation correction coefficient is greater than 1 when the fluctuation intensity is lower than a first threshold and less than 1 when the fluctuation intensity is higher than a second threshold. The dynamic confirmation time is obtained by multiplying the basic confirmation time by the confirmation time correction factor and the wind power fluctuation correction factor.

7. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 6, characterized in that, The mode switching and management module is also used to: in the fifth mode, adaptively adjust the confirmation time for exiting the fifth mode according to the intensity of active power fluctuation in the wind farm; Within the medium power grid intensity range, discrete mode switching is not triggered. Instead, the basic virtual inertia and the basic virtual damping coefficient are generated by continuous interpolation between the parameter sets of adjacent modes based on the short-circuit ratio.

8. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 7, characterized in that, It also includes a motion post-processing and safety limiting module; The post-processing and safety limiting module is used to: receive the original adjustment amount output by the policy network, limit the original adjustment amount according to the allowable range of adjustment amount corresponding to the current operating mode issued by the mode switching and management module, and then output it to the virtual synchronizer control module.

9. The SVG dynamic response system for power grid transient voltage support based on AI reinforcement learning according to claim 8, characterized in that, The power grid state detection module uses a recursive least squares method with a forgetting factor to estimate the equivalent impedance of the power grid, and performs recursive updates only when a valid excitation condition is detected. The wind power access status detection module quantifies the output fluctuation intensity by analyzing the mean and peak power change of the active power of the wind farm, and judges the wind power access status based on the mean, standard deviation and peak power change of the active power of the wind farm. The wind power access status is used by the mode switching and management module to adjust the operation mode judgment logic and generate the pre-bias related to the operating condition.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the SVG dynamic response system based on AI reinforcement learning for power grid transient voltage support as described in any one of claims 1-9.

Images