Dynamic reactive power control method and device for new energy plant station

Abstract

The embodiment of the invention relates to a dynamic reactive power control method and device for a new energy plant station, and the method comprises the steps: calculating a dynamic safety domain of virtual impedance based on a real-time estimation value of power grid impedance; based on a multi-time scale reinforcement learning algorithm, generating a virtual impedance control quantity under the constraint of the dynamic security domain; obtaining a reactive power output error value, and correcting the virtual impedance control quantity by using the reactive power output error value; and performing reactive power output control on the inverter of the plant station based on the corrected virtual impedance control quantity. According to the technical scheme provided by the embodiment of the invention, the power grid impedance is predicted through the multi-time scale reinforcement learning algorithm, the virtual impedance control quantity is generated accordingly, the upper layer prediction model can predict the impedance change trend in advance and provide a prospective reference for lower layer dynamic correction, and the lower layer dynamic correction model can realize instantaneous disturbance suppression. And rapid fine tuning of the virtual impedance is realized, so that more accurate virtual impedance control quantity can be obtained.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of power system control, and in particular, to a dynamic reactive power control method and device for a new energy power station. Background Art

[0002] With the continuous increase in the penetration rate of new energy in the power system, the reactive power control of new energy power stations is becoming increasingly crucial for maintaining the safe and stable operation of the power system and providing large-scale power grid security. In traditional reactive power control methods for new energy power stations, control strategies are usually designed based on fixed power grid model parameters and a control model with a single characteristic.

[0003] However, the actual operation conditions of the power grid are complex and changeable. The power grid impedance will change in real time due to factors such as load fluctuations, power grid topology adjustments, and the randomness of new energy power generation. Control strategies designed based on fixed parameters are difficult to adapt to this dynamic change, which easily leads to inaccurate reactive power control and inability to effectively maintain the system voltage stability. At the same time, the operation of new energy power stations involves both dealing with rapid power fluctuations on a short time scale and ensuring the stable operation of the system on a long time scale. A control model with a single characteristic cannot comprehensively coordinate control objectives at different time scales, making it difficult to achieve optimal reactive power control in actual operation.

[0004] Therefore, how to achieve optimal control of the reactive power of new energy power stations and improve control accuracy has become an urgent problem to be solved. Summary of the Invention

[0005] Based on the above situation of the prior art, the purpose of the embodiments of the present invention is to provide a dynamic reactive power control method and device for a new energy power station. By predicting the power grid impedance through a multi-time scale reinforcement learning algorithm, precise control of the reactive power of the new energy power station is achieved, providing a strong guarantee for the safe operation of the large-scale power grid.

[0006] To achieve the above object, according to one aspect of the present invention, a dynamic reactive power control method for a new energy power station is provided, including the steps of:

[0007] Calculating the dynamic safety domain of the virtual impedance based on the real-time estimated value of the power grid impedance;

[0008] Generating a virtual impedance control quantity under the constraint of the dynamic safety domain based on a multi-time scale reinforcement learning algorithm; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model;

[0009] Obtaining the reactive power output error value, and correcting the virtual impedance control quantity by using the reactive power output error value;

[0010] Reactive power output control of the inverter at the substation is performed based on the modified virtual impedance control quantity.

[0011] Furthermore, the dynamic security region includes impedance constraints; the impedance constraints are expressed as:

[0012] X min ≤X virtual ≤X max

[0013] Among them, X virtual represents the virtual impedance, X min represents the minimum value of the virtual impedance, X max represents the maximum value of the virtual impedance.

[0014] Furthermore, the minimum value of the virtual impedance is expressed as:

[0015]

[0016] The maximum value of the virtual impedance is expressed as:

[0017]

[0018] Among them, ΔV max represents the upper limit of voltage deviation, ΔV min represents the lower limit of voltage deviation, ΔQ max represents the maximum allowable reactive power change, V pcc represents the voltage at the point of common coupling, and X represents the equivalent reactance of the power grid.

[0019] Furthermore, the upper-layer model includes an LSTM-PINN hybrid prediction model, and the LSTM-PINN hybrid prediction model is used to predict the future power grid impedance. The hybrid prediction model is expressed as:

[0020]

[0021] Among them, Z grid (t-n:t) represents the historical impedance sequence, n represents the time step, P pv represents the real-time photovoltaic output value, ω LSTM and ω PINN respectively represent the weight coefficients of the LSTM model and the PINN model, represents the predicted value of the power grid impedance at the kth second in the future.

[0022] Furthermore, the lower-layer virtual impedance dynamic correction model includes an SAC model. The state space s and action space a of the SAC model are expressed as:

[0023]

[0024] a = ΔX virtual ∈[-0.1, 0.1] p.u.

[0025] where Q current represents the current reactive power output; ΔX virtual represents the virtual impedance correction amount; represents the predicted value of the future k-step grid impedance predicted by the upper-layer model;

[0026] The reward function of the SAC model is expressed as:

[0027]

[0028] where α represents the voltage deviation penalty coefficient, β represents the reactive power deviation penalty coefficient, η represents the action change rate penalty coefficient, λ represents the safety limit violation penalty coefficient, ΔV pcc represents the voltage deviation value, ΔQ represents the reactive power deviation value; Ω represents the dynamic security region, and Violation(Ω) represents the degree of violation of the dynamic security region constraint conditions.

[0029] Furthermore, the upper-layer prediction model and the lower-layer dynamic correction model are nested in the time scale. The upper-layer prediction model updates the virtual impedance reference value every first period, and the lower-layer dynamic correction model updates the virtual impedance correction amount every second period. The outputs of the upper-layer prediction model and the lower-layer dynamic correction model are fused to obtain the virtual impedance control amount;

[0030] where the first period is less than the second period.

[0031] Furthermore, using the reactive power output error value to correct the virtual impedance control amount includes:

[0032] Calculating the virtual impedance adjustment value based on the reactive power output error value;

[0033] Correcting the virtual impedance control amount based on the virtual impedance adjustment value.

[0034] Furthermore, the virtual impedance adjustment value ΔX(k) is expressed as:

[0035]

[0036] where K p represents the proportionality coefficient, K i represents the integral coefficient, and e(k) represents the reactive power output error value.

[0037] According to another aspect of the present invention, a new energy power station dynamic reactive power control device is provided, including:

[0038] A dynamic safety region calculation module, configured to calculate a dynamic safety region of a virtual impedance based on a real-time estimated value of a grid impedance;

[0039] A virtual impedance control quantity calculation module, configured to generate a virtual impedance control quantity based on a multi-time scale reinforcement learning algorithm under the constraint of the dynamic safety region; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model;

[0040] A virtual impedance control quantity correction module, configured to obtain a reactive power output error value and correct the virtual impedance control quantity by using the reactive power output error value;

[0041] A reactive power output control module, configured to perform reactive power output control on an inverter of a substation based on the corrected virtual impedance control quantity.

[0042] In summary, the embodiments of the present invention provide a method and device for dynamic reactive power control of a new energy substation. The method includes the steps of: calculating a dynamic safety region of a virtual impedance based on a real-time estimated value of a grid impedance; generating a virtual impedance control quantity based on a multi-time scale reinforcement learning algorithm under the constraint of the dynamic safety region; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model; obtaining a reactive power output error value and correcting the virtual impedance control quantity by using the reactive power output error value; performing reactive power output control on an inverter of a substation based on the corrected virtual impedance control quantity. The technical solution provided by the embodiments of the present invention predicts the grid impedance through a multi-time scale reinforcement learning algorithm, and generates a virtual impedance control quantity accordingly. The upper-layer prediction model can predict the impedance change trend in advance and provide a forward-looking benchmark for the lower-layer dynamic correction. The lower-layer dynamic correction model can achieve instantaneous disturbance suppression and realize rapid fine-tuning of the virtual impedance, so as to obtain a more accurate virtual impedance control quantity. On this basis, the reactive power output error value is used to correct the virtual impedance control quantity, forming a hierarchical control architecture of "coarse adjustment + fine adjustment", realizing precise control of the reactive power of the new energy substation, and providing a strong guarantee for the safe operation of a large-scale power grid. Description of the Drawings

[0043] Figure 1 is a flowchart of the method for dynamic reactive power control of a new energy substation provided by the embodiments of the present invention. Detailed Embodiments

[0044] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts of the present invention.

[0045] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in one or more embodiments of the present invention should have the ordinary meanings understood by those with ordinary skills in the field to which the present invention pertains. The terms "first", "second" and similar words used in one or more embodiments of the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "comprising" or "including" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0046] The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings. In an embodiment of the present invention, a method for dynamic reactive power control in a new energy power station is provided. In the embodiment of the present invention, the new energy power station is a photovoltaic power station. Figure 1 The flowchart of the method for dynamic reactive power control in a new energy power station according to an embodiment of the present invention is shown in Figure 1 As shown, the control method includes the following steps:

[0047] S202. Calculate the dynamic safety region of the virtual impedance based on the real-time estimated value of the grid impedance.

[0048] S2021. Collect the voltage, current, frequency and phase angle of the point of common coupling (PCC) of the grid, and calculate the real-time estimated value of the current grid impedance. Perform a synchronous rotating coordinate transformation (dq transformation) on the collected voltage and current of the point of common coupling of the grid to obtain the dq-axis coordinate values of the voltage and current. Construct a linear regression model:

[0049]

[0050] where V d 、V q 、I d and I q respectively represent the dq-axis components of the voltage and current. represents the observed value, represents the regression matrix, represents the parameter vector, and R and X respectively represent the real part and the imaginary part of the grid impedance.

[0051] Each time when sampling and updating, substitute the regression matrix and the observed value at the current moment into the recurrence formula:

[0052] θ k+1 =θ k +K k+1 (Y k+1 -Φ k+1T θ k )

[0053] The Kalman gain update formula is as follows:

[0054]

[0055] Among them, P represents the covariance matrix, λ is the forgetting factor, and the typical value is 0.99.

[0056] The initial value of the parameter vector θ0 = [R0, X0] T , and the values of R0 and X0 can be set according to the nominal impedance of the power grid. For example, R0 = 0.01 p.u., X0 = 0.05 p.u.; the initial value of the covariance matrix can be set as a diagonal matrix:

[0057] P0 = σ 2 I

[0058] Among them, I represents the identity matrix, and σ takes a relatively large value such as 10 6 , indicating a relatively high initial estimation uncertainty. According to the predetermined sampling period (e.g., 0.1 ms), the parameters are updated every time a sampling data is received, so that the estimated value Z grid (k) of the equivalent impedance of the power grid at the current moment can be obtained:

[0059] Z grid (k) = R(k) + jX(k)

[0060] S2022. Calculate the dynamic safety region of the virtual impedance based on the estimated value of the equivalent impedance of the power grid. The dynamic safety region of the virtual impedance is the impedance constraint generated according to the voltage constraint, which is used to constrain the value of the virtual impedance control quantity. This impedance constraint is expressed as:

[0061] X min ≤ X virtual ≤ X max

[0062] Among them, X virtual represents the virtual impedance, X min and X max represent the minimum and maximum values of the virtual impedance obtained according to the voltage constraint. The photovoltaic power station is connected to the power grid through the virtual impedance, and its equivalent circuit satisfies the equivalent circuit equation:

[0063] V pcc = V grid + (R + jX)·I pcc + jX virtual ·I pcc

[0064] Among them, V grid represents the voltage on the power grid side, Vpcc Denotes the voltage at the common connection point, I pcc Denotes the current at the common connection point. Ignoring the active component, the PV reactive power approximately satisfies:

[0065] Q out ≈V pcc ·I Q

[0066] Where, I Q Denotes the vertical component of the current (i.e., the reactive component). Substituting the current into the equivalent circuit equation, the voltage deviation can be expressed as:

[0067]

[0068] Where, P out Denotes the PV active power output (usually decoupled from the reactive power control). Therefore, when ignoring the active influence (P out changes slowly), the voltage deviation is mainly related to the reactive power:

[0069] ΔV pcc ≈K V ·Q out

[0070] Where, K V Denotes the sensitivity coefficient, which is expressed by the following formula:

[0071]

[0072] Therefore, according to the voltage constraint, X min and X max are calculated:

[0073]

[0074] Where, ΔV max Denotes the upper limit of the voltage deviation, ΔV min Denotes the lower limit of the voltage deviation, ΔQ max Denotes the maximum allowable reactive power change.

[0075] S204. Based on the multi-time scale reinforcement learning algorithm, under the constraints of the dynamic security domain, generate the virtual impedance control quantity. The multi-time scale reinforcement learning algorithm of the embodiment of the present invention includes an upper-layer prediction model and a lower-layer dynamic correction model. The upper-layer prediction model predicts the change of the power grid impedance in the next k seconds (for example, k = 3) based on the LSTM-PINN hybrid model, outputs the predicted value of the power grid impedance, and accordingly obtains the virtual impedance reference value X virtual_base . This hybrid model can be expressed as:

[0076]

[0077] The PINN loss function includes a physical constraint term:

[0078]

[0079] Among them, represents the predicted value of the grid impedance at the kth second in the future. In the LSTM model, the input historical impedance sequence Z grid (t-n:t), where n represents the time step, to obtain the predicted value of the grid impedance at the kth second in the future. In the PINN model, the input real-time photovoltaic output value P pv and the real-time voltage value V pcc of the grid common connection point are used to obtain the predicted increment of the grid impedance at the kth second in the future, ω LSTM and ω PINN represent the weight coefficients of the LSTM model and the PINN model respectively. In this embodiment, ω LSTM = 0.7, ω PINN = 0.3. In the physical constraint term, the first term is used to constrain the mean square error between the predicted voltage and the measured voltage, and the second term is used to constrain the residual between the voltage differential and the theoretical value of the physical model. The physical model is, for example, a voltage differential equation based on Kirchhoff's law. By jointly training the above hybrid model and minimizing the prediction error and physical residual using the physical constraint term, the predicted value of the grid impedance can be solved, thereby reflecting the trend change of the grid impedance.

[0080] According to the predicted value of the grid impedance, the virtual impedance reference value X virtual_base is obtained according to the following formula:

[0081]

[0082] Among them, Z target represents the target total impedance, which can be obtained based on the voltage-reactive power droop characteristic. By setting the target total impedance Z target , the reactive power-voltage characteristic of the photovoltaic substation is made to meet the droop control requirements:

[0083] Q = K·(V ref - V pcc )

[0084]

[0085] Among them, K represents the droop coefficient. According to the grid dispatching requirements or equipment capacity, the maximum reactive power is determined to be Q max , and the allowable voltage deviation range is set to ΔV max , so that the target total impedance can be expressed as:

[0086]

[0087] The lower-layer dynamic correction model adopts the SAC (Soft Actor-Critic) model to generate the virtual impedance correction amount ΔX. virtual Define the state space s and action space a of the SAC (Soft Actor-Critic) model as follows:

[0088]

[0089] a = ΔX virtual ∈[-0.1, 0.1] p.u.

[0090] Where R and X represent the equivalent resistance and equivalent reactance of the power grid respectively, which are the real and imaginary parts of the equivalent impedance of the power grid. The estimated value of the equivalent impedance of the power grid obtained in step S2021 can be used. Q current represents the current reactive power output, and ΔX virtual represents the virtual impedance correction amount, that is, the output of the lower-layer dynamic correction model; represents the predicted value of the future k-step power grid impedance predicted by the upper-layer model. The reward function r is:

[0091]

[0092] Where α represents the voltage deviation penalty coefficient, β represents the reactive power deviation penalty coefficient, η represents the action change rate penalty coefficient, λ represents the safety limit violation penalty coefficient, Ω represents the dynamic security region, and Violation(Ω) represents the degree of violation of the constraints of the dynamic security region, which can be quantified according to, for example, the percentage of the degree of violation. In this embodiment, the typical values of each penalty coefficient are α = 1.0, β = 0.5, η = 0.2, λ = 10.0. Use the trained SAC (Soft Actor-Critic) model to generate the action a according to the real-time state s, that is, obtain the virtual impedance correction amount.

[0093] The upper-layer prediction model and the lower-layer dynamic correction model are nested in the time scale. The upper-layer prediction model updates the virtual impedance reference value every first period, and the lower-layer dynamic correction model updates the virtual impedance correction amount every second period. The first period is less than the second period. For example, the upper-layer prediction model receives the latest prediction result every 1 second and updates the virtual impedance reference value X virtual_base , and the lower-layer dynamic correction model generates the virtual impedance correction amount ΔX according to the real-time state every 10 seconds virtual . Fuse the outputs of the upper-layer prediction model and the lower-layer dynamic correction model, that is, fuse the virtual impedance reference value and the virtual impedance correction amount, to obtain the virtual impedance control amount:

[0094] X virtual = X virtual_base ·(1 + tanh(ΔX virtual ))

[0095] In the multi-time-scale reinforcement learning algorithm provided by the embodiments of the present invention, the upper-layer prediction model can predict the impedance change trend in advance, providing a forward-looking benchmark for the lower-layer dynamic correction. For example, when it is predicted that the impedance is increasing, the virtual impedance reference value can be adjusted upward in advance to reserve reactive power capacity; when a decreasing trend is detected, the virtual impedance reference value can be reduced to avoid overcompensation. The lower-layer dynamic correction model can achieve instantaneous disturbance suppression (such as cloud occlusion, switch operation, etc.), and can realize rapid fine-tuning of the virtual impedance through the virtual impedance control quantity.

[0096] S206. Obtain the reactive power output error value, and use the reactive power output error value to correct the virtual impedance control quantity. The reactive power output error value e(k) is expressed as the difference between the predicted reactive power value Q pred (k) and the measured reactive power value Q meas (k):

[0097] e(k) = Q pred (k) - Q meas (k)

[0098] In the embodiments of the present invention, the predicted reactive power value is obtained through moving average prediction. To improve the prediction accuracy, other more complex algorithms can also be used. According to the moving average prediction, the reactive power data of the nearest N sampling points (for example, N = 5) are used, and the predicted reactive power value at the next moment is predicted as:

[0099]

[0100] Among them, Q pred (k + 1) represents the predicted reactive power value at the next moment, Q(k - i) represents the reactive power value of the nearest sampling point, and the weight distribution coefficient w i can be set such that the weight of the adjacent moment is higher.

[0101] According to the reactive power output error value, the virtual impedance is corrected by using the incremental PI feedback method. The virtual impedance adjustment value ΔX(k) is expressed as:

[0102]

[0103] Among them, K p represents the proportional coefficient, and the typical value is 0.05 - 0.1. The proportional term is used to quickly respond to the current error and immediately adjust the virtual impedance; K i represents the integral coefficient, and the typical value is 0.01 - 0.05. The integral term is used to eliminate the historical cumulative error and avoid the steady-state deviation. In the embodiments of the present invention, the proportional term directly uses the current error e(k), and the integral term uses the historical cumulative error to eliminate the long-term steady-state deviation. If e(k) > 0, it indicates that the measured reactive power is insufficient, and X needs to be reduced virtual to increase the reactive power output; if e(k) < 0, it indicates that the measured reactive power is excessive, and X needs to be increased virtual to reduce the reactive power output.

[0104] S208. Perform reactive power output control on the inverters of the substation based on the calibrated virtual impedance control quantity. In a multi-inverter scenario, the virtual impedance values of each inverter can be allocated according to the contribution degree weights of each inverter. Input the final virtual impedance control quantity into the inverter control loop to generate a reactive power reference value, and realize dynamic tracking output through a PI controller.

[0105] An embodiment of the present invention also provides a new energy substation dynamic reactive power control device, including:

[0106] A dynamic security region calculation module, configured to calculate the dynamic security region of the virtual impedance based on the real-time estimated value of the grid impedance;

[0107] A virtual impedance control quantity calculation module, configured to generate a virtual impedance control quantity under the constraint of the dynamic security region based on a multi-time scale reinforcement learning algorithm; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model;

[0108] A virtual impedance control quantity correction module, configured to obtain a reactive power output error value and correct the virtual impedance control quantity by using the reactive power output error value;

[0109] A reactive power output control module, configured to perform reactive power output control on the inverters of the substation based on the corrected virtual impedance control quantity.

[0110] The specific steps for each module in the new energy substation dynamic reactive power control device in this embodiment of the present invention to implement its functions are the same as the steps provided in the new energy substation dynamic reactive power control method in the above embodiment of the present invention, and the repeated description thereof will be omitted here.

[0111] In summary, the embodiments of the present invention relate to a method and device for dynamic reactive power control in a new energy power station. The method includes the steps of: calculating the dynamic safety region of the virtual impedance based on the real-time estimated value of the grid impedance; generating a virtual impedance control quantity under the constraint of the dynamic safety region based on a multi-time scale reinforcement learning algorithm, where the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model; obtaining the reactive power output error value and using the reactive power output error value to correct the virtual impedance control quantity; and performing reactive power output control on the inverter of the power station based on the corrected virtual impedance control quantity. The technical solution provided by the embodiments of the present invention predicts the grid impedance through a multi-time scale reinforcement learning algorithm, and accordingly generates a virtual impedance control quantity. The upper-layer prediction model can predict the impedance change trend in advance and provide a forward-looking benchmark for the lower-layer dynamic correction. The lower-layer dynamic correction model can achieve instantaneous disturbance suppression and realize rapid fine-tuning of the virtual impedance, so as to obtain a more accurate virtual impedance control quantity. On this basis, the reactive power output error value is used to correct the virtual impedance control quantity, forming a hierarchical control architecture of "coarse adjustment + fine adjustment", and realizing precise control of the reactive power in the new energy power station.

[0112] It should be understood that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the present invention (including the claims) is limited to these examples; under the concept of the present invention, the technical features in the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations in different aspects of one or more embodiments of the present invention as described above, and they are not provided in detail for the sake of brevity. The above specific embodiments of the present invention are only used for exemplary illustration or explanation of the principle of the present invention and do not constitute a limitation to the present invention. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the present invention shall be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundary of the appended claims, or equivalent forms of such scope and boundary.

Claims

1. A dynamic reactive power control method for a new energy plant and substation, characterized in that, Including the steps: Calculate the dynamic safety region of the virtual impedance based on the real-time estimated value of the grid impedance; Based on the multi-time scale reinforcement learning algorithm, generate the virtual impedance control quantity under the constraint of the dynamic safety region; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model; Obtain the reactive power output error value, and use the reactive power output error value to correct the virtual impedance control quantity; Perform reactive power output control on the inverter of the substation based on the corrected virtual impedance control quantity.

2. The method according to claim 1, wherein The dynamic safety region includes impedance constraints; the impedance constraints are expressed as: X min ≤X virtual ≤X max Among them, X virtual represents the virtual impedance, X min represents the minimum value of the virtual impedance, X max represents the maximum value of the virtual impedance.

3. The method according to claim 2, wherein The minimum value of the virtual impedance is expressed as: The maximum value of the virtual impedance is expressed as: Among them, ΔV max represents the upper limit of voltage deviation, and ΔV min represents the lower limit of voltage deviation. ΔQ max represents the maximum allowable reactive power change, and V pcc represents the voltage at the common connection point, and X represents the equivalent reactance of the power grid.

4. The method according to claim 1, wherein The upper-layer model includes an LSTM-PINN hybrid prediction model, and the LSTM-PINN hybrid prediction model is used to predict the future grid impedance. The hybrid prediction model is expressed as: Among them, Z grid (t-n:t) represents the historical impedance sequence, n represents the time step, P pv represents the real-time photovoltaic output value, ω LSTM and ω PINN represent the weight coefficients of the LSTM model and the PINN model respectively, represents the predicted value of the grid impedance at the k-th second in the future.

5. The method according to claim 4, characterized in that The lower-layer virtual impedance dynamic correction model includes an SAC model. The state space s and action space a of the SAC model are expressed as: a = ΔX virtual ∈ [-0.1, 0.1] p.u. Among them, Q current represents the current reactive power output; ΔX virtual represents the virtual impedance correction amount; represents the predicted value of the future k-step power grid impedance predicted by the upper-layer model; The reward function of the SAC model is expressed as: Among them, α represents the voltage deviation penalty coefficient, β represents the reactive power deviation penalty coefficient, η represents the action change rate penalty coefficient, λ represents the safety limit violation penalty coefficient, and ΔV pcc represents the voltage deviation value, and ΔQ represents the reactive power deviation value; Ω represents the dynamic security region, and Violation(Ω) represents the degree of violation of the constraint conditions of the dynamic security region.

6. The method according to claim 5, characterized in that, The upper-layer prediction model and the lower-layer dynamic correction model are nested in terms of time scale. The upper-layer prediction model updates the virtual impedance reference value every first period, and the lower-layer dynamic correction model updates the virtual impedance correction quantity every second period. The outputs of the upper-layer prediction model and the lower-layer dynamic correction model are fused to obtain the virtual impedance control quantity; Wherein, the first period is less than the second period.

7. The method according to any one of claims 1-6, characterized in that, Using the reactive power output error value to correct the virtual impedance control quantity includes: Calculate the virtual impedance adjustment value based on the reactive power output error value; Correct the virtual impedance control quantity based on the virtual impedance adjustment value.

8. The method according to claim 7, characterized in that, The virtual impedance adjustment value ΔX(k) is expressed as: Among them, K p represents the proportionality coefficient, and K i represents the integral coefficient, and e(k) represents the reactive power output error value.

9. A dynamic reactive power control device for a new energy power station, characterized in that, Including: A dynamic safety region calculation module, which is used to calculate the dynamic safety region of the virtual impedance based on the real-time estimated value of the grid impedance; A virtual impedance control quantity calculation module, which is used to generate the virtual impedance control quantity under the constraint of the dynamic safety region based on the multi-time scale reinforcement learning algorithm; the multi-time scale reinforcement learning algorithm includes an upper-layer prediction model and a lower-layer dynamic correction model; A virtual impedance control quantity correction module, which is used to obtain the reactive power output error value and use the reactive power output error value to correct the virtual impedance control quantity; A reactive power output control module, which is used to perform reactive power output control on the inverter of the substation based on the corrected virtual impedance control quantity.

Images