Power grid voltage stability margin optimization control method based on data driving

By adopting a data-driven grid voltage stability margin optimization control method combined with the TD3 algorithm, the time-consuming problem of traditional methods is solved, and real-time monitoring and optimization control of grid voltage stability are realized, thereby improving the adaptability and reliability of the power system.

CN119029909APending Publication Date: 2024-11-26YIBIN POWER SUPPLY COMPANY STATE GRID SICHUAN ELECTRIC POWER +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411134149.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Traditional continuous power flow calculation methods are time-consuming in assessing voltage stability in power systems, making it difficult to meet the requirements of online calculations and unable to monitor the grid voltage stability margin in real time, leading to grid voltage instability and unreliable power supply.

Method used

A data-driven power grid voltage stability margin optimization control method is adopted. By defining reactive power reserves and constructing influence models of various control actions, and combining the dual-delay deep deterministic strategy gradient algorithm (TD3), a voltage stability margin optimization model is established to achieve real-time monitoring and optimization control.

Benefits of technology

It enables real-time monitoring and optimized control of grid voltage stability, improves the adaptability and real-time performance of the power system, reduces human intervention, enhances system reliability and operational efficiency, and adapts to operation under various extreme conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119029909A_ABST
    Figure CN119029909A_ABST
Patent Text Reader

Abstract

The invention discloses a power grid voltage stability margin optimization control method based on data driving, which comprehensively considers the influence of various control actions on reactive power compensation variation, linearizes the reactive power compensation reserve and the voltage stability margin, establishes an optimization model for calculating the voltage stability margin, and improves the power grid voltage stability margin optimization control accuracy. And the model is converted into a deep reinforcement learning model to obtain an instant response value of the large power grid voltage stability margin under various working conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and more specifically, relates to a data-driven power grid voltage stability margin optimization control method. Background Art

[0002] The voltage stability margin control of large power grids is one of the key factors for the safe operation of power systems. With the deepening of the green energy transformation, the power system presents the "double high" characteristics of a high proportion of renewable energy and a high proportion of power electronic equipment, which has caused profound changes in the dynamic characteristics of the power system and gradually highlighted new stability issues.

[0003] However, for voltage stability margin control, although the traditional continuous power flow calculation method can evaluate the voltage stability of the system, it is time-consuming and difficult to meet the requirements of online calculation. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a data-driven grid voltage stability margin optimization control method to achieve real-time monitoring of the grid voltage stability margin, provide a reliable basis for the adjustment of reactive compensation equipment and the regulation of grid operation mode, and ensure the stability of grid voltage and the reliable and high-quality supply of electric energy.

[0005] To achieve the above-mentioned object of the invention, the present invention provides a data-driven power grid voltage stability margin optimization control method, which is characterized by comprising the following steps:

[0006] (1) Define reactive power reserve, denoted as Q R,t , where t represents the time, t = 1, 2, 3, ...;

[0007] Q R,t =Q max -Q g,t

[0008] Among them, Q R,t is the reactive power reserve in the generator at time t, Q max is the maximum reactive power limit of the generator, Q g,t is the reactive power generated by the unit at time t;

[0009] (2) Construct the effects of four different actions on reactive power reserve;

[0010] Action 1: The impact of active power generation on reactive power reserve is;

[0011]

[0012] Among them, P gi,t is the active power generated by generator set i at time t;

[0013] Action 2: The impact of parallel compensation on reactive power reserve is;

[0014]

[0015] Among them, B shkt is the parallel capacitor bank on bus k at time t;

[0016] Action 3: The impact of active load shedding on reactive power reserve is:

[0017]

[0018] Among them, P Lkt is the active load reduction power on bus k at time t;

[0019] Action 4: The impact of reactive load shedding on reactive power reserve is:

[0020]

[0021] Among them, Q Lkt is the active load reduction power on bus k at time t;

[0022] (3) Define that there is a linear relationship between reactive power reserve and voltage stability margin;

[0023] VSM t =αQ R,t +β

[0024]

[0025] Among them, VSM t represents the voltage stability margin at time t, α represents the change ratio between the voltage stability margin and the sum of reactive power reserve, and β is a constant;

[0026] (4) Optimize reactive power reserve and voltage stability margin using four control actions, and construct the objective function and constraint conditions of the voltage stability margin control model for large power grids;

[0027] The objective function is:

[0028]

[0029] in, is the weight for active power generation action, is the adjustment amount of the active power generation of generator set i at time t, is the weight for the parallel capacitor action, is the amount of parallel capacitors connected to bus k at time t, is the weight for active load shedding action, is the active power load adjustment on bus k at time t, is the weight for reactive load shedding action, is the active load reduction power adjustment on bus k at time t, is the number of generating units, Indicates the number of parallel capacitor banks, N k Indicates the number of buses;

[0030] The constraints are:

[0031]

[0032] in, Represents the minimum and maximum active power input of generator set i at time t; Indicates the minimum and maximum parallel capacitor input on bus k at time t; Indicates the minimum and maximum active power load adjustment on bus k at time t; Indicates the minimum and maximum reactive power load adjustment on bus k at time t; V i,t represents the terminal voltage of generator set i at time t, V i min 、V i max Indicates the minimum and maximum terminal voltage of generator set i, V i 0 represents the initial terminal voltage of generator set i;

[0033] (5) Build and train a large power grid voltage margin control model based on the double-delay deep deterministic policy gradient algorithm TD3;

[0034] (5.1) The voltage stability margin control objective function and constraints of the large power grid within an operation cycle are transformed into a Markov decision process containing the state space S, the action space A and the reward function R;

[0035] Among them, S contains all the states of the decision-making process, and the state s at time t t = {Q g,t ,V i,t ,VSM t}; A contains all actions in the decision-making process, the action at time t The specific actions are as follows:

[0036] Action 1

[0037] Action 2

[0038] Action 3

[0039] Action 4

[0040] Among them, action 1 is to generate active power, action 2 is to connect the reactive compensator equipment in parallel, action 3 is to reduce the active load, and action 4 is to reduce the reactive load;

[0041] The reward function R is the cumulative reward obtained by trying to combine the four actions:

[0042]

[0043] (5.2) Construct the four groups of six neural networks with the same structure required for the TD3 algorithm:

[0044] Construct two action networks, denoted as μ and μ', and their parameter sets are denoted as θ μ ,θ μ ', used to implement the input state s t To input action a t ;

[0045] Construct four evaluation networks, denoted as Q1, Q2, Q1', Q2', and their parameter sets are Used to implement input state t , output action a t To the action value function Q π (s t ,a t ), π is the mapping strategy;

[0046] (5.3) Set the total number of iterations N of the data-driven large-scale power grid voltage stability margin optimization control model and the number of iterations T of the Markov process; set the memory bank, record its capacity as M, and initialize it to empty; initialize the parameter set of all neural networks, initialize n = 1, initialize the learning rate α, and initialize the counter m

[0047] (5.4) Reset the large power grid system, set t = 1 and obtain the current state s t , and then execute a Markov process;

[0048] (5.5) Determine whether t is less than T. If t<T, proceed to step (5.6). Otherwise, proceed to step (5.14).

[0049] (5.6), s t Input to the action network μ, and get the output action a t , then in action a t Add noise, get motion Among them, ε is Gaussian noise;

[0050] (5.7) Execute actions get The next moment state after the action s t+1 ;

[0051] (5.8) Construct tuple information And store it in the m%M position of the memory bank, and then assign m=m+1;

[0052] (5.9) Determine whether m is greater than M. If so, proceed to step (5.10); otherwise, proceed to step (5.13);

[0053] (5.10), based on the double-delayed deep deterministic policy gradient update action network μ parameter set θ μ ;

[0054]

[0055] Among them, b represents the number of tuple information sampled with equal probability from the memory bank with replacement, s i ,a i Indicates the corresponding state and action in the i-th tuple information; Indicates that in state s i and action a i Next, we need to find a for the action value function i The gradient of Indicates that in state s i The gradient of the lower network μ;

[0056] (5.11), update the parameter set θ of the evaluation network Q1 and Q2 by minimizing the loss function Q1 ,θ Q2 ;

[0057]

[0058] Among them, L(θ Q1 ) is the loss function of network Q1; Indicates that in state s i and action a i Next, evaluate the action value function value output by network Q1; Indicates that in state s i and action a i The action value function value output by the next evaluation network Q2; y i Indicates that the smaller estimate is obtained after comparing networks Q1' and Q2'; γ is the discount factor, represents the gradient of the loss function of network Q1, Represents the gradient of the loss function of network Q2;

[0059] (5.12), Similarly, update the parameter set of the action network μ' and the evaluation network Q1', Q2';

[0060] θ μ' =τθ μ +(1-τ)θ μ'

[0061]

[0062] Among them, τ is the update coefficient;

[0063] (5.13) Update the current state, assign t = t + 1 and s t =s t+1 , then return to step (5.6);

[0064] (5.14), let n = n + 1, and then determine whether n is greater than N. If so, go to step (5.15); otherwise, go to step (5.5);

[0065] (5.15), the iteration stops, and the neural network parameter set is output, thereby obtaining the online scheduling model of the solar-storage system;

[0066] (6) Real-time collection of reactive power generated by the unit and terminal voltage Then determine the real-time And build the real-time status at time t Finally Input to the large power grid voltage stability margin optimization control model to obtain real-time output action Then follow Realize the optimal control of the voltage stability margin of large power grid.

[0067] The object of the invention of the present invention is achieved in this way:

[0068] The present invention is based on a data-driven grid voltage stability margin optimization control method, which comprehensively considers the impact of various control actions on the change in reactive power compensation, and linearizes the reactive power compensation reserve and the voltage stability margin, establishes an optimization model for calculating the voltage stability margin, and converts it into a deep reinforcement learning model to obtain the instant response value of the large grid voltage stability margin under various working conditions.

[0069] At the same time, the data-driven grid voltage stability margin optimization control method of the present invention also has the following advantages:

[0070] Beneficial effects:

[0071] (1) Compared with traditional methods, the model based on deep reinforcement learning of this invention has significant advantages in adaptability and real-time performance, and can quickly respond to dynamic changes in the power system; it has strong generalization ability and can effectively handle high-dimensional complex system problems; the robustness of the algorithm enables it to remain efficient in the face of incomplete data or noise; the optimization decision-making ability further ensures the stability and performance of the power system; in addition, the flexibility and data-driven characteristics of the reinforcement learning algorithm reduce the dependence on manual intervention and complex models, and improve the reliability and operational efficiency of the system. ;

[0072] (2) The present invention takes into account the impact of various actions on the change in reactive power compensation. The model has high stability and can adapt to operations under various extreme conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the data-driven power grid voltage stability margin optimization control method of the present invention;

[0074] Figure 2 It is a flowchart for building and training a large power grid voltage margin control model based on a double-delay deep deterministic policy gradient algorithm. DETAILED DESCRIPTION

[0075] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0076] Example

[0077] Figure 1 It is a flow chart of the data-driven grid voltage stability margin optimization control method of the present invention.

[0078] In this embodiment, if Figure 1 As shown, the present invention provides a data-driven power grid voltage stability margin optimization control method, comprising the following steps:

[0079] S1. Define reactive power reserve, denoted as Q R,t , where t represents the time, t = 1, 2, 3, ...;

[0080] Q R,t =Q max -Q g,t

[0081] Among them, Q R,t is the reactive power reserve in the generator at time t, Q max is the maximum reactive power limit of the generator, which is 80 MVAr, Qg,t is the reactive power generated by the unit at time t;

[0082] S2, construct the impact of four different actions on reactive power reserve;

[0083] In this embodiment, four control variables are taken into consideration: the adjustment amount of active power generation, the amount of parallel capacitor input, the active power load adjustment amount and the active load reduction power adjustment amount, all of which will affect the reactive power reserve amount. Therefore, it is necessary to define the impact of the reactive power reserve amount under different actions.

[0084] Action 1: The impact of active power generation on reactive power reserve is;

[0085]

[0086] Among them, P gi,t is the active power generated by generator set i at time t;

[0087] Action 2: The impact of parallel compensation on reactive power reserve is;

[0088]

[0089] Among them, B shkt is the parallel capacitor bank on bus k at time t;

[0090] Action 3: The impact of active load shedding on reactive power reserve is:

[0091]

[0092] Among them, P Lkt is the active load reduction power on bus k at time t;

[0093] Action 4: The impact of reactive load shedding on reactive power reserve is:

[0094]

[0095] Among them, Q Lkt is the active load reduction power on bus k at time t;

[0096] S3. Define that there is a linear relationship between reactive power reserve and voltage stability margin;

[0097] Assuming reactive power reserve Q R,t and voltage stability margin VSM t There is a certain relationship. When the reactive power compensation changes slightly, it can be approximately considered that this relationship is a linear relationship, that is,

[0098] VSM t =αQR,t +β

[0099]

[0100] Among them, VSM t represents the voltage stability margin at time t, α represents the change ratio between the voltage stability margin and the sum of reactive power reserve, and its value is 5.1, and β is a constant, and its value is 2.0pu;

[0101] S4. Optimize reactive power reserve and voltage stability margin using four control actions, and construct the objective function and constraint conditions of the voltage stability margin control model for large power grids;

[0102] The objective function is:

[0103]

[0104] in, is the weight for active power generation action, which is 0.4. is the adjustment amount of the active power generation of generator set i at time t, is the weight for the parallel capacitor action, which is 0.3. is the amount of parallel capacitors connected to bus k at time t, is the weight for active load shedding action, and its value is 0.2. is the active power load adjustment on bus k at time t, is the weight for reactive load shedding action, and its value is 0.2. is the active load reduction power adjustment on bus k at time t, is the number of generator sets, which is 3. Indicates the number of parallel capacitor groups, the value is 3, N k Indicates the number of buses, the value is 14;

[0105] The basic constraints include: The first is the voltage stability margin constraint, which ensures that the voltage stability margin VSM of each node t The first is to meet a specific safety threshold to avoid voltage collapse; the second is the reactive power reserve constraint, which is to ensure that at each critical bus or node, the reactive power reserve Q R,t Not less than the minimum safety requirement; therefore, the constraints are constructed as follows:

[0106]

[0107] in, Indicates the minimum and maximum active power input of generator set i; Indicates the minimum and maximum parallel capacitor input on bus k; Indicates the minimum and maximum active power load adjustment on bus k; Indicates the minimum and maximum reactive power load adjustment on bus k; V i,t represents the terminal voltage of generator set i at time t, V i min 、V i max Indicates the minimum and maximum terminal voltage of generator set i, V i 0 represents the initial terminal voltage of generator set i;

[0108] S5, such as Figure 2 As shown, a large power grid voltage margin control model based on the Twin Delayed Deep Deterministic policy gradient algorithm (TD3) is built and trained;

[0109] S5.1. Convert the voltage stability margin control objective function and constraints of the large power grid within an operation cycle into a Markov decision process including state space S, action space A and reward function R;

[0110] Among them, S contains all the states of the decision-making process, and the state s at time t t = {Q g,t ,V i,t ,VSM t}; A contains all actions in the decision-making process, the action at time t The specific actions are as follows:

[0111] Action 1

[0112] Action 2

[0113] Action 3

[0114] Action 4

[0115] Among them, action 1 is to generate active power, action 2 is to connect the reactive compensator equipment in parallel, action 3 is to reduce the active load, and action 4 is to reduce the reactive load;

[0116] The reward function R is the cumulative reward obtained by trying to combine the four actions:

[0117]

[0118] S5.2. Construct the four groups of six neural networks with the same structure required for the TD3 algorithm:

[0119] Construct two action networks, denoted as μ and μ', and their parameter sets are denoted as θ μ ,θ μ ', used to implement the input state s t To input action a t ;

[0120] Construct four evaluation networks, denoted as Q1, Q2, Q1', Q2', and their parameter sets are Used to implement input state t , output action a t To the action value function Q π (s t ,a t ), π is the mapping strategy;

[0121] S5.3, set the total number of iterations N = 10000 of the data-driven large power grid voltage stability margin optimization control model and the number of iterations of the Markov process T = 24; set the memory bank, record its capacity as M = 48000, and initialize it to empty; initialize the parameter set of all neural networks, initialize n = 1, initialize the learning rate α = 0.0015, and initialize the counter m = 1;

[0122] S5.4. Reset the large power grid system, set t = 1 and obtain the current state s t , and then execute a Markov process;

[0123] S5.5, determine whether t is less than T, if t<T, proceed to step S5.6, otherwise, proceed to step S5.14;

[0124] S5.6、s t Input to the action network μ, and get the output action a t , then in action a t Add noise, get motion Among them, ε is Gaussian noise;

[0125] S5.7. Execution of actions get The next moment state after the action s t+1 ;

[0126] S5.8. Constructing tuple information And store it in the m%M position of the memory bank, and then assign m=m+1;

[0127] S5.9, determine whether m is greater than M, if so, proceed to step S5.10; otherwise, proceed to step S5.13; S5.10, update the parameter set θ of the action network μ based on the double-delayed deep deterministic policy gradient μ ;

[0128]

[0129] Among them, b represents the number of tuple information sampled with equal probability from the memory bank with replacement, s i ,a i Indicates the corresponding state and action in the i-th tuple information; Indicates that in state s i and action a i Next, we need to find a for the action value function i The gradient of Indicates that in state s i The gradient of the lower network μ;

[0130] S5.11. Update the parameter set of evaluation networks Q1 and Q2 by minimizing the loss function

[0131]

[0132] in, is the loss function of network Q1; Indicates that in state s i and action a i Next, evaluate the action value function value output by network Q1; Indicates that in state s i and action a i The action value function value output by the next evaluation network Q2; y i Indicates that the smaller estimate is obtained after comparing networks Q1' and Q2'; γ is the discount factor, represents the gradient of the loss function of network Q1, Represents the gradient of the loss function of network Q2;

[0133] S5.12, Similarly, update the parameter set of the action network μ' and the evaluation network Q1', Q2';

[0134] θ μ' =τθ μ +(1-τ)θ μ'

[0135]

[0136] Among them, τ is the update coefficient;

[0137] S5.13, update the current state, assign t = t + 1 and s t =s t+1 , then return to step S5.6;

[0138] S5.14, let n=n+1, and then determine whether n is greater than N. If so, proceed to step S5.15; otherwise, proceed to step S5.5;

[0139] S5.15, the iteration stops, and a neural network parameter set is output, thereby obtaining an online scheduling model of the photovoltaic-storage system;

[0140] S6. Optimal control of voltage stability margin of large power grid;

[0141] Real-time acquisition of reactive power generated by the unit and terminal voltage Then determine the real-time And build the real-time status at time t Finally Input to the large power grid voltage stability margin optimization control model to obtain real-time output action Then follow Realize the optimal control of large power grid voltage stability margin.

[0142] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A data-driven power grid voltage stability margin optimization control method, characterized in that: The following steps are involved: (1) Define reactive power reserve, denoted as Q R,t , where t represents the time, t = 1, 2, 3, ...; Q R,t =Q max -Q g,t Among them, Q R,t is the reactive power reserve in the generator at time t, Q max is the maximum reactive power limit of the generator, Q g,t is the reactive power generated by the unit at time t; (2) Construct the effects of four different actions on reactive power reserve; Action 1: The impact of active power generation on reactive power reserve is; in, is the active power generated by generator set i at time t; Action 2: The impact of parallel compensation on reactive power reserve is; in, is the parallel capacitor bank on bus k at time t; Action 3: The impact of active load shedding on reactive power reserve is: in, is the active load reduction power on bus k at time t; Action 4: The impact of reactive load shedding on reactive power reserve is: in, is the active load reduction power on bus k at time t; (3) Define that there is a linear relationship between reactive power reserve and voltage stability margin; VSM t =αQ R,t +b Among them, VSM t represents the voltage stability margin at time t, α represents the change ratio between the voltage stability margin and the sum of reactive power reserve, and β is a constant; (4) Optimize reactive power reserve and voltage stability margin using four control actions, and construct the objective function and constraint conditions of the voltage stability margin control model for large power grids; The objective function is: in, is the weight for active power generation action, is the adjustment amount of the active power generation of generator set i at time t, is the weight for the parallel capacitor action, is the amount of parallel capacitors connected to bus k at time t, is the weight for active load shedding action, is the active power load adjustment on bus k at time t, is the weight for reactive load shedding action, is the active load reduction power adjustment on bus k at time t, is the number of generating units, Indicates the number of parallel capacitor banks, N k Indicates the number of buses; The constraints are: in, Represents the minimum and maximum active power input of generator set i at time t; Indicates the minimum and maximum parallel capacitor input on bus k at time t; Indicates the minimum and maximum active power load adjustment on bus k at time t; Indicates the minimum and maximum reactive power load adjustment on bus k at time t; V i,t represents the terminal voltage of generator set i at time t, Indicates the minimum and maximum terminal voltage of generator set i, represents the initial terminal voltage of generator set i; (4) Build and train a large power grid voltage margin control model based on the double-delay deep deterministic policy gradient algorithm TD3; (4.1) The voltage stability margin control objective function and constraints of the large power grid within an operation cycle are transformed into a Markov decision process containing a state space S, an action space A and a reward function R; Among them, S contains all the states of the decision-making process, and the state s at time t t = {Q g,t ,V i,t ,VSM t }; A contains all actions in the decision-making process, the action at time t The specific actions are as follows: action action action action Among them, action 1 is to generate active power, action 2 is to connect the reactive compensator equipment in parallel, action 3 is to reduce the active load, and action 4 is to reduce the reactive load; The reward function R is the cumulative reward obtained by trying to combine the four actions: (4.2) Construct the four neural networks with the same structure required by the TD3 algorithm: Construct two action networks, denoted as μ and μ', and their parameter sets are denoted as θ μ ,θ μ' , used to implement the input state s t To input action a t ; Construct four evaluation networks, denoted as Q1, Q2, Q1', Q2', and their parameter sets are Used to implement input state t , output action a t To the action value function Q π (s t ,a t ), π is the mapping strategy; (4.3) Set the total number of iterations N of the data-driven large-scale power grid voltage stability margin optimization control model and the number of iterations T of the Markov process; set the memory bank, record its capacity as M, and initialize it to empty; initialize the parameter set of all neural networks, initialize n = 1, initialize the learning rate α, and initialize the counter m (4.4) Reset the large power grid system, set t = 1 and obtain the current state s t , and then execute a Markov process; (4.5) Determine whether t is less than T. If t<T, proceed to step (4.6). Otherwise, proceed to step (4.14). (4.6), s t Input to the action network μ, and get the output action a t , then in action a t Add noise, get motion Among them, ε is Gaussian noise; (4.7) Execute actions get The next moment state after the action s t+1 ; (4.8) Construct tuple information And store it in the m%M position of the memory bank, and then assign m=m+1; (4.9) Determine whether m is greater than M. If so, proceed to step (4.10); otherwise, proceed to step (4.13); (4.10), based on the double-delayed deep deterministic policy gradient update action network μ parameter set θ μ ; Among them, b represents the number of tuple information sampled with equal probability from the memory bank with replacement, s i ,a i Indicates the corresponding state and action in the i-th tuple information; Indicates that in state s i and action a i Next, we need to find a for the action value function i The gradient of Indicates that in state s i The gradient of the lower network μ; (4.11) Update the parameter set of the evaluation network Q1 and Q2 by minimizing the loss function in, is the loss function of network Q1; Indicates that in state s i and action a i Next, evaluate the action value function value output by network Q1; Indicates that in state s i and action a i The action value function value output by the next evaluation network Q2; y i Indicates that the smaller estimate is obtained after comparing networks Q1' and Q2'; γ is the discount factor, represents the gradient of the loss function of network Q1, Represents the gradient of the loss function of network Q2; (4.12), Similarly, update the parameter set of the action network μ' and the evaluation network Q1', Q2'; i μ′ =tθ μ +(1-τ)θ μ′ Among them, τ is the update coefficient; (4.13) Update the current state, assign t = t + 1 and s t =s t+1 , then return to step (4.6); (4.14), let n = n + 1, and then determine whether n is greater than N. If so, go to step (4.15); otherwise, go to step (4.5); (4.15), the iteration stops, and the neural network parameter set is output, thereby obtaining the online scheduling model of the solar-storage system; (5) Real-time collection of reactive power generated by the unit and terminal voltage Then determine the real-time And build the real-time status at time t Finally Input to the large power grid voltage stability margin optimization control model to obtain real-time output action Then follow Realize the optimal control of large power grid voltage stability margin.