Overload line emergency control strategy continuous learning method suitable for changing scenes

Abstract

The application provides an overload line emergency control strategy continuous learning method suitable for changing scenes. First, a large strategy model and a strategy network model are designed. Then, for the overload operation mode in a continuous time period, a large strategy model and an initial strategy network in the current time period are learned based on the large strategy model in the last time period and the initial strategy network, and the specific process includes searching for the overload operation mode in the time period; a decision-making process is implemented by using the large strategy model to screen out the overload operation mode that cannot be processed by the large strategy model; the overload operation mode that cannot be processed by the large strategy model is learned by using a reinforcement learning method to obtain a strategy network; and the overload operation mode processed by the strategy network and an emergency control scheme are taken as new samples of the large strategy model to further learn the large strategy model. The application can realize the continuous learning of the overload line emergency control strategy in the changing scene and learn the emergency control strategy suitable for each time period.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of emergency control of power system, and particularly relates to an overload line emergency control strategy continuous learning method suitable for changing scenarios. BACKGROUND

[0002] In the field of emergency control of power system, for the operation mode and its potential risk scenario of each time period, an emergency control strategy that can be used for the overload operation mode processing of the time period needs to be generated in advance. With the development of artificial intelligence technology, reinforcement learning has been gradually applied to the generation of emergency control strategy, which can learn the emergency control strategy based on the potential risk operation mode searched in advance under the time period.

[0003] However, the emergency control strategy generated based on the reinforcement learning method for a certain time period has limited scope of application. When the operation mode changes, the emergency control strategy is not applicable, at which time the emergency control strategy suitable for the time period needs to be relearned. Most current researches follow the above principle. However, after learning a new control strategy for the changed operation mode and its potential risk scenario each time, the value of the old strategy is discarded, and the training cost is wasted. Obviously, the existing emergency control strategy learning method cannot retain the value of the old strategy, and the training cost is wasted, and there is an urgent need for an overload line emergency control strategy continuous learning method that can continuously accumulate overload line emergency control experience. SUMMARY

[0004] In order to at least solve one of the problems existing in the prior art, the present application provides an overload line emergency control strategy continuous learning method suitable for changing scenarios.

[0005] In order to achieve the purpose of the present application, the present application provides an overload line emergency control strategy continuous learning method suitable for changing scenarios, which comprises the following steps:

[0006] S1: design a large strategy model Π and a strategy network model π, and perform the emergency control strategy continuous learning process of S2-S5 based on the large strategy model Π and the strategy network model π;

[0007] S2: define a symbol T representing the continuous learning time period, the number of time periods n, and the sample set D learned by the large strategy model Π s ;

[0008] S3: initialize the time period index i of continuous learning i=1, and initialize the initial large strategy model Π0 and the initial strategy network π0;

[0009] S4: perform emergency control strategy learning on the overload operation mode under T i time period based on the large strategy model Π i-1and the initial policy network i-1 learns the large policy model Π i and the initial policy network i , and outputs the latest sample set D' learned by the large policy model s ;

[0010] S5: Determine whether i is greater than n? If yes, end and output the large policy model Π after continuous learning i and the policy network model i ; if not, i = i + 1, D s = D' s , and jump to S4 to continue the policy learning of the next time period.

[0011] The large policy model Π and the policy network model π are both constructed based on a neural network, and are used to represent a high-dimensional emergency control scheme.

[0012] Further, the neural network used to represent the large policy model Π and the policy network model π is a Transformer-based graph neural network, and the large policy model Π and the policy network model π both include a feature extraction module and a fitting module, the feature extraction module includes alternately arranged Transformer convolution layers and batch normalization layers, the Transformer convolution layers are used to extract graph features, and the fitting module includes a full connection network.

[0013] Further, the input of the large policy model Π and the policy network model π is the state s = {X, A} of the operation mode, where X is the grid node feature, V is the node voltage, is the node phase, P is the node injected active power, Q is the node injected reactive power, and A represents the adjacency matrix of the grid topology.

[0014] Further, the output of the large policy model Π and the policy network model π is an emergency control measure vector a of adjusting line power, including adjusting generator power measures a G , load shedding measures a L and generator tripping measures a GC , a G , a L and a GC are all in [0, 1], respectively representing the percentage of the adjustment amount of each measure, assuming that there are N M generators and N D loads, then:

[0015]

[0016] wherein, is the N MThe adjustment amount of the power adjustment measure of the i-th generator node, The cut-off amount of the i-th load node, D The cut-off amount of the i-th generator node. M The cut-off amount of the i-th generator node.

[0017] Further, the steps of the emergency control strategy learning method include:

[0018] S4-1: search overload operation modes under the T i time period, and record the set of overload operation modes as D i ;

[0019] S4-2: determine whether i is equal to 1? If yes, D' i = D i , and jump to S4-4; if not, directly jump to S4-3;

[0020] S4-3: implement the decision-making process on the overload operation modes in the set D i-1 using the large strategy model Π i , obtain the decision-making effect, and screen out scenarios that cannot be handled by the large strategy model Π i-1 , and record the set as D' i ;

[0021] S4-4: learn the overload operation modes in the set D' i using the reinforcement learning method, and learn the strategy network model π i ;

[0022] S4-5: implement the decision-making process on the overload operation modes in the set D' i using the strategy network model π i , and form a mapping pair {s k , a k}, wherein s k is the state of the overload operation mode, a k is the emergency control scheme, and k is the index of the operation mode in the set D' i ; store all the mapping pairs {s k , a k} in the large strategy model sample set D s , and form a sample set D' s ;

[0023] S4-6: for the samples in the sample set D' s , perform deep learning on the large strategy model Π i-1 using the graph deep learning method, and learn the large strategy model Π i . ​

[0024] Further, in step S4-1, the search T i The overload operation mode under the time period is obtained by the method for obtaining the basic operation mode under the time period T i The basic operation mode under the time period is obtained by the method for obtaining the basic operation mode under the time period T

[0025] Further, the specific method for implementing the decision-making process is as follows: first, the state of the overload operation mode is extracted, including the node voltage V, the node phase The node active power P and the node reactive power Q, and the adjacent matrix of the current power grid are extracted to form the state s = {X, A} of the operation mode; then, the state s is input into the large strategy model Π i-1 to obtain the emergency control measure vector a of the adjusted line power; finally, the operation mode of the power grid is modified according to the emergency control measure vector a of the adjusted line power, including the generator node power and the load power, and the operation mode is recalculated by power flow calculation and time domain simulation to obtain the decision-making effect, that is, whether the power grid still appears line overload or node voltage out-of-limit after the decision-making.

[0026] Further, the convergence condition of the reinforcement learning method is set as the strategy network model π i can process all overload operation modes in the set D' i , so that all line overloads in the operation mode are controlled to be within the limit value.

[0027] Further, the graph deep learning method is to take the large strategy model Π i-1 of the previous time period T i-1 as the basic model, take the mean square error function as the loss function, and constantly learn the new large strategy model Π i by multiple iterations, and when the number of iterations is reached, the model with the largest success rate in each iteration in the past is selected as the large strategy model Π i .

[0028] Compared with the prior art, the present application can at least achieve the following beneficial effects:

[0029] The application provides an overload line emergency control strategy continuous learning method suitable for changing scenes, and a new continuous learning process is designed to solve the problems that current emergency control strategy learning methods are difficult to retain the value of old strategies and waste training costs. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 FIG. 1 is a step flowchart of the overload line emergency control strategy continuous learning method suitable for changing scenes provided by the embodiment of the application.

[0031] Figure 2 FIG. 2 is a structural schematic diagram of the big strategy model Π or the strategy network model π in the embodiment of the application (both have the same structure).

[0032] Figure 3 FIG. 3 is a flowchart of the emergency control strategy learning in the embodiment of the application.

[0033] Figure 4 FIG. 4 is an IEEE-39 node system topology diagram in the embodiment of the application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, those skilled in the art can obtain without creative labor.

[0035] Please refer to Figure 1 The overload line emergency control strategy continuous learning method suitable for changing scenes provided by the embodiment of the application comprises the following steps.

[0036] S1: Construct the big strategy model Π and the strategy network model π, and then perform the emergency control strategy continuous learning process of S2-S5 based on the big strategy model Π and the strategy network model π.

[0037] To ensure the generalization performance of the emergency control strategy and represent the high-dimensional space, the neural network is used to represent the large strategy model Π and the strategy network π. Since the functions of the large strategy model Π and the strategy network model π are similar, both of them are used to calculate the emergency control scheme according to the operation state of the power grid, and thus the two models are defined as the same structure.

[0038] For the large strategy model Π and the strategy network model π, the input of both models is defined as the state of the operation mode s = {X, A}, wherein X is the node feature of the power grid, and the node voltage V, the node phase , the node active power injection P and the node reactive power injection Q are considered, that is, wherein the voltage is the per-unit value, and the phase is expressed in radians; A represents the adjacent matrix of the power grid topology, and the adjacent matrix A will change with the on-off state of the line or the main transformer. Only when the line and the main transformer are in the connected state, the element in the corresponding position of A is 1, otherwise, the element is 0.

[0039] For the large strategy model Π and the strategy network model π, the output of both models is the emergency control measure vector of the adjusted line power, which is denoted by a, and the expression is shown in equation (1). The emergency control measure vector a of the adjusted line power includes the generator power adjustment measure a G , the load shedding measure a L and the generator tripping measure a GC , wherein a G , a L and a GC are all in [0, 1], and respectively represent the percentage of the adjustment amount of each measure. Assuming that there are N M generator nodes and N D load nodes, then is the adjustment amount of the power adjustment measure of the N M th generator node, is the amount of the load shedding measure of the N D th load node, is the amount of the generator tripping measure of the N M th generator node.

[0040]

[0041] In one of the embodiments of the present application, for the network structure of the large policy model Π and the policy network model π, considering that the power grid operation state is strongly related to the topology, the network structure of both is selected as a graph neural network based on Transformer which has strong ability to extract features of the power grid topology, the graph neural network based on Transformer includes a feature extraction module and a fitting module, the activation function uses a Relu function, the feature extraction module includes a Transformer convolution (TC) layer, a batch normalization (BN) layer, the fitting module includes a fully connected network (FCN), wherein the TC layer is used to extract graph features, the BN layer is used to adjust the distribution of data to make it more suitable for training, and the FCN layer is used to fit the data, the fully connected network has two, as shown in Figure 2 It should be noted that the number of each layer needs to be determined according to the business situation, and the current mainstream method for determining the number of each layer in the power field is the trial method, that is, trying different configurations and selecting the number of layers with better effect. In one of the embodiments of the present application, the TC layer is set to three layers, the BN layer is set to four layers, and the BN layer and the TC layer are alternately arranged, and in other embodiments, other numbers can be set. The input of the network model is the state s = {X, A}, after the feature extraction module and the fitting module, the emergency control measure vector a of adjusting the line power is output. Wherein, the graph data after feature extraction is expanded into a one-dimensional vector as the input of FCN1 and FCN2, and the emergency control measure vector a of adjusting the line power is fitted. Wherein, FCN1 fits continuous variables a G and a L , FCN2 fits discrete variables a GC , and finally a is synthesized, wherein a G is a vector composed of the generator power adjustment measures a G , a L is a vector composed of the load shedding measures a L , and a GC is a vector composed of the machine tripping measures a GC .

[0042] S2: define a symbol T representing a continuous learning time period, the number of time periods is n, and the sample set learned by the large policy model Π is D s .

[0043] S3: initialize the time period index i = 1 of continuous learning, and initialize the initial large policy model Π0 and the initial policy network model π0.

[0044] S4: T iEmergency control strategies are learned based on overload operation modes over a specific time period, using a large strategy model Π. i-1 And policy network model π i-1 Learn the big strategy model Π i And policy network model π i It outputs the latest sample set D' for large policy model learning. s .

[0045] S5: Determine if i is greater than n. If yes, end the process and output the large policy model Π after continuous learning. i With policy network model π i If not, then let i = i + 1, D s =D' s Then, it jumps to S4 to continue learning the strategy for the next time period.

[0046] The specific process for learning the emergency control strategy in step S4 above can be found in [link to relevant documentation]. Figure 3 The process can be summarized as follows:

[0047] S4-1: Search T i The set of overload operation modes over a given time period is denoted as D. i .

[0048] S4-2: Determine if i equals 1? If yes, then D' i =D i If not, then jump to S4-4; otherwise, jump directly to S4-3.

[0049] S4-3: Employing the Big Strategy Model Π i-1 For the set D of overload operation modes i The overload operation mode in the process is implemented to determine the decision-making effect. The large strategy model Π is selected. i-1 Scenarios that cannot be handled are denoted as set D'. i .

[0050] S4-4: Using reinforcement learning methods on set D' i The overload operation mode is learned to obtain the policy network model π. i .

[0051] S4-5: Employing the policy network model π i For set D' i The decision-making process for overload operation mode is implemented to form a mapping pair {s} k ,a k}, where s k In the overload operation mode, a k For emergency control plans, k is a set D' i The index of the running mode in the middle. All mapping pairs {sk a k} into the sample set D s , and form the sample set D' s .

[0052] S4-6: For the samples in the sample set D' s , the graph deep learning method is used to learn the grand strategy model Π i-1 , and the learned grand strategy model is denoted as Π i .

[0053] Details of the above steps are described below.

[0054] The function of step S4-1 is to find the overload operation mode under T i . First, a large number of operation modes are generated under the basic operation mode of T i , considering the uncertainty factors such as line opening, generator node opening and source-load fluctuation; then, the power flow calculation and time domain simulation are performed on the generated large number of operation modes, the full network line current under each operation mode is calculated, and the current I j of each line is compared with the current limit I j,limit of the line to determine whether it is overloaded, wherein if formula (2) is satisfied, it is considered to be overloaded, and j is the jth line; finally, all the overload scenes are screened out to form the set D i of overload operation modes.

[0055] I j >I j,limit (2)

[0056] The function of step S4-3 is to filter out the operation modes that can be handled in the set D i of overload operation modes through the grand strategy model, and to perform reinforcement learning again for other unhandleable scenes, so as to reduce the number of operation modes to be learned and improve the learning efficiency.

[0057] In one embodiment of the present application, the specific decision-making method of the grand strategy model Π i-1 for the overload operation mode in the set D i of overload operation modes is as follows: first, the state of the overload operation mode is extracted, including the node voltage V, the node phase , the node active power P and the node reactive power Q, and the adjacency matrix of the current power grid is extracted to form the state s={X,A} of the operation mode. Then, the grand strategy model Π i-1Inputting the state s, obtaining the emergency control measure vector a of regulating line power. Finally, modifying the operation mode of the power grid according to the emergency control measure vector a of regulating line power, including the generator node power and the load power, and re-performing power flow calculation and time domain simulation on the operation mode to obtain the decision effect, i.e., whether the line overload or node voltage out-of-limit condition still exists in the power grid after the decision.

[0058] In the above steps, the modification of the operation mode of the power grid according to the emergency control measure vector a of regulating line power is different according to the control object. The emergency control measure vector a of regulating line power contains three measures, the generator power regulation measure corresponds to a G , the power regulation expression of the mth generator node under this measure is shown in formula (3), i.e., the maximum power adjustable amount P m,max of the generator node m is adjusted according to the product of a G,m ; the load shedding measure corresponds to a L , the power regulation expression of the dth load node under this measure is shown in formula (4) and (5); the generator tripping measure corresponds to a GC , the power regulation expression of the mth generator node under this measure is shown in formula (6), i.e., when a GC is greater than or equal to 0.5, P G,m is set to 0, otherwise it remains unchanged.

[0059] P G,m =a G,m ·P m,max (3)

[0060] P′ L,d =a L,d ·(P L,d -P L,d,min ) (4)

[0061] Q′ L,d =a L,d ·(Q L,d -Q L,d,min ) (5)

[0062]

[0063] Wherein, P G,m represents the power of the mth generator node; a G,m represents the adjustment amount of the power regulation measure of the mth generator node; a L,d represents the amount of load shedding of the dth load node; P L,d represents the active power of the dth load node; P L,d,min represents the minimum active power of the dth load node; P′ L,dActive power after cutting off the load of the dth load node; Q L,d Reactive power of the dth load node; Q L,d,min Minimum reactive power of the dth load node, Q' L,d Reactive power after cutting off the load of the dth load node.

[0064] In one embodiment of the present application, in S4-4, the reinforcement learning method used is a common reinforcement learning method, which learns a policy network model π i for the overload operating modes in set D' i through the reinforcement learning method. It should be noted that the convergence condition of the reinforcement learning method is set as the policy network model π i can handle all overload operating modes in set D' i , so that all line overloads in the operating modes are controlled within the limit.

[0065] In one embodiment of the present application, in S4-5, the specific decision-making manner of using the policy network model π i to make decisions for the overload operating modes in set D' i is as follows: first, the state of the overload operating mode is extracted, including the node voltage V, the node phase , the node active power injection P and the node reactive power injection Q, and the adjacency matrix of the current power grid is extracted to form the state s = {X, A} of the operating mode. Then, the state s is input to the large policy model Π i-1 to obtain an emergency control measure vector a of the adjusted line power. Finally, the operating mode of the power grid is modified according to the emergency control measure vector a of the adjusted line power, including the generator node power and the load power, and the operating mode is recalculated for power flow calculation and time domain simulation to obtain the decision-making effect, which is whether the power grid still has line overload or node voltage out-of-limit condition after the decision. After the decision is made, the corresponding control scheme a k for each operating mode state s k is obtained, i.e., the mapping pair {s k , a k}, and all mapping pairs are stored in the large policy model sample set D s to form D' s .

[0066] In one embodiment of the present application, in S4-6, the graph deep learning method used is a common deep learning method. The deep learning method takes the large policy model Π i-1 of the previous time period T i-1 as the base model, takes the mean square error function as the loss function, and continuously learns a new large policy model Π iIt is noted that when the number of iterations is reached, the model with the largest success rate among the models of each iteration in the past is selected as the large strategy model Π i .

[0067] In one embodiment of the present application, the above continuous learning method is verified in an IEEE-39 node system.

[0068] As shown in Figure 4 is a 39-node system topology diagram. Given four time periods T1, T2, T3, and T4, search for overload operating modes under T i periods T1, T2, T3, and T4, and the number of overload operating modes in the four sets is shown in Table 1. It can be seen that the number of overload scenarios in set D1 is 20, of which 10 are line overload operating modes after line opening, 6 are line overload operating modes after generator node opening, and 4 are line overload operating modes after source load fluctuation; the number of overload scenarios in D2 is 21, of which 10 are line overload operating modes after line opening, 6 are line overload operating modes after generator node opening, and 5 are line overload operating modes after source load fluctuation; the number of overload scenarios in D3 is 22, of which 9 are line overload operating modes after line opening, 6 are line overload operating modes after generator node opening, and 7 are line overload operating modes after source load fluctuation; the number of overload scenarios in D4 is 18, of which 9 are line overload operating modes after line opening, 6 are line overload operating modes after generator node opening, and 3 are line overload operating modes after source load fluctuation.

[0069] Table 1 Overload scenario related information in the continuous learning case of the 39-node system

[0070]

[0071] The continuous learning method of the overload line emergency control strategy is used to continuously learn sets D1, D2, D3, and D4 in turn, and the large strategy models learned in T1, T2, T3, and T4 are defined as Π1, Π2, Π3, and Π4, respectively, and the strategy network models are π1, π2, π3, and π4.

[0072] The large strategy models Π i and the strategy network models π i are tested in T i periods for learning effect of D i , where i is 1, 2, 3, and 4. The test steps are as follows:

[0073] 1. First, through step S4-3, use T i-1 Large policy model Π trained before the time period i-1 To process T i The set of overload operation modes D for a given time period i Identify the execution method that caused the failure and denote it as D'. i And set the indicator "Use Π" i-1 Process D i The success rate of the medium-scale operation mode is used to evaluate the large strategy model for T. i Time period D i The ability to handle overload operation.

[0074] 2. Then, for set D' i In the middle scene, reinforcement learning in step S4-4 is performed to learn the tractable set D'. i Strategy network model π for medium overload operation mode i Testing the large strategy model Π i-1 And policy network model π i Can set D be processed jointly? i All overload operation modes are configured, and the indicator "Use strategy network model π" is set. i Process D i "Number of successful runs in the medium mode" and "Using the large strategy model Π" i-1 And policy network model π i Process D i The effectiveness of coprocessing is evaluated by the number of successful runs.

[0075] 3. Next, add the processing set D' to the large strategy model sample. i New sample D of medium overload operation mode s , forming D' s And through steps S4-6, the large strategy model Π i-1 Deep learning was used to obtain the large policy model Π i Testing the large strategy model Π i In set D i The performance of the model was evaluated, and the metric "Using the large strategy model" was set. i Able to process D i The success rate of the operation mode is used to evaluate the effect of the large policy model after learning and to verify the effectiveness of the large policy model learning.

[0076] 4. Finally, use the Grand Strategy Model Π4 to process set D. i The system describes the operating mode, counts the number of successful processes using each mode, and sets an indicator: "Can the large strategy model Π4 process D?" i The number of successful runs is used to verify the effectiveness of the large strategy model's learning.

[0077] The index results are shown in Table 2. It can be seen that:

[0078] 1) For the 20 overload operating modes in set D1, the learned strategy network model π1 can handle all operating modes in set D1, and the large strategy model Π1 can also handle 19 scenarios.

[0079] 2) For the 21 overload operating modes in set D2, the large strategy model Π1 can handle 19 overload operating modes, so set D'2 contains 2 overload operating modes. Reinforcement learning is performed on the 2 overload operating modes in set D2, and the learned strategy network model π2 can handle two scenarios in O 2,L All overload operating modes in set D2 can be jointly handled by the large strategy model Π1 and the strategy network model π2. The decision samples of the 2 overload operating modes in D'2 are added to the training samples of the large strategy model Π, and the large strategy model Π1 is further learned to obtain Π2, which can handle all overload operating modes in set D2.

[0080] 3) For the 22 overload scenarios in set D3, the large strategy model Π2 can handle 20 overload operating modes, so set D'3 contains 2 overload operating modes that cannot be handled. After reinforcement learning is performed on the 2 overload operating modes in D'3, the large strategy model Π2 and the learned strategy network model π3 can jointly handle all overload operating modes in set D3. The decision samples of the 2 scenarios in D'3 are added to the training samples of the large strategy model Π, and the large strategy model Π2 is further learned to obtain the large strategy model Π3, which can handle all scenarios in set D3.

[0081] 4) For the 18 overload scenarios in set D4, the large strategy model Π3 can handle 17 overload operating modes, so set D'4 contains 1 overload operating mode that cannot be handled. After reinforcement learning is performed on the 1 scenario in D'4, the large strategy model Π3 and the learned strategy network model π4 can jointly handle all overload operating modes in set D4. The decision sample of the 1 scenario in D'4 is added to the training samples of the large strategy model Π, and the large strategy model Π3 is further learned to obtain the large strategy model Π4, which can handle all scenarios in set D4.

[0082] 5) The final learned large strategy model Π4 is used to make decision control for the 81 overload operating modes in sets D1, D2, D3, and D4, and all scenarios can be successfully controlled.

[0083] Table 2: Effect of continuous learning of a 39-node system

[0084]

[0085] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An overload line emergency control strategy persistent learning method suitable for changing scenes, characterized in that, The method comprises the following steps: S1: design grand strategy model Π with strategy network model π , based on grand strategy model Π with strategy network model π emergency control strategy continuous learning process of S2~S5; S2: define a symbol representing a continuous learning time period as T , the number of time periods is n , the large strategy model Π , the sample set for learning is D s ; S3: initialize the time period index of the continual learning i = 1, and initialize the initial large policy model Π 0 and the initial policy network π 0; S4: to T i Emergency control strategy learning is performed on the overload operation mode under the time period, and a large strategy model is learned based on the large strategy model Π i-1 And a strategy network model π i-1 Learning a large strategy model Π i And an initial strategy network model π i And output the latest sample set learned by the large strategy model D' s ​ S5: determine i whether it is greater than n , if so, end, output the large strategy model after continuous learning Π i and the policy network π i ; if not, i = i +1, D s = D' s , and jump to S4 to continue policy learning in the next time period; Wherein, the step of learning the emergency control strategy comprises: S4-1: search T i The overload operation mode under the time period is denoted as D i ; S4-2: judge whether or not i is equal to 1, if yes, then D' i = 0, and jump to S4-4; if not, then directly jump to S4-3; D i , and jump to S4-4; if not, then directly jump to S4-3; S4-3: adopt large strategy model Π i-1 overload operation mode implementation decision flow in the set D i The decision effect is obtained, and the large strategy model is screened out Π i-1 The scene that can't be handled is recorded as a set D' i ; S4-4: learning the overload operation mode in the set by using a reinforcement learning method, learning a policy network model D' i π i ;​ S4-5: Adopting a strategy network model π i to the set D' i , the decision flow of the overload operation mode is implemented, and a mapping pair { s k , a k} is formed, wherein s k is the state of the overload operation mode, a k is the emergency control scheme, k is the set D' i , and the operation mode index; all mapping pairs { s k , a k} are stored in a large strategy model sample set D s , and a sample set D' s is formed; S4-6: For the sample set D' s Deep learning method is used to large strategy model Π i-1 Deep learning is performed on the large strategy model Π i .

2. The overload line emergency control strategy persistent learning method suitable for varying scenarios according to claim 1, characterized in that, The grand strategy model Π With strategy network model π Both are based on neural networks, for representing high-dimensional emergency control scheme. 3.The overload line emergency control strategy persistent learning method suitable for varying scenarios according to claim 2, characterized in that, For representing large policy models Π With policy network models π The structure of the neural network is a Transformer-based graph neural network, the large policy model Π With policy network models π Both include a feature extraction module and a fitting module, the feature extraction module includes alternately arranged Transformer convolutional layers and batch normalization layers, the Transformer convolutional layers are used for extracting graph features, and the fitting module includes a fully connected network.

4. The overload line emergency control strategy persistent learning method suitable for varying scenarios according to claim 2, characterized in that, Big strategy model Π The input to the strategy network model π is the state of the operating mode s ={ X , A} where X is the grid node characteristics, X ={ V , φ , P , Q}, V is the node voltage, φ is the node phase, P is the node injected active power, Q is the node injected reactive power, A denotes the adjacency matrix of the grid topology.

5. The method of claim 2-4, wherein, Large-scale strategy model Π The output of the strategy network model π is an emergency control measure vector for regulating line power a , including generator power regulation measures , load shedding measures and generator tripping measures , , and are all between [0, 1], representing the percentage of the regulation amount of each measure respectively, assuming that there are N M generator nodes and N D load nodes, and there are: In the formula, For the first N M The adjustment amount of power regulation measures for each generator node. For the first N D The amount of load shedding measures to be implemented at each load node. For the first N M The amount of generator node disconnection measures to be implemented.

6. The overload line emergency control policy persistent learning method suitable for varying scenarios according to claim 1, wherein, In step S4-1, the search T i The overload operation mode under the time period is generated in a manner that the base operation mode under the time period is generated in a manner that a large number of operation scenarios are generated after considering various uncertain factors. T i The base operation mode under the time period, a large number of operation scenarios are generated after considering various uncertain factors.

7. The overload line emergency control strategy persistent learning method suitable for varying scenarios according to claim 1, characterized in that, The specific implementation decision-making process is as follows: First, extract the overload operation mode status, including node voltage. V Node phase φ Node injection active P and node injection reactive power Q And extract the adjacency matrix of the current power grid to form the state of the operating mode. s ={ X , A Then, towards the large strategy model. Π i-1 Input status s The emergency control measure vector for adjusting line power is obtained. a Finally, based on the emergency control measures vector for adjusting line power... a Modify the power grid's operating mode, including generator node power and load power, and re-perform power flow calculation and time-domain simulation of this operating mode to obtain the decision effect. This decision effect is whether the power grid still experiences line overload or node voltage exceeding the limit after the decision. 8.The overload line emergency control policy persistent learning method suitable for varying scenarios of claim 1, wherein, The convergence condition of the reinforcement learning method is set as a policy network model π i The ability to handle all overload operating modes in the collection D' i All lines in the operating mode are controlled to within the limit value.

9. The method of claim 1-8, wherein, The graph deep learning method is to learn a new large strategy model in each time period T i-1 The large strategy model Π i-1 As a basic model, the mean square error function is used as the loss function, and the new large strategy model is learned through multiple iterations Π i When the number of iterations is reached, the model with the largest success rate in each iteration is selected as the large strategy model Π i .

Images