Operation mode intelligent confrontation adjustment method based on physical embedded neural network

Abstract

The invention discloses an operation mode intelligent confrontation adjustment method based on a physical embedded neural network, and the method comprises the following steps: 1), building an operation mode N / N-1 power flow coupling relation model based on the physical embedded neural network; 2) constructing a to-be-adjusted operation mode input feature X, and inputting the to-be-adjusted operation mode input feature X to the operation mode N / N-1 power flow coupling relation model to obtain an output feature Yo; 3) calculating and outputting an out-of-limit vector delta Y; 4) judging whether an output out-of-limit vector delta Y is 0, if delta Y is equal to 0, skipping to a step 6), otherwise, entering a step 5); 5) calculating according to a confrontation process formula to obtain an input adjustment strategy vector delta X, adjusting the input features, and returning to the step 2); and 6) taking the input feature X corresponding to the current output feature Yo as an operation mode meeting static N-1 security verification. According to the method, the accurate adjustment strategy of the operation mode can be automatically obtained, and the adjustment process of the operation mode before and after adjustment is analyzed, so that the operation mode adjustment of power flow out-of-limit in the N / N-1 state is realized.

Description

Technical Field

[0001] The invention relates to the field of power systems and automation thereof, and in particular to an intelligent confrontation adjustment method for operating modes based on a physical embedded neural network. Background Art

[0002] Adjusting the operation mode is one of the key measures to ensure the safe and stable operation of the power system. With the "double high" trend of a high proportion of renewable energy and a high proportion of power electronic equipment, the scale and complexity of the power grid are increasing, and the actual adjustment of the power grid operation mode is difficult. In this regard, the General Office of the National Energy Administration issued the "Notice on Improving the Power System Operation Mode Analysis System and Strengthening the Safety Risk Management of Power System Operation", highlighting the urgency of adjusting the power system operation mode. In practice, the operation mode of over-limit flow under the N / N-1 state may lead to safety risks such as equipment overload, line damage, and system instability. Therefore, it is of vital importance to accurately formulate an operation mode adjustment strategy to eliminate the over-limit flow under the N / N-1 state (that is, to meet the N-1 safety check).

[0003] Regarding the adjustment of the operation mode, the industry still mainly relies on manual adjustment. The staff repeatedly adjusts the operation mode of the flow exceeding the limit under the N / N-1 state through trial and error based on manual experience. This method has strong interpretability, but there are two problems: 1) It relies on manual experience and it is difficult to form a standardized and unified theoretical guidance rule; 2) A large number of N / N-1 state flow calculations and repeated adjustments of the operation mode are required, which is inefficient and time-consuming. Therefore, for new power systems, it is urgent to further study the operation mode adjustment method with standardized theoretical guidance and efficiency.

[0004] The academic community has conducted in-depth research on the adjustment of the operation mode, which can be mainly divided into direct optimization method and data-driven method. Among them, the direct optimization method can solve the optimization model through the optimization algorithm and directly obtain the operation mode adjustment strategy. Compared with the manual adjustment method, it has obvious high efficiency advantages. For example, some studies use the sensitivity method to find the adjustment node, and then use the interior point method to solve the operation mode optimization model, and take measures such as adjusting the generator node output and shedding the load to directly obtain the operation mode adjustment strategy for the flow exceeding the limit under the N state; some scholars use the optimization algorithm based on swarm intelligence theory (such as improved particle swarm algorithm, genetic simulated annealing particle swarm algorithm, improved bird flock algorithm, etc.) to solve the power grid optimization dispatching model, and obtain the optimization adjustment strategy such as generator node output and load shedding that meets the flow constraints under the N state. In addition to considering only the adjustment of the operation mode in the N state, some researchers have established a power flow optimization model for adjusting the over-limit power flow in the N-1 state, and used the interior point method to solve it, which can directly control the generator output and load to participate in the power flow adjustment after the N-1 fault. Since all N-1 constraints are considered, the above power flow optimization model is large in scale and the solution efficiency is low. Therefore, some researchers have intelligently screened the effective N-1 safety constraints based on deep learning methods, and further constructed a power flow optimization model to quickly solve the optimal power flow. The solution algorithms of the above methods are all iterative solutions. In order to further and more efficiently solve the problem of adjusting the operation mode, researchers have begun to apply data-driven methods to this field. Some scholars have used data-driven methods such as fully connected neural networks (FCNN), convolutional neural networks (CNN), Bayesian neural networks (Bayesian networks), and deep Q networks (DQN) to directly characterize the power flow coupling relationship in the N state, and further derive the adjustment scheme of the operation mode considering the N state. Although the above data-driven methods can effectively extract the nonlinear characteristics of the power flow, their ability to extract topological features is limited. Therefore, some scholars use graph convolutional neural networks (GCNNs) embedded in the power flow physical model for power flow optimization, which can not only effectively extract topological features, but also provide adjustment strategies for the power flow over-limit operation mode under the N state. However, the above data-driven methods lack research on the adjustment of the power flow over-limit operation mode under the N-1 state. More importantly, although the direct optimization method and the data-driven method can obtain the final adjustment strategy of the operation mode, they cannot provide the transition process and internal connection of the operation mode before and after the adjustment, which makes it difficult for staff to explain and analyze the problem of operation mode adjustment, affecting decision-making efficiency. Therefore, the following two issues need to be further studied in order to deeply analyze the transition process and internal connection of the operation mode before and after the adjustment: 1) Effectively characterize the power flow coupling relationship under the N / N-1 state; 2) Accurately analyze the adjustment process between the operation modes before and after the adjustment. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent confrontation adjustment method for operation mode based on a physical embedded neural network, comprising the following steps:

[0006] 1) Based on the physical embedded neural network, a N / N-1 power flow coupling relationship model is constructed to characterize the data-driven power flow coupling relationship under the N / N-1 state;

[0007] 2) Construct the input feature X of the operation mode to be adjusted, and input it into the operation mode N / N-1 power flow coupling relationship model to obtain the output feature Y o ;

[0008] 3) Based on the output feature Y o , calculate the output limit-crossing vector ΔY;

[0009] 4) Determine whether the output over-limit vector ΔY is 0. If ΔY=0, it is determined that the operation mode to be adjusted does not exceed the limit, and jump to step 6). Otherwise, it is determined that the operation mode to be adjusted exceeds the limit, and enter step 5);

[0010] 5) Calculate the input adjustment strategy vector ΔX according to the adversarial process formula, adjust the input feature X=X-ΔX, and return to step 2);

[0011] 6) With the current output feature Y o The corresponding input feature X is used as the operation mode that satisfies the static N-1 safety check.

[0012] Further, the operation mode N / N-1 power flow coupling relationship model includes a power flow equation embedded graph convolution submodule, a power flow equation embedded graph convolution submodule, a branch power equation embedded graph convolution submodule, a node feature prediction convolution submodule and a branch feature prediction convolution submodule;

[0013] The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under N states;

[0014] The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under the N / N-1 state;

[0015] The graph convolution submodule embedded in the branch power equation is used to extract branch power features under the N / N-1 state;

[0016] The node feature prediction convolution submodule is used to predict node voltage features in N states and K N-1 states;

[0017] The branch feature prediction convolution submodule is used to predict branch power features in N states and K N-1 states.

[0018] Furthermore, the input of the graph convolution submodule embedded in the power flow equation is as follows:

[0019] X={U in ,θ in , P in , Q in} (1)

[0020] In the formula, U i is the voltage amplitude of the ith node; U in ,θ in , P in , They represent the voltage amplitude characteristic vector, voltage phase angle characteristic vector, injected active power characteristic vector, and injected reactive power characteristic vector of the power flow node under the N state respectively; n is the number of nodes in the power system;

[0021] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is derived by the power flow equation under N states;

[0022] The power flow equation in N state is as follows:

[0023]

[0024] Where Pi, Q i are the injected active power and injected reactive power of the ith node respectively; G ij , B ij Respectively represent the elements of the i-th row and j-th column in the conductance matrix G and the susceptance matrix B; θ ij represents the voltage phase angle difference from node i to node j; N(i) represents the adjacent set of node i;

[0025] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows:

[0026]

[0027] Where l represents the number of graph convolution layers, l = 1, 2, ..., m, and m is the total number of graph convolution layers in the graph convolution submodule 1 embedded in the power flow equation; are the voltage amplitude characteristics and phase angle characteristics of the N-th state output by the ,th graph convolution layer, F l is the dimension of the output feature of the lth layer; σ is the activation function; is the weight coefficient corresponding to the voltage amplitude characteristic and phase angle characteristic; U Nl+1 ,θ Nl+1 The voltage amplitude and phase angle features under N states output by the l+1th graph convolution layer;

[0028] The output Y1 of the graph convolution submodule embedded in the power flow equation is as follows:

[0029] Y1={UNm ,θ Nm} (4)

[0030] In the formula, is the output of the mth convolutional layer.

[0031] Further, the input of the power flow equation embedded graph convolution submodule is the output of the power flow equation embedded graph convolution submodule;

[0032] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows:

[0033]

[0034] In the formula, θ=[θ1,θ1,…,θ n ] T ; P = [P1, P2, ... P n ] T ; Q = [Q1, Q2, ... Q n ] T ; D(*) is to extract the diagonal elements in * and construct a diagonal matrix; x, x N-1k is the propagation matrix;

[0035] The output Y2 of the graph convolution submodule embedded in the power flow equation is as follows:

[0036] Y2={U N-1，k ,θ N-1，k} (6)

[0037] Where, k = 1, 2, ..., K; K is the number of states considered N-1; They respectively represent the node voltage amplitude characteristic matrix and phase angle characteristic matrix in the N-1 state.

[0038] Further, the input of the branch power equation embedded graph convolution submodule includes the output of the power flow equation embedded graph convolution submodule and the output of the power flow equation embedded graph convolution submodule;

[0039] The forward propagation expression of the graph convolution submodule embedded in the branch power equation is derived by the branch power equation in N states;

[0040] The branch power equation in the N state is as follows:

[0041]

[0042] In the formula, PF ij , QF ij are the active power and reactive power of branch ij respectively;

[0043] The forward propagation expression of the graph convolution submodule embedded in the branch power equation is as follows:

[0044]

[0045] In the formula, PF = [PF ij ] b×1 ; QF=[QF ij ] b×1 ; G br =[G ij ] b×1 ; B br =[B ij ] b×1 ; θ=[θ i ] n×1 ; b is the total number of branches; diag(*) is a diagonal matrix constructed with * as the diagonal element; ω N ,ω N-1，k are the corresponding propagation matrices in the N state and the kth N-1 state respectively; A a is the current;

[0046] The output Y3 of the graph convolution submodule embedded in the branch power equation is as follows:

[0047] Y3={P N , Q N , P N-1，k , Q N-1，k} (9)

[0048] In the formula, They respectively represent the branch active power characteristic matrix and reactive power characteristic matrix in the N state, and the branch active power characteristic matrix and reactive power characteristic matrix in the N-1 state.

[0049] Further, the input of the node feature prediction convolution submodule includes the output of the power flow equation embedded graph convolution submodule and the output of the power flow equation embedded graph convolution submodule;

[0050] The forward propagation formula of the node feature prediction convolution submodule is as follows:

[0051]

[0052] Where, l = 1, 2, ..., c1, c1 is the total number of convolutional layers in the node feature prediction convolutional submodule; U l ,θ l They are respectively the voltage amplitude feature and phase angle feature of the th convolutional layer input; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, When l>1, g l is the number of convolution kernel channels of the ,th layer;

[0053] The output of the node feature prediction convolution submodule is as follows:

[0054] When l = 1, U 1 =[U Nm , U N-1，k ],θ 1 =[θ Nm ,θ N-1，k ], k = 1, 2, ..., K, and the forward propagation formula is used to propagate from the 1st layer to the c1th layer layer by layer, and finally the output feature is obtained: the node voltage amplitude feature vector under the N state Phase angle eigenvector And the node voltage amplitude characteristic vectors under K N-1 states Phase angle eigenvector Right now:

[0055]

[0056] In the formula, They represent the node voltage amplitude characteristic vector, phase angle characteristic vector in the N state, the node voltage amplitude characteristic vector, phase angle characteristic vector in the N-1 state respectively; U o ,θ o is the output feature.

[0057] Further, the input of the branch feature prediction convolution submodule is the output of the branch power equation embedded graph convolution submodule;

[0058] The convolution forward propagation formula of the branch feature prediction convolution submodule is as follows:

[0059]

[0060] Where l = 1, 2, ..., c2, c2 is the total number of convolutional layers in the branch feature prediction convolutional submodule; PF k , QF l They are the branch active power features and branch reactive power features input to the ,th convolutional layer respectively; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, PF l , When l>1, r l is the number of convolution kernel channels of the lth layer;

[0061] The branch feature predicts the output PF of the convolution submodule o , QFo As shown below:

[0062]

[0063] In the formula, They respectively represent the branch active power characteristic vector, branch reactive power characteristic vector in the N state, the branch active power characteristic vector, and branch reactive power characteristic vector in the N-1 state.

[0064] Furthermore, the simplified expression of the N / N-1 power flow coupling relationship model of operation mode is as follows:

[0065] Y o =f(X) (14)

[0066] In the formula, X={U in ,θ in , P in , Q in} is input, Y o = {U o ,θ o , PF o , QF o} is the output. f() is the function.

[0067] Furthermore, the output limit crossing vector ΔU=[ΔU, ΔPF], that is:

[0068]

[0069] Where U o，i PF o，i They are the node voltage amplitude characteristic vector U o And the branch active power characteristic vector PF o The i-th element of up，i , U down,i U o,i Upper and lower limits of PF up，i PF down,i PF o，i The upper and lower limits of ΔU i , ΔPF i is the i-th out-of-limit vector.

[0070] Further, the input adjustment strategy vector ΔX is as follows:

[0071]

[0072] In the formula, ε is the adjustment coefficient; It is the sensitivity of the power flow over-limit ΔY to the input characteristic X under the N / N-1 state.

[0073] It is worth noting that the present invention constructs an N / N-1 flow data-driven coupling relationship model based on the graph convolutional neural network, which can characterize the flow coupling relationship under the N / N-1 state with high precision, and further automatically obtain the precise adjustment strategy of the operation mode based on the confrontation process, and accurately analyze the adjustment process of the operation mode before and after the adjustment, so as to realize the adjustment of the operation mode of the flow exceeding the limit under the N / N-1 state to meet the N-1 safety check.

[0074] The technical effect of the present invention is undoubted, and the beneficial effects of the present invention are as follows:

[0075] The present invention proposes an N / N-1 power flow coupling relationship model based on a physically embedded neural network operation mode, by embedding the power flow equation and the branch power equation into the neural network forward propagation to guide the neural network to more effectively characterize the data-driven power flow coupling relationship under the N / N-1 state.

[0076] The present invention proposes an intelligent confrontation adjustment method for the operation mode based on the N / N-1 power flow coupling relationship. By driving the power flow coupling relationship and the confrontation process with data under the N / N-1 state, an accurate adjustment strategy for the operation mode can be automatically obtained, and the adjustment process of the operation mode before and after the adjustment can be accurately analyzed to achieve adjustment of the operation mode under the N / N-1 state when the power flow exceeds the limit to meet the N-1 safety check.

[0077] The present invention can be widely used in the intelligent adjustment of operation modes in the power system industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is the N / N-1 power flow coupling relationship model architecture of the operation mode;

[0079] Figure 2 Adjust the method flow for intelligent confrontation of operation mode. DETAILED DESCRIPTION

[0080] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0081] Embodiment 1:

[0082] See also Figure 1 to Figure 2 , an intelligent confrontation adjustment method for operation mode based on physical embedded neural network, comprising the following steps:

[0083] 1) Based on the physical embedded neural network, a N / N-1 power flow coupling relationship model is constructed to characterize the data-driven power flow coupling relationship under the N / N-1 state;

[0084] 2) Construct the input feature X of the operation mode to be adjusted, and input it into the operation mode N / N-1 power flow coupling relationship model to obtain the output feature Y o ;

[0085] 3) Based on the output feature Y o , calculate the output limit-crossing vector ΔY;

[0086] 4) Determine whether the output over-limit vector ΔY is 0. If ΔY=0, it is determined that the operation mode to be adjusted does not exceed the limit, and jump to step 6). Otherwise, it is determined that the operation mode to be adjusted exceeds the limit, and enter step 5);

[0087] 5) Calculate the input adjustment strategy vector ΔX according to the adversarial process formula, adjust the input feature X=X-ΔX, and return to step 2);

[0088] 6) With the current output feature Y o The corresponding input feature X is used as the operation mode that satisfies the static N-1 safety check.

[0089] Embodiment 2:

[0090] An operation mode intelligent confrontation adjustment method based on a physical embedded neural network, the technical content is the same as that of Example 1, further, the operation mode N / N-1 power flow coupling relationship model includes a power flow equation embedded graph convolution submodule, a power flow equation embedded graph convolution submodule, a branch power equation embedded graph convolution submodule, a node feature prediction convolution submodule and a branch feature prediction convolution submodule;

[0091] The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under N states;

[0092] The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under the N / N-1 state;

[0093] The graph convolution submodule embedded in the branch power equation is used to extract branch power features under the N / N-1 state;

[0094] The node feature prediction convolution submodule is used to predict node voltage features in N states and K N-1 states;

[0095] The branch feature prediction convolution submodule is used to predict branch power features in N states and K N-1 states.

[0096] Embodiment 3:

[0097] A method for adjusting the operation mode of an intelligent confrontation based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-2, and further, the input of the graph convolution submodule embedded in the power flow equation is as follows:

[0098] X={U in ,θ in , P in , Q in} (1)

[0099] In the formula, U i is the voltage amplitude of the i-th node; They represent the voltage amplitude characteristic vector, voltage phase angle characteristic vector, injected active power characteristic vector, and injected reactive power characteristic vector of the power flow node under the N state respectively; n is the number of nodes in the power system;

[0100] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is derived by the power flow equation under N states;

[0101] The power flow equation in N state is as follows:

[0102]

[0103] Where P i , Q i are the injected active power and injected reactive power of the ith node respectively; G ij , B ij Represent the elements of the i-th row and j-th column in the conductance matrix G and the susceptance matrix B respectively; θ ij represents the voltage phase angle difference from node i to node j; N(i) represents the adjacent set of node i;

[0104] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows:

[0105]

[0106] Where l represents the number of graph convolution layers, l = 1, 2, ..., m, and m is the total number of graph convolution layers in the graph convolution submodule 1 embedded in the power flow equation; are the voltage amplitude characteristics and phase angle characteristics of the N-th state output by the ,th graph convolution layer, F l is the dimension of the output feature of the lth layer; σ is the activation function; is the weight coefficient corresponding to the voltage amplitude characteristic and phase angle characteristic; U Nl+1、 θ Nl+1 The voltage amplitude and phase angle features under N states output by the l+1th graph convolution layer;

[0107] The output Y1 of the graph convolution submodule embedded in the power flow equation is as follows:

[0108] Y1={U Nm ,θ Nm} (4)

[0109] In the formula, is the output of the mth convolutional layer.

[0110] Embodiment 4:

[0111] A method for adjusting the operation mode of an intelligent confrontation based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-3, and further, the input of the graph convolution submodule embedded in the power flow equation is the output of the graph convolution submodule embedded in the power flow equation;

[0112] The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows:

[0113]

[0114] In the formula, θ=[θ1,θ1,…,θ n ] T ; P = [P1, P2, ... P n ] T ; Q = [Q1, Q2, ... Q n ] T ; D(*) is to extract the diagonal elements in * and construct a diagonal matrix; x, x N-1，k is the propagation matrix;

[0115] The output Y2 of the graph convolution submodule embedded in the power flow equation is as follows:

[0116] Y2={U N-1，k ,θ N-1，k} (6)

[0117] Where, k = 1, 2, ..., K; K is the number of states considered N-1; They respectively represent the node voltage amplitude characteristic matrix and phase angle characteristic matrix in the N-1 state.

[0118] Embodiment 5:

[0119] A method for adjusting the operation mode of an intelligent countermeasure based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-4, further, the input of the embedded graph convolution submodule of the branch power equation includes the output of the embedded graph convolution submodule of the power flow equation, and the output of the embedded graph convolution submodule of the power flow equation;

[0120] The forward propagation expression of the graph convolution submodule embedded in the branch power equation is derived by the branch power equation in N states;

[0121] The branch power equation in the N state is as follows:

[0122]

[0123] In the formula, PF ij , QF ij are the active power and reactive power of branch ij respectively;

[0124] The forward propagation expression of the graph convolution submodule embedded in the branch power equation is as follows:

[0125]

[0126] In the formula, PF = [PF ij ] b×1 ; QF=[QF ij ] b×1 ; G br =[G ij ] b×1 ; B br =[B ij ] b×1 ; θ=[θ i ] n×1 ; b is the total number of branches; diag(*) is a diagonal matrix constructed with * as the diagonal element; ω N ,ω N-1，k are the corresponding ω in the N state and the kth N-1 state respectively; A a is the current;

[0127] The output Y3 of the graph convolution submodule embedded in the branch power equation is as follows:

[0128] Y3={P N , Q N , P N-1，k , Q N-1，k} (9)

[0129] In the formula, They respectively represent the branch active power characteristic matrix and reactive power characteristic matrix in the N state, and the branch active power characteristic matrix and reactive power characteristic matrix in the N-1 state.

[0130] Embodiment 6:

[0131] A method for adjusting the operation mode of an intelligent confrontation based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-5, further, the input of the node feature prediction convolution submodule includes the output of the graph convolution submodule embedded in the power flow equation, and the output of the graph convolution submodule embedded in the power flow equation;

[0132] The forward propagation formula of the node feature prediction convolution submodule is as follows:

[0133]

[0134] Where, l = 1, 2, ..., c1, c1 is the total number of convolutional layers in the node feature prediction convolutional submodule; U l ,θ l They are respectively the voltage amplitude feature and phase angle feature of the th convolutional layer input; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, When l>1, g l is the number of convolution kernel channels of the lth layer;

[0135] The output of the node feature prediction convolution submodule is as follows:

[0136] When l = 1, U 1 =[U Nm , U N-1，k ],θ 1 =[θ Nm ,θ N-1，k ], k = 1, 2, ..., K, and the forward propagation formula is used to propagate from the 1st layer to the c1th layer layer by layer, and finally the output feature can be obtained: the node voltage amplitude feature vector under the N state Phase angle eigenvector And the node voltage amplitude characteristic vectors under K N-1 states Phase angle eigenvector Its available U o and θ o It is expressed as follows.

[0137]

[0138] In the formula, They represent the node voltage amplitude characteristic vector, phase angle characteristic vector in the N state, and the node voltage amplitude characteristic vector, phase angle characteristic vector in the N-1 state respectively.

[0139] Embodiment 7:

[0140] A method for adjusting the operation mode of an intelligent confrontation based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-6, and further, the input of the branch feature prediction convolution submodule is the output of the branch power equation embedded graph convolution submodule;

[0141] The convolution forward propagation formula of the branch feature prediction convolution submodule is as follows:

[0142]

[0143] Where l = 1, 2, ..., c2, c2 is the total number of convolutional layers in the branch feature prediction convolutional submodule; PFl , QF l They are the branch active power features and branch reactive power features input to the ,th convolutional layer respectively; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, When l>1, r l is the number of convolution kernel channels of the lth layer;

[0144] The branch feature predicts the output PF of the convolution submodule o , QF o As shown below:

[0145]

[0146] In the formula, They respectively represent the branch active power characteristic vector, branch reactive power characteristic vector in the N state, the branch active power characteristic vector, and branch reactive power characteristic vector in the N-1 state.

[0147] Embodiment 8:

[0148] An intelligent countermeasure adjustment method for an operation mode based on a physically embedded neural network, the technical content of which is the same as any one of Embodiments 1-7, and further, a simplified expression of the operation mode N / N-1 power flow coupling relationship model is as follows:

[0149] Y o =f(X) (14)

[0150] In the formula, X={U in ,θ in , P in , Q in} is input, Y o = {U o ,θ o , PF o , QF o} is the output. f() is the function.

[0151] Embodiment 9:

[0152] A method for adjusting the operation mode of intelligent confrontation based on a physical embedded neural network, the technical content is the same as any one of embodiments 1-8, further, the output of the limit crossing vector ΔY=[ΔU, ΔPF], that is:

[0153]

[0154] Where U o，i PF o，i They are the node voltage amplitude characteristic vector Uo And the branch active power characteristic vector PF o The i-th element of up，i , U down，i U o,i Upper and lower limits of PF up，i PF down，i PF o，i The upper and lower limits of ΔU i , ΔPF i is the i-th out-of-limit vector.

[0155] Embodiment 10:

[0156] A method for adjusting the operation mode of an intelligent confrontation based on a physical embedded neural network, the technical content of which is the same as any one of Embodiments 1-9, and further, the input adjustment strategy vector ΔX is as follows:

[0157]

[0158] In the formula, ε is the adjustment coefficient; It is the sensitivity of the power flow over-limit ΔY to the input characteristic X under the N / N-1 state.

[0159] Embodiment 11:

[0160] A method for intelligent confrontation adjustment of operation mode based on physically embedded neural network, the contents are as follows:

[0161] Firstly, the data-driven power flow coupling relationship under the N / N-1 state is characterized by the N / N-1 power flow coupling relationship model of the operation mode based on the physical embedded neural network; secondly, for the operation mode with power flow exceeding the limit under the N / N-1 state, an intelligent confrontation adjustment method of the operation mode based on the N / N-1 power flow coupling relationship is proposed to obtain the precise adjustment strategy of the operation mode to ensure that it meets the static N-1 safety check; finally, through the case analysis on the 341-node system of an actual large power grid of the IEEE 30-node system, the proposed method can accurately analyze the adjustment process of the operation mode before and after the adjustment, and realize the adjustment of the operation mode with power flow exceeding the limit under the N / N-1 state to meet the N-1 safety check, which verifies the effectiveness of the method proposed in the present invention; the specific contents are as follows.

[0162] N / N-1 power flow coupling relationship model based on physical embedded neural network operation mode

[0163] The operation mode N / N-1 power flow coupling relationship model architecture designed by the present invention is as follows Figure 1 As shown, it includes a flow equation embedded graph convolution submodule 1, a flow equation embedded graph convolution submodule 2, a branch power equation embedded graph convolution submodule, a node feature prediction convolution submodule and a branch feature prediction convolution submodule.

[0164] For the embedded graph convolution submodule 1 of the power flow equation, its function is to effectively extract the node voltage characteristics under the N state. Its input characteristics include the power flow node voltage amplitude characteristic vector U under the N state. in , voltage phase angle characteristic vector θ in , injected active power characteristic vector P in and injected reactive power eigenvector Q in ,The input feature X of the embedded graph convolution submodule 1 of the power flow equation is specifically expressed as follows.

[0165] X={U in ,θ in , P in , Q in} (1)

[0166] In the formula, U i is the voltage amplitude of the ith node; U in ,θ in , P in , n is the number of nodes in the power system.

[0167] The forward propagation expression (3) of the embedded graph convolution submodule 1 in the power flow equation embedded in the N-state power flow physical model can be derived from the power flow equation in the N-state as shown in (2).

[0168]

[0169] Where P i , Q i are the injected active power and injected reactive power of the ith node respectively; G ij , B ij Represent the elements of the i-th row and j-th column in the conductance matrix G and the susceptance matrix B respectively; θ ij represents the voltage phase angle difference from node i to node j; N(i) represents the adjacent set of node i.

[0170]

[0171] Where l represents the number of graph convolution layers, l = 1, 2, ..., m, and m is the total number of graph convolution layers in the graph convolution submodule 1 embedded in the power flow equation; are the voltage amplitude characteristics and phase angle characteristics of the N-th state output by the ,th graph convolution layer, F l is the dimension of the output feature of the lth layer; σ is the activation function.

[0172] When l = 1, U N1 ,θ N1 is the input feature of the graph convolution submodule 1 embedded in the power flow equation, that is, UN1 =U in ,θ N1 =θ in , we can further forward propagate from the 1st layer to the mth layer based on formula (3), and obtain the output characteristics of the graph convolution submodule 1 embedded in the power flow equation: It can be expressed by Y1 as follows.

[0173] Y1={U Nm ,θ Nm} (4)

[0174] For the embedded graph convolution submodule 2 of the power flow equation, its function is to effectively extract the node voltage characteristics under the N / N-1 state. Its input feature matrix is ​​the feature matrix Y1 output by the embedded graph convolution submodule 1 of the power flow equation. Based on the power flow equation under the N state shown in formula (2), the forward propagation expression of the embedded graph convolution submodule 2 of the power flow equation is derived as shown in formula (5).

[0175]

[0176] In the formula, θ=[θ1,θ1,…,θ n ] T ; P = [P1, P2, ... P n ] T ; Q = [Q1, Q2, ... Q n ] T ; D(*) extracts the diagonal elements in * and constructs a diagonal matrix; * N-1，k Indicates the * corresponding to the N state when the kth branch is disconnected.

[0177] Through the forward propagation expression shown in equation (5), the output features of the graph convolution submodule 2 embedded in the power flow equation include: K node voltage amplitude feature matrices U under N-1 states N-1，k and phase angle characteristic matrix θ N-1，k ,in It can be expressed by Y2 as follows.

[0178] Y2={U N-1，k ,θ N-1，k} (6)

[0179] In the formula, k = 1, 2, ..., K, K is the number of N-1 states considered.

[0180] For the embedded graph convolution submodule of the branch power equation, its function is to effectively extract the branch power features under the N / N-1 state. Its input features are the feature matrix Y1 output by the embedded graph convolution submodule 1 of the power flow equation and the feature matrix Y2 output by the embedded graph convolution submodule 2 of the power flow equation. Based on the branch power equation under the N state shown in formula (7), the forward propagation expression of the embedded graph convolution submodule of the branch power equation is derived as shown in formula (8).

[0181]

[0182] In the formula, PF ij , QF ij are the active power and reactive power of branch ij respectively.

[0183]

[0184] In the formula, PF = [PF ij ] b×1 ; QF=[QF ij ] b×1 ; G br =[G ij ] b×1 ; B br =[B ij ] b×1 ; θ=[θ i ] n×1 ; b is the total number of branches; diag(*) is a diagonal matrix constructed with * as the diagonal element; ω N ,ω N-1，k are the corresponding ω in the N state and the kth N-1 state respectively.

[0185] Through the forward propagation expression shown in equation (8), the output features of the graph convolution submodule embedded in the branch power equation include: the branch active power feature matrix P under N states N , reactive power characteristic matrix Q N And the branch active power characteristic matrix P under K N-1 states N-1，k , reactive power characteristic matrix QN-1,k, where It can be expressed by Y3 as follows.

[0186] Y3={P N , Q N , P N-1，k , Q N-1，k} (9)

[0187] The node feature prediction convolution submodule is used to predict the node voltage features in N states and K N-1 states. The input feature is the feature matrix Y1 output by the convolution submodule 1 embedded in the power flow equation. Nm ,θ Nm} and the characteristic matrix Y2 output by the graph convolution submodule 2 embedded in the power flow equation = {U N-1，k ,θ N-1，k}, the convolution forward propagation formula is shown in formula (10).

[0188]

[0189] Where, l = 1, 2, ..., c1, c1 is the total number of convolutional layers in the node feature prediction convolutional submodule; U l ,θ l They are respectively the voltage amplitude feature and phase angle feature of the th convolutional layer input; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, When l>1, g l is the number of convolution kernel channels in layer l.

[0190] When l = 1, U 1 =[U Nm , U N-1，k ],θ 1 =[θ Nm ,θ N-1，k ], k = 1, 2, ..., K, and forward propagation formula (23) is used to propagate from the 1st layer to the c1th layer layer by layer, and finally the output feature can be obtained: the node voltage amplitude feature vector under the N state Phase angle eigenvector And the node voltage amplitude characteristic vectors under K N-1 states Phase angle eigenvector Its available U o and θ o It is expressed as follows.

[0191]

[0192] The branch feature prediction convolution submodule is used to predict the branch power features in N states and K N-1 states. Its input feature is the feature matrix Y3 output by the embedded graph convolution submodule of the branch power equation = {PF N , QF N , PF N-1，k , QF N-1，k}, and its convolution forward propagation formula is shown in formula (12).

[0193]

[0194] Where l = 1, 2, ..., c2, c2 is the total number of convolutional layers in the branch feature prediction convolutional submodule; PF l , QF l They are the branch active power features and branch reactive power features input to the ,th convolutional layer respectively; are the convolution kernel and bias of the convolution layer, which need to be obtained through training. When l=1, When l>1, r l is the number of convolution kernel channels in layer l.

[0195] When l = 1, PF 1 =[PF N , PF N-1，k ], QF 1 =[QF N , QF N-1，k ], k = 1, 2, ..., K, according to the forward propagation formula (25), forward propagation is performed layer by layer from the 1st layer to the c2th layer, and the output feature is finally obtained: the branch active power feature vector under the N state Branch reactive power characteristic vector And the branch active power characteristic vector in the N-1 state Branch reactive power characteristic vector Its available PF o and QF o It is expressed as follows.

[0196]

[0197] based on Figure 1 The N / N-1 power flow coupling relationship model architecture of the operation mode and the forward propagation expression of each module can be obtained by using the power flow training samples under the N / N-1 state corresponding to the operation mode and using training algorithms such as gradient descent: The N / N-1 power flow coupling relationship model f of the operation mode can be obtained by the input feature vector X={U in ,θ in , P in , Q in} Directly map to obtain the output feature vector Y in the N / N-1 state o = {U o ,θ o , PF o , QF o}, as shown in formula (27).

[0198] Y o =f(X) (14)

[0199] Intelligent confrontation adjustment method of operation mode based on N / N-1 power flow coupling relationship

[0200] (1N / N-1 state, the formula for the confrontation process of the flow over-limit to the input characteristics is derived

[0201] Based on the N / N-1 power flow coupling relationship, the present invention can derive the formula for the countermeasure process of power flow over-limit to input characteristics under the N / N-1 state. First, the node voltage amplitude output vector U is calculated by formula (14). o And the branch active power output vector PF o , the node voltage over-limit vector can be further calculated and branch active power over-limit vector Specifically, as shown in formula (15), the output over-limit vector ΔY=[ΔU, ΔPF] can be obtained.

[0202]

[0203] Where U o，i PF o，i They are the node voltage amplitude characteristic vector U o And the branch active power characteristic vector PF o The i-th element of up，i , U down，i U o,i Upper and lower limits of PF up，i PF down，i PF o，i The upper and lower limits of .

[0204] When the output over-limit vector ΔY≠0, the current operation mode does not meet the static N-1 safety check, then the generator node injected active power, the generator node injected reactive power, the load, the voltage, etc. in the operation mode input characteristic X are gradually adjusted until the output over-limit vector ΔY=0. In order to obtain the adjustment strategy, this paper takes minJ=ΔY as the optimization objective function and introduces the idea of ​​the confrontation process to achieve it. Specifically, based on the sensitivity of the flow over-limit ΔY to the input characteristic X in the N / N-1 state And the steepest descent method, the formula for the confrontation process of the flow over-limit ΔY against the input feature X under the N / N-1 state is shown in formula (16), where ΔX is the input adjustment strategy vector obtained at each step.

[0205]

[0206] In the formula, ε is the adjustment coefficient.

[0207] (2 Operation mode adjustment strategy intelligent generation method

[0208] The process of the intelligent countermeasure adjustment method for the operation mode based on the N / N-1 power flow coupling relationship constructed by the present invention is as follows: Figure 2 As shown, the specific steps are as follows.

[0209] Step 1: Construct the N / N-1 power flow coupling relationship model f of the operation mode. First, collect the power flow training samples under the N / N-1 state corresponding to the operation mode; secondly, build Figure 1 The operating mode N / N-1 power flow coupling relationship model architecture is shown; then, the operating mode N / N-1 power flow coupling relationship model f is obtained based on training algorithms such as gradient descent.

[0210] Step 2: Determine the limit of the operation mode to be adjusted. First, based on the operation mode N / N-1 power flow coupling relationship model f and the input characteristic X of the operation mode to be adjusted, the output characteristic Y is obtained. o ; Secondly, calculate the output over-limit vector ΔY through formula (15); then, determine whether ΔY is 0. If ΔY = 0, it means that the operating mode to be adjusted does not exceed the limit, and jump to step 4. Otherwise, it means that the operating mode to be adjusted exceeds the limit (that is, it violates the static N-1 safety check), and step 3 needs to be continued.

[0211] Step 3: Update the input characteristic X of the operation mode to be adjusted based on the countermeasure process formula. First, calculate the sensitivity of the flow over-limit ΔY to the input characteristic X under the N / N-1 state Secondly, the adversarial process formula (16) is used to obtain the input adjustment strategy vector ΔX; then, the input feature X of the operating mode to be adjusted is updated by X=X-ΔX, and then jump to step 2.

[0212] Step 4: First, stop adjusting the operating mode to be adjusted; second, obtain the input adjustment strategy vector ΔX obtained in each step to accurately analyze the adjustment process between the operating modes before and after the adjustment; then, obtain the final input feature X, which is the operating mode that meets the static N-1 safety check after adjustment.

[0213] Embodiment 12:

[0214] A verification of an intelligent adversarial adjustment method based on the operation mode of physically embedded neural networks is as follows:

[0215] (1) Introduction to the simulation process

[0216] In this embodiment, the IEEE-30 system and a certain actual large power grid 341 node system are used for simulation. For power systems of different scales, the number of graph convolution / convolution layers of the graph convolution submodule 1 embedded in the power flow equation, the node feature prediction convolution submodule and the branch feature prediction convolution submodule in the operation mode N / N-1 power flow coupling relationship model architecture are different. The operation mode N / N-1 power flow coupling relationship model architecture settings constructed by the present invention for power systems of different scales are specifically shown in Table 1.

[0217] Table 1 Operation modes of power systems of different scales N / N-1 power flow coupling relationship model architecture settings

[0218]

[0219] In this simulation, Rectified Linear Unit (ReLU) is selected as the activation function, 10,000 training samples and 1,000 test samples are set, and the output feature of each sample considers 11 states, that is, K=I0.

[0220] (2) Simulation results and analysis

[0221] In order to verify the effectiveness of the intelligent adjustment method for the power grid operation mode proposed in this paper, this method is used to adjust the operation mode of the IEEE 30-node system and a real large power grid 341-node system with power flow exceeding the limit in the N / N-1 state. Each time, the injected active power and injected reactive power of the generator node in the input adjustment strategy vector ΔX are mainly selected for adjustment.

[0222] First, the proposed method is applied to adjust the pending operation mode with power flow exceeding the limit in the IEEE 30-bus system, and the adjustment steps are shown in Table 2. Table 2 shows the initial exceeding of the limit in the pending operation mode in the IEEE 30-bus system: the voltage of node 7 exceeds the limit in the N state and in the N-1 state with the 3rd, 9th and 10th branches disconnected, the maximum voltage exceeding the limit per unit value is 0.0032, and the overall voltage exceeding the limit per unit value is 0.0078; the active power of branches 4-6 exceeds the limit in the N-1 state with the 1st, 4th, 6th and 7th branches disconnected, the maximum active power exceeding the limit is 15MW, and the overall active power exceeding the limit is 32MW. Through the intelligent adjustment method proposed in this paper, five input adjustment strategies are obtained as shown in Table 2. After adjustment according to the first three input adjustment strategies, the overall voltage limit per unit value becomes 0.0196, and the overall active power limit is reduced from 32MW to 0MW, then the branch active power is no longer over-limited. After continuing to adjust the operation mode according to the last two input adjustment strategies, the overall voltage limit per unit value is reduced from 0.0196 to 0, then the node voltage is no longer over-limited. Finally, the Newton-Raphson method is used to verify the adjusted operation mode, which proves that the operation mode adjusted according to these five input adjustment strategies meets all static N-1 safety checks.

[0223] Table 2 Adjustment steps for the IEEE 30-bus system to be adjusted operation mode

[0224]

[0225] The total over-limit is the sum of all node voltage over-limit or branch active power over-limit

[0226] Similarly, the proposed method in this paper is applied to the operation mode to be adjusted in a certain actual large power grid 341 node system with power flow exceeding the limit, and the adjustment steps are shown in Table 3. Table 3 shows the initial exceeding of the limit of the operation mode to be adjusted in the 341 node system: the active power of branch 47-186 exceeds the limit in the N state and in the N-1 state of disconnecting the 3rd, 6th, 9th and 10th branches, and the active power of branch 52-306 exceeds the limit in the N state and in the N-1 state of disconnecting the 4th, 6th and 7th branches. The maximum active power exceeding the limit is 38MW, and the overall active power exceeding the limit is 181MW. The intelligent adjustment method proposed in this paper obtains the four input adjustment strategies shown in Table 3. After adjusting the operation mode according to the first input adjustment strategy, the overall active power exceeding the limit is reduced from 181MW to 84MW, and the maximum active power exceeding the limit is reduced from 38MW to 13MW. After the third adjustment according to the input adjustment strategy, the active power of branch 52-306 no longer exceeds the limit, and the overall active power limit is reduced from the initial 181MW to 12MW. Then, after the fourth input adjustment strategy given by this method is adjusted, the active power of branch 47-186 no longer exceeds the limit, and the overall active power limit is directly reduced from 12MW to 0MW. Finally, the Newton-Raphson method is used to verify the adjusted operation mode, which proves that after adjustment according to the four input adjustment strategies in Table 3, the operation mode meets all static N-1 safety checks.

[0227] Table 3 Adjustment steps of the waiting operation mode of the 341 node system of a real large power grid

[0228]

[0229] By applying the present invention to the IEEE 30-node system and a certain actual large power grid 341-node system, its effectiveness in adjusting the flow over-limit operation mode under the N / N-1 state is proved. Compared with the optimal flow method that can only obtain the final adjustment strategy, this method can accurately analyze the adjustment process between the operation modes before and after the adjustment, and provide detailed basis for the staff to make operation mode adjustment decisions.

[0230] In summary, the present invention proposes an intelligent confrontation adjustment method for operating mode based on a physical embedded neural network. First, the N / N-1 power flow coupling relationship model of the operating mode based on the physical embedded neural network is used to characterize the data-driven power flow coupling relationship under the N / N-1 state; secondly, for the operating mode with power flow exceeding the limit under the N / N-1 state, an intelligent confrontation adjustment method for the operating mode based on the N / N-1 power flow coupling relationship is proposed to obtain a precise adjustment strategy for the operating mode to ensure that it meets the static N-1 safety check; finally, the effectiveness of the method proposed in the present invention is verified through example simulation analysis. Therefore, the present invention can be used for intelligent adjustment of the operating mode in the power system.

Claims

1. A method for adjusting the operation mode of intelligent confrontation based on a physical embedded neural network, characterized in that: The following steps are involved: 1) Based on the physical embedded neural network, a N / N-1 power flow coupling relationship model is constructed to characterize the data-driven power flow coupling relationship under the N / N-1 state; 2) Construct the input feature X of the operation mode to be adjusted, and input it into the operation mode N / N-1 power flow coupling relationship model to obtain the output feature Y o ; 3) Based on the output feature Y o , calculate the output limit-crossing vector ΔY; 4) Determine whether the output over-limit vector ΔY is 0. If ΔY=0, it is determined that the operation mode to be adjusted does not exceed the limit, and jump to step 6). Otherwise, it is determined that the operation mode to be adjusted exceeds the limit, and enter step 5); 5) Calculate the input adjustment strategy vector ΔX according to the adversarial process formula, adjust the input feature X=X-ΔX, and return to step 2). 6) With the current output feature Y o The corresponding input feature X is used as the operation mode that satisfies the static N-1 safety check.

2. According to claim 1, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network is characterized in that: The operation mode N / N-1 power flow coupling relationship model includes a power flow equation embedded graph convolution submodule, a power flow equation embedded graph convolution submodule, a branch power equation embedded graph convolution submodule, a node feature prediction convolution submodule and a branch feature prediction convolution submodule; The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under N states; The graph convolution submodule embedded in the power flow equation is used to extract node voltage characteristics under the N / N-1 state; The graph convolution submodule embedded in the branch power equation is used to extract branch power features under the N / N-1 state; The node feature prediction convolution submodule is used to predict node voltage features in N states and K N-1 states; The branch feature prediction convolution submodule is used to predict branch power features in N states and K N-1 states.

3. According to claim 2, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network, characterized in that: The input of the graph convolution submodule embedded in the power flow equation is as follows: X={U in ,θ in ,P in ,Q in } (1) In the formula, U i is the voltage amplitude of the ith node; U in ,θ in , P in , They represent the voltage amplitude characteristic vector, voltage phase angle characteristic vector, injected active power characteristic vector, and injected reactive power characteristic vector of the power flow node under the N state respectively; n is the number of nodes in the power system; The forward propagation expression of the graph convolution submodule embedded in the power flow equation is derived by the power flow equation under N states; The power flow equation in N state is as follows: Where P i , Q i are the injected active power and injected reactive power of the ith node respectively; G ij , B ij denote the elements of the i-th row and j-th column in the conductance matrix G and the susceptance matrix B respectively; θ ij represents the voltage phase angle difference from node i to node j; N(i) represents the adjacent set of node i; U i , U j is the voltage amplitude of the i-th node; The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows: Where l represents the number of graph convolution layers, l = 1, 2, ..., m, m is the total number of graph convolution layers in the graph convolution submodule 1 embedded in the power flow equation; U Nl , are the voltage amplitude characteristics and phase angle characteristics of the N states output by the lth graph convolution layer, F l The dimension of the output feature of the lth layer; σ is the activation function; is the weight coefficient corresponding to the voltage amplitude characteristic and phase angle characteristic; U Nl+1 ,θ Nl+1 The voltage amplitude and phase angle features under N states output by the l+1th graph convolution layer; The output Y1 of the graph convolution submodule embedded in the power flow equation is as follows: Y1={U Nm ,θ Nm } (4) Where U Nm , is the output of the mth convolutional layer.

4. According to claim 2, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network, characterized in that: The input of the power flow equation embedded graph convolution submodule is the output of the power flow equation embedded graph convolution submodule; The forward propagation expression of the graph convolution submodule embedded in the power flow equation is as follows: In the formula, θ=[θ1,θ1,…,θ n ] T ; P = [P1, P2, ... P n ] T ; Q=[Q1,Q2,…Q n ] T ; D(*) is to extract the diagonal elements in * and construct a diagonal matrix; χ, χ N-1,k is the propagation matrix; The output Y2 of the graph convolution submodule embedded in the power flow equation is as follows: Y2={U N-1,k ,θ N-1,k } (6) Where k = 1, 2, ..., K; K is the number of states considered N-1; U N-1,k , They respectively represent the node voltage amplitude characteristic matrix and phase angle characteristic matrix in the N-1 state.

5. According to claim 2, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network, characterized in that: The input of the branch power equation embedded graph convolution submodule includes the output of the power flow equation embedded graph convolution submodule and the output of the power flow equation embedded graph convolution submodule; The forward propagation expression of the graph convolution submodule embedded in the branch power equation is derived by the branch power equation in N states; The branch power equation in the N state is as follows: In the formula, PF ij , QF ij are the active power and reactive power of branch ij respectively; The forward propagation expression of the graph convolution submodule embedded in the branch power equation is as follows: In the formula, PF = [PF ij ] b×1 ; QF=[QF ij ] b×1 ; G br =[G ij ] b×1 ; B br =[B ij ] b×1 ; b is the total number of branches; diag(*) is a diagonal matrix constructed with * as the diagonal element; ω N ,ω N-1,k are the corresponding propagation matrices in the N state and the kth N-1 state respectively; A a is the current; The output Y3 of the graph convolution submodule embedded in the branch power equation is as follows: Y3={P N ,Q N ,P N-1,k ,Q N-1,k } (9) Where P N , Q N , P N-1,k , They respectively represent the branch active power characteristic matrix and reactive power characteristic matrix in the N state, and the branch active power characteristic matrix and reactive power characteristic matrix in the N-1 state.

6. According to claim 2, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network, characterized in that: The input of the node feature prediction convolution submodule includes the output of the power flow equation embedded graph convolution submodule and the output of the power flow equation embedded graph convolution submodule; The forward propagation formula of the node feature prediction convolution submodule is as follows: Where l = 1, 2, ..., c1, c1 is the total number of convolutional layers in the node feature prediction convolutional submodule; U l ,θ l They are the voltage amplitude feature and phase angle feature of the input of the lth convolutional layer respectively; are the convolution kernel and bias of the lth convolution layer, which need to be obtained through training. When l = 1, U l , When l>1, g l is the number of convolution kernel channels of the lth layer; The output of the node feature prediction convolution submodule is as follows: Current l = 1 hour, u 1 =[u Nm ,U N-1,k ], θ 1 =[θ Nm ,θ N-1,k ], k = 1, 2, …,; Through the forward propagation formula, we propagate forward layer by layer from the 1st layer to the c1th layer, and finally get the output feature: the node voltage amplitude feature vector under the N state Phase angle eigenvector And K node voltage amplitude characteristic vectors under N-1 states Phase angle eigenvector Right now: In the formula, They represent the node voltage amplitude characteristic vector, phase angle characteristic vector in the N state, the node voltage amplitude characteristic vector, phase angle characteristic vector in the N-1 state respectively; U o ,θ o is the output feature.

7. According to claim 2, a method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network, characterized in that: The input of the branch feature prediction convolution submodule is the output of the branch power equation embedded graph convolution submodule; The convolution forward propagation formula of the branch feature prediction convolution submodule is as follows: Where l = 1, 2, ..., c2, c2 is the total number of convolutional layers in the branch feature prediction convolutional submodule; PF l , QF l They are the branch active power features and branch reactive power features input to the lth convolutional layer respectively; are the convolution kernel and bias of the lth convolution layer, which need to be obtained through training. When l = 1, PF l , When l>1, r l is the number of convolution kernel channels of the lth layer; The branch feature predicts the output PF of the convolution submodule o , QF o As shown below: In the formula, They respectively represent the branch active power characteristic vector, branch reactive power characteristic vector in the N state, the branch active power characteristic vector, and branch reactive power characteristic vector in the N-1 state.

8. According to claim 1, the method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network is characterized in that: The simplified expression of the N / N-1 power flow coupling relationship model of operation mode is as follows: Y o =f(X) (14) In the formula, X={U in ,θ in ,P in ,Q in } is input, Y o = {U o ,θ o ,PF o ,QF o } is the output. f() is the function.

9. The method for adjusting the operation mode of intelligent confrontation based on physical embedded neural network according to claim 1 is characterized in that: Output limit-crossing vector ΔY=[ΔU,ΔPF], that is: Where U o,i PF o,i They are the node voltage amplitude characteristic vector U o And the branch active power characteristic vector PF o The i-th element of up,i , U down,i U o,i Upper and lower limits of PF up,i PF down,i PF o,i The upper and lower limits of ΔU i , ΔPF i is the i-th out-of-limit vector.

10. The method for adjusting the operation mode of an intelligent confrontation based on a physical embedded neural network according to claim 1, characterized in that: The input adjustment strategy vector ΔX is as follows: In the formula, ε is the adjustment coefficient; It is the sensitivity of the power flow over-limit ΔY to the input characteristic X under the N / N-1 state.

Images