An intelligent adjustment method for power flow over-limit based on confrontation process

Abstract

A method for intelligent adjustment of over-limit current based on a confrontation process, comprising the following steps: 1) constructing current data features characterizing a variety of N-1 scenarios to obtain a training set; 2) constructing a current calculation model based on a fully connected neural network; 3) using the training set to train the current calculation model to construct an input-output function relationship; 4) obtaining a current state to be adjusted that does not meet the N-1 check as an over-limit current sample; 5) inputting the over-limit current sample into the current calculation model, calculating the output and over-limit value of the current sample, and judging whether the N-1 check is met, if not, proceeding to step 6); 6) calculating the input adjustment amount based on the gradient descent method and the iterative strategy, adjusting the over-limit current sample, and returning to step 5), until the over-limit current sample meets the N-1 check, the confrontation process is terminated. The present invention can automatically generate adjustment measures to adjust the over-limit current to a current state that meets the N-1 check based on the confrontation process.

Description

Technical Field

[0001] The present invention relates to the field of power systems and automation thereof, and in particular to an intelligent adjustment method for power flow over-limit based on a confrontation process. Background Art

[0002] Flow regulation plays an important role in the daily operation of power systems. Its goal is to generate a flow regulation method to adjust the node injection power when the voltage amplitude or branch power exceeds the limit. However, with the expansion of the power system scale and the increase of renewable energy, the power system has become more and more complex, which makes the flow exceeding the limit more complicated. In actual operation, the operator of the power system determines the adjustment plan based on expert experience and then performs N-1 verification. For such a complex power system, it is difficult to formulate an adjustment plan that meets the N-1 verification based on expert experience alone.

[0003] At present, power flow adjustment methods can be mainly divided into two categories, one is direct optimization method and the other is data-driven method. Among them, the direct optimization method is committed to optimizing the adjustment strategy through numerical calculation and modeling of N-1 verification to achieve the adjustment of power flow state. However, the model constructed by the direct optimization method contains many nonlinear and non-convex power flow constraints, making it difficult to solve in a limited time. Existing studies simplify these nonlinear and non-convex power flow bundles by linearization to accelerate the solution. Although the linearization error is very small, there is a loss of accuracy, and the iterative calculation in the numerical solution method is very time-consuming during N-1 verification. In actual power systems, the calculation time is limited, so the direct optimization method is difficult to be widely used in actual scenarios.

[0004] Data-driven methods can use real-time data and advanced technologies to quickly analyze power flow states. However, existing research on data-driven methods mainly focuses on the power flow analysis level and cannot directly generate power flow adjustment strategies. The adversarial process can achieve the goal of directly generating power flow adjustment methods based on power flow calculation neural networks. The adversarial process is a widely used technique in machine learning and is often used to generate adversarial samples. It calculates the gradient of the neural network error and the input features and updates the input features by maximizing the error of the neural network. According to this idea, the power flow states that do not meet the N-1 check can be reduced by optimizing the input features. However, existing data-driven methods have not yet fully utilized the adversarial process to generate power flow adjustment methods. Summary of the invention

[0005] The object of the present invention is to provide a method for intelligently adjusting power flow over-limit based on a confrontation process, comprising the following steps:

[0006] 1) Construct the power flow input and output features that represent various N-1 scenarios, and use the Monte Carlo method to obtain the training set.

[0007] 2) Construct a power flow calculation model based on a fully connected neural network to characterize the functional relationship between power flow input and output.

[0008] 3) Use the training set to train the power flow calculation model to obtain the trained power flow calculation model.

[0009] 4) Obtain the to-be-adjusted power flow state that does not satisfy the N-1 check as an out-of-limit power flow sample.

[0010] 5) Input the over-limit power flow sample into the power flow calculation model, calculate the output and over-limit value of the power flow sample, and determine whether the N-1 check is satisfied. If not, proceed to step 6).

[0011] 6) Based on the gradient descent method and the iterative strategy, the input adjustment amount is calculated, and after adjusting the over-limit power flow sample, the process returns to step 5) until the over-limit power flow sample satisfies the N-1 check, and the confrontation process ends.

[0012] Furthermore, the power flow input feature X that characterizes multiple N-1 scenarios is as follows:

[0013] X={P1,P2,···,P n ,Q1,Q2,···,Q n} (1)

[0014] Where P j represents the injected active power of the jth node. j represents the injected reactive power of the jth node. j = 1, 2, ..., n, where n represents the number of nodes in the power system.

[0015] The power flow output characteristics representing various N-1 scenarios are as follows:

[0016] Y out ={y0,y1,y2,···,y t} (2)

[0017] y i = {v i1 ,v i2 ,···,v in ,θ i1 ,θ i2 ,···,θ in ,p fi1 ,p fi2 ,···,p fib ,q fi1 ,q fi2 ,···,q fib} (3)

[0018] Where Y outTo characterize the power flow output characteristics of various N-1 scenarios, y i represents the power flow output characteristics under different topological structures, y0 is the output characteristic under the original topological structure, when i>0, y i is the power flow output characteristics under different N-1 scenarios. i=0,1,2...,t, t is the number of N-1 scenarios. ij Represents the voltage amplitude of the jth node in the i-th scenario. ij represents the phase angle of the jth node in the i-th scenario. fia It represents the active power of the ath branch in the i-th scenario. fia It represents the branch reactive power of the ath branch in the ith scenario. a=1,2...,b, where b represents the number of branches in the power system.

[0019] Furthermore, the output of the power flow calculation model is as follows:

[0020]

[0021]

[0022] Where Y out To characterize the power flow output characteristics of various N-1 scenarios. X is the power flow sample input into the input layer. r-1 Represents the input features entering the rth neuron layer. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. s is the activation function. W r 、b r They represent the weight and bias of the r-th layer respectively.

[0023] Furthermore, the weights and biases of the power flow calculation model are as follows:

[0024]

[0025]

[0026] Where W r 、b r They represent the weight and bias of the r-th layer respectively. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. L is the loss function of the neural network. η represents the learning rate.

[0027] Furthermore, the calculation formula of the over-limit value is as follows:

[0028] Δy r ={Δy[0],Δy[1],…,Δy[m]} (8)

[0029]

[0030] In the formula, Δy r Y is the threshold value vector of the power flow sample output calculated at the rth iteration. Δy[x] represents the threshold value corresponding to the power flow output value indexed as x. x=1,2,...,m, where m represents the total number of output features. out [x] indicates the output feature with index x. outmax [x] represents the upper limit value of the output feature with index x. outmin [x] represents the lower limit of the output feature indexed by x.

[0031] Furthermore, the power flow input adjustment amount is calculated through a confrontation process of a neural network.

[0032] Further, the method for determining whether the over-limit power flow samples satisfy the N-1 check includes that the over-limit values ​​are all 0.

[0033] Furthermore, the calculation formula of the power flow input adjustment amount is as follows:

[0034] Δx r = J·Δy r (10)

[0035] In the formula, Δx r is the input adjustment of the rth layer neural network. Δy r It is the over-limit value vector output by the power flow sample of the r-th layer neural network.

[0036] Among them, the Jacobian matrix J is as follows:

[0037]

[0038] In the formula, a r is the output of the rth hidden layer. r=1,2,...,k, where k represents the number of layers in the neural network. out is the output feature. X is the power flow sample input into the input layer. is the derivative of the output layer activation function. is the derivative of the activation function of the r1th hidden layer with respect to the r2th hidden layer. z represents the weighted input of the input layer. is the hidden layer weight of the input layer.

[0039] Furthermore, after adjusting the over-limit power flow samples, the input of the over-limit power flow samples is as follows:

[0040] X r+1 =X r -Δx r+1 (12)

[0041] Where, X r , X r+1 are the power flow inputs of the rth and r+1th layers respectively. r+1 is the input adjustment amount of the r+1th layer.

[0042] Furthermore, the final adjustment amount of the over-limit power flow sample is as follows:

[0043] ΔX=X e -X0 (13)

[0044] In the formula, e represents the final number of iterations of the adversarial process. e It is the final power flow input that satisfies N-1 verification after adjustment by the confrontation process. X0 represents the original input of the over-limit power flow sample without the adjustment process. ΔX represents the power flow adjustment amount finally found by the adjustment method.

[0045] The technical effect of the present invention is unquestionable. The power flow calculation model based on the fully connected neural network (FCN) proposed in the present invention constructs a functional relationship between power flow input and output that satisfies a variety of N-1 scenarios, and can quickly and accurately analyze the power flow status under a variety of N-1 scenarios at the same time.

[0046] The intelligent adjustment method for over-limit power flow based on the confrontation process proposed in the present invention can automatically generate an over-limit power flow adjustment strategy for the over-limit power flow, so that it can be quickly adjusted to the power flow state that meets the N-1 verification, providing a new perspective for solving the power flow adjustment problem.

[0047] The present invention can be widely applied to power flow adjustment analysis in various industries of power systems.

[0048] The power flow over-limit adjustment method proposed in the present invention can intelligently adjust the over-limit power flow to satisfy the N-1 check, and no manual adjustment of node injection power is required. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The schematic diagram of the framework for intelligent adjustment of power flow over-limit based on the confrontation process;

[0050] Figure 2 It is a flow chart of the power flow over-limit adjustment method based on the confrontation process;

[0051] Figure 3 It is the over-limit situation diagram of the over-limit power flow sample; Figure 3 (a) is the limit-crossing diagram of sample 23; Figure 3 (b) is the limit-crossing diagram of sample 174;

[0052] Figure 4 This is a graph showing the change of the over-limit value with the number of iterations; Figure 4 (a) is the change diagram of the over-limit value change of sample 23; Figure 4 (b) is a graph showing the variation of the limit value of sample 174. DETAILED DESCRIPTION

[0053] 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.

[0054] Embodiment 1:

[0055] See also Figures 1 to 4 , a power flow over-limit intelligent adjustment method based on a confrontation process, comprising the following steps:

[0056] 1) Construct the power flow input and output features that represent various N-1 scenarios, and use the Monte Carlo method to obtain the training set.

[0057] 2) Construct a power flow calculation model based on a fully connected neural network to characterize the functional relationship between power flow input and output.

[0058] 3) Use the training set to train the power flow calculation model to obtain the trained power flow calculation model.

[0059] 4) Obtain the to-be-adjusted power flow state that does not satisfy the N-1 check as an out-of-limit power flow sample.

[0060] 5) Input the over-limit power flow sample into the power flow calculation model, calculate the output and over-limit value of the power flow sample, and determine whether the N-1 check is satisfied. If not, proceed to step 6).

[0061] 6) Based on the gradient descent method and the iterative strategy, the input adjustment amount is calculated, and after adjusting the over-limit power flow sample, the process returns to step 5) until the over-limit power flow sample satisfies the N-1 check, and the confrontation process ends.

[0062] Embodiment 2:

[0063] A method for intelligently adjusting power flow over-limit based on a confrontation process, the main technical content of which is shown in Example 1. Further, the power flow input feature X characterizing multiple N-1 scenarios is as follows:

[0064] X={P1,P2,···,P n ,Q1,Q2,···,Q n} (1)

[0065] Where P j represents the injected active power of the jth node. jrepresents the injected reactive power of the jth node. j = 1, 2, ..., n, where n represents the number of nodes in the power system.

[0066] The power flow output characteristics representing various N-1 scenarios are as follows:

[0067] Y out ={y0,y1,y2,···,y t} (2)

[0068] y i = {v i1 ,v i2 ,···,v in ,θ i1 ,θ i2 ,···,θ in ,p fi1 ,p fi2 ,···,p fib ,q fi1 ,q fi2 ,···,q fib} (3)

[0069] Where Y out To characterize the power flow output characteristics of various N-1 scenarios, y i represents the power flow output characteristics under different topological structures, y0 is the output characteristic under the original topological structure, when i>0, y i is the power flow output characteristics under different N-1 scenarios. i=0,1,2...,t, t is the number of N-1 scenarios. ij Represents the voltage amplitude of the jth node in the i-th scenario. ij represents the phase angle of the jth node in the i-th scenario. fia It represents the active power of the ath branch in the i-th scenario. fia It represents the branch reactive power of the ath branch in the ith scenario. a=1,2...,b, where b represents the number of branches in the power system.

[0070] Embodiment 3:

[0071] A method for intelligently adjusting power flow over-limit based on a confrontation process, the main technical content of which is shown in any one of Embodiments 1 to 2. Further, the output of the power flow calculation model is as follows:

[0072]

[0073]

[0074] Where Y outTo characterize the power flow output characteristics of various N-1 scenarios. X is the power flow sample input into the input layer. r-1 Represents the input features entering the rth neuron layer. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. s is the activation function. W r 、b r They represent the weight and bias of the r-th layer respectively.

[0075] Embodiment 4:

[0076] A method for intelligently adjusting power flow over-limit based on a confrontation process, the main technical content of which is shown in any one of Embodiments 1 to 3. Further, the weight and bias of the power flow calculation model are as follows:

[0077]

[0078]

[0079] Where W r 、b r They represent the weight and bias of the r-th layer respectively. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. L is the loss function of the neural network. η represents the learning rate.

[0080] Embodiment 5:

[0081] A method for adjusting power flow over-limit intelligently based on a confrontation process, the main technical content of which is shown in any one of Embodiments 1 to 4. Further, the calculation formula of the over-limit value is as follows:

[0082] Δy r ={Δy[0],Δy[1],…,Δy[m]} (8)

[0083]

[0084] In the formula, Δy r Y is the threshold value vector of the power flow sample output calculated at the rth iteration. Δy[x] represents the threshold value corresponding to the power flow output value indexed as x. x=1,2,...,m, where m represents the total number of output features. out [x] indicates the output feature with index x. outmax [x] represents the upper limit value of the output feature with index x. outmin [x] represents the lower limit of the output feature indexed by x.

[0085] Embodiment 6:

[0086] A method for intelligently adjusting power flow over-limit based on an adversarial process, the main technical content of which can be found in any one of embodiments 1 to 5. Furthermore, the power flow input adjustment amount is calculated through an adversarial process of a neural network.

[0087] Embodiment 7:

[0088] A method for intelligently adjusting an over-limit flow based on a confrontation process, wherein the main technical content is shown in any one of embodiments 1 to 6. Further, a method for determining that an over-limit flow sample satisfies the N-1 check includes that all over-limit values ​​are 0.

[0089] Embodiment 8:

[0090] A method for adjusting power flow exceeding the limit intelligently based on a confrontation process, the main technical content of which is shown in any one of Embodiments 1 to 7. Further, the calculation formula of the power flow input adjustment amount is as follows:

[0091] Δx r = J·Δy r (10)

[0092] In the formula, Δx r is the input adjustment of the rth layer neural network. Δy r It is the over-limit value vector output by the power flow sample of the r-th layer neural network.

[0093] Among them, the Jacobian matrix J is as follows:

[0094]

[0095] In the formula, a r is the output of the rth hidden layer. r=1,2,...,k, where k represents the number of layers in the neural network. out is the output feature. X is the power flow sample input into the input layer. is the derivative of the output layer activation function. is the derivative of the activation function of the r1th hidden layer with respect to the r2th hidden layer. z represents the weighted input of the input layer. is the hidden layer weight of the input layer.

[0096] Embodiment 9:

[0097] A method for adjusting power flow exceeding limit intelligently based on a confrontation process, the main technical content of which is shown in any one of embodiments 1 to 8. Further, after adjusting the exceeding limit power flow sample, the input of the exceeding limit power flow sample is as follows:

[0098] X r+1 =X r -Δx r+1 (12)

[0099] In the formula, X r, X r+1 are the power flow inputs of the rth and r+1th layers respectively. r+1 is the input adjustment amount of the r+1th layer.

[0100] Embodiment 10:

[0101] A method for adjusting power flow exceeding the limit based on a confrontation process, the main technical content of which is shown in any one of embodiments 1 to 9. Further, the final adjustment amount of the exceeding power flow sample is as follows:

[0102] ΔX=X e -X0 (13)

[0103] In the formula, e represents the final number of iterations of the adversarial process. e It is the final power flow input that satisfies N-1 verification after adjustment by the confrontation process. X0 represents the original input of the over-limit power flow sample without the adjustment process. ΔX represents the power flow adjustment amount finally found by the adjustment method.

[0104] Embodiment 11:

[0105] See also Figures 1 to 4 , a power flow over-limit intelligent adjustment method based on a confrontation process, comprising the following steps:

[0106] 1) Construct the power flow input and output features that represent various N-1 scenarios, and use the Monte Carlo method to obtain the training set.

[0107] 2) Construct a power flow calculation model based on a fully connected neural network (FCN) to characterize the functional relationship between power flow input and output.

[0108] 3) Use the training set to train the power flow calculation model to obtain the trained power flow calculation model.

[0109] 4) Obtain the to-be-adjusted power flow state that does not satisfy the N-1 check as an out-of-limit power flow sample.

[0110] 5) Input the over-limit power flow sample into the power flow calculation model, calculate the output and over-limit value of the power flow sample, and determine whether the N-1 check is satisfied. If not, proceed to step 6).

[0111] 6) Based on the gradient descent method and the iterative strategy, the input adjustment amount is calculated, and after adjusting the over-limit power flow sample, the process returns to step 5) until the over-limit power flow sample satisfies the N-1 check, and the confrontation process ends.

[0112] Finally, by simulating the IEEE 30 system, the proposed method can automatically generate a power flow over-limit adjustment strategy for the over-limit power flow, so that it can be quickly adjusted to the power flow state that meets the N-1 check, verifying the effectiveness of the method proposed in this invention.

[0113] The power flow over-limit adjustment framework designed by the present invention based on the confrontation process is as follows: Figure 1 As shown in Figure 1, the proposed method has two stages, and its specific contents are as follows: In the first stage, it is necessary to train a power flow calculation model that can simultaneously calculate the power flow results in multiple topologies to reflect the functional relationship between the power flow input characteristics and output characteristics. Therefore, the Monte Carlo method is first used to generate power flow samples, and then the functional relationship between the power flow input and output under various N-1 topological structures is constructed based on the fully connected neural network (FCN); in the second stage, when any power flow state does not meet the N-1 check, the power flow calculation model constructed in the first stage is used to predict the power flow results and calculate the over-limit value, and the power flow adjustment strategy is generated by using the confrontation process of the neural network to adjust the node injection power of the power system so that the power flow state meets the N-1 check.

[0114] Embodiment 12:

[0115] A method for intelligently adjusting flow over-limit based on a confrontation process. The main technical content is shown in Example 11. Furthermore, in order to simultaneously characterize the input and output relationship of the flow in multiple N-1 scenarios, special input and output feature vectors are first constructed, and then a flow calculation model based on a neural network is obtained through training.

[0116] In order to adjust the over-limit power flow to a power flow state that satisfies the N-1 check by adjusting the node injection power, the input feature vector is selected as the node injection active power and the node injection reactive power. The power flow input feature X that characterizes multiple N-1 scenarios is as follows:

[0117] X={P1,P2,···,P n ,Q1,Q2,···,Q n} (1)

[0118] Where P j represents the injected active power of the jth node. j represents the injected reactive power of the jth node. j = 1, 2, ..., n, where n represents the number of nodes in the power system.

[0119] In order to simultaneously calculate the power flow results under various N-1 scenarios, the voltage amplitude, phase angle, branch active power and branch reactive power under the original topology and various N-1 scenarios are selected as output features. The power flow output features characterizing various N-1 scenarios are as follows:

[0120] Yout ={y0,y1,y2,···,y t} (2)

[0121] y i = {v i1 ,v i2 ,···,v in ,θ i1 ,θ i2 ,···,θ in ,p fi1 ,p fi2 ,···,p fib ,q fi1 ,q fi2 ,···,q fib} (3)

[0122] Where Y out To characterize the power flow output characteristics of various N-1 scenarios, y i represents the power flow output characteristics under different topological structures, y0 is the output characteristic under the original topological structure, when i>0, y i is the power flow output characteristics under different N-1 scenarios. i=0,1,2...,t, t is the number of N-1 scenarios. ij Represents the voltage amplitude of the jth node in the i-th scenario. ij represents the phase angle of the jth node in the i-th scenario. fia It represents the active power of the ath branch in the i-th scenario. fia It represents the branch reactive power of the ath branch in the ith scenario. a=1,2...,b, where b represents the number of branches in the power system.

[0123] Embodiment 13:

[0124] A method for intelligently adjusting power flow over-limit based on an adversarial process, the main technical content of which is shown in any one of Examples 11 to 12. Further, FCN is a basic neural network structure. It consists of multiple neuron layers, each neuron is connected to each neuron in the previous layer, which means that each neuron receives input from all neurons in the previous layer. In addition, each neuron has a nonlinear activation function, which enables FCN to handle complex nonlinear relationships. The output of the power flow calculation model is shown below:

[0125]

[0126]

[0127] Where Y out To characterize the power flow output characteristics of various N-1 scenarios. X is the power flow sample input into the input layer.r-1 Represents the input features entering the rth neuron layer. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. s is the activation function. W r 、b r They represent the weight and bias of the r-th layer respectively.

[0128] Embodiment 14:

[0129] A method for intelligent adjustment of power flow over-limit based on an adversarial process, the main technical content of which is shown in any one of Examples 11 to 13. Further, after the structure of the neural network is established, training is required to determine the internal parameters of the neural network. The training of a fully connected neural network is divided into two processes: forward propagation and backward propagation. Forward propagation is the calculation of the loss function value after the data is input to the output. During the forward propagation process, the input data is propagated through each layer of the neural network until the predicted output is obtained. Backward propagation is an optimization process that uses the gradient descent method to reduce the loss function value generated by the forward propagation, thereby optimizing and updating the parameters. The weights and biases of the power flow calculation model are as follows:

[0130]

[0131]

[0132] Where W r 、b r They represent the weight and bias of the r-th layer respectively. is the feedforward propagation function of the rth layer. r=1,2,...,k, where k represents the number of layers of the neural network. L is the loss function of the neural network. η represents the learning rate.

[0133] Finally, the trained neural network can characterize the functional relationship between power flow input and output, that is, the relationship between node injection power and voltage amplitude, phase angle, branch active power, and branch reactive power in N and N-1 cases, laying the foundation for the power flow over-limit adjustment method based on the confrontation process.

[0134] Embodiment 15:

[0135] A method for adjusting power flow exceeding the limit based on the confrontation process, the main technical content of which is shown in any one of embodiments 11 to 14. Further, the power flow inequality constraint is used to check whether the power flow exceeds the limit, including the branch power flow constraint and the voltage amplitude constraint in the N-1 scenario, and the over-limit value vector Δy output by the power flow sample can be calculated. i The calculation formula of the over-limit value is as follows:

[0136] Δy r={Δy[0],Δy[1],…,Δy[m]} (8)

[0137]

[0138] In the formula, Δy r Y is the threshold value vector of the power flow sample output calculated at the rth iteration. Δy[x] represents the threshold value corresponding to the power flow output value indexed as x. x=1,2,...,m, where m represents the total number of output features. out [x] indicates the output feature with index x. outmax [x] represents the upper limit value of the output feature with index x. outmin [x] represents the lower limit of the output feature indexed by x.

[0139] Embodiment 16:

[0140] A method for intelligently adjusting over-limit power flow based on an adversarial process, the main technical content of which is shown in any one of embodiments 11 to 15. Further, the training of the neural network is to optimize the internal parameters of the network, and the adversarial process is to further adjust the input characteristics according to the determined internal parameters of the neural network so that the input characteristics achieve the expected purpose. Therefore, the adversarial process of the neural network is introduced into the power flow adjustment, and the node injection power can be continuously adjusted using the gradient iteration strategy, so that the over-limit power flow is finally adjusted to the power flow state that satisfies the N-1 check.

[0141] The flowchart of the power flow over-limit adjustment method based on the confrontation process is as follows: Figure 2 As shown, the power flow input adjustment amount is calculated through the adversarial process of the neural network.

[0142] Embodiment 17:

[0143] A method for adjusting the power flow over-limit based on the confrontation process, the main technical content of which is shown in any one of Embodiments 11 to 16. Further, if the output element Δy i If it is positive, the output value corresponding to the position is higher than the specified upper limit; if it is negative, it is less than the specified lower limit; if it is equal to 0, it means that it does not exceed the limit.

[0144] The method of judging that the over-limit power flow samples satisfy the N-1 check includes that the over-limit values ​​are all 0.

[0145] Embodiment 18:

[0146] A method for intelligently adjusting power flow over-limit based on an adversarial process, the main technical content of which is shown in any one of Examples 11 to 17. Furthermore, according to the adversarial processing technology, the gradient of the power flow calculation model at the current input reflects how to adjust the power flow sample input, so that the power flow state can be adjusted in the direction of satisfying the N-1 check. The adjustment amount of the power injected by the calculation node is shown in formula (10). The gradient is the Jacobian matrix J. The Jacobian matrix of each iteration is calculated as shown in formula (11). The Jacobian matrix can link the over-limit value of the power flow sample with the input adjustment amount, thereby calculating the input adjustment amount corresponding to the over-limit value of the power flow under the current iteration.

[0147] The calculation formula of the power flow input adjustment amount is as follows:

[0148] Δx r = J·Δy r (10)

[0149] In the formula, Δx r is the input adjustment of the rth layer neural network. Δy r It is the over-limit value vector output by the power flow sample of the r-th layer neural network.

[0150] Among them, the Jacobian matrix J is as follows:

[0151]

[0152] In the formula, a r is the output of the rth hidden layer. r=1,2,...,k, where k represents the number of layers in the neural network. out is the output feature. X is the power flow sample input into the input layer. is the derivative of the output layer activation function. is the derivative of the activation function of the r1th hidden layer with respect to the r2th hidden layer. z represents the weighted input of the input layer. is the hidden layer weight of the input layer.

[0153] Embodiment 19:

[0154] A method for intelligently adjusting power flow over-limit based on a countermeasure process, the main technical content of which can be found in any one of Examples 11 to 18. Furthermore, based on the gradient descent method, the calculated power flow input adjustment amount is added to the sample input of the previous iteration as the power flow input of this iteration process.

[0155] After adjusting the over-limit power flow samples, the input of the over-limit power flow samples is as follows:

[0156] X r+1 =X r -Δx r+1 (12)

[0157] In the formula, X r , X r+1 are the power flow inputs of the rth and r+1th layers respectively. r+1 is the input adjustment amount of the r+1th layer.

[0158] Embodiment 20:

[0159] A method for adjusting power flow exceeding the limit based on a confrontation process, the main technical content of which is shown in any one of embodiments 11 to 19. Further, the final adjustment amount of the exceeding power flow sample is as follows:

[0160] ΔX=X e -X0 (13)

[0161] In the formula, e represents the final number of iterations of the adversarial process. e It is the final power flow input that satisfies N-1 verification after adjustment by the confrontation process. X0 represents the original input of the over-limit power flow sample without the adjustment process. ΔX represents the power flow adjustment amount finally found by the adjustment method.

[0162] Through the above power flow over-limit adjustment method, the over-limit power flow can be intelligently adjusted to meet the N-1 check, and there is no need to manually adjust the node injection power.

[0163] Embodiment 21:

[0164] See also Figures 1 to 4 , a power flow over-limit intelligent adjustment method based on confrontation process, including the following contents:

[0165] Firstly, the functional relationship between power flow input and output under various N-1 topological structures is constructed based on the power flow calculation model of fully connected neural network (FCN). Then, the power flow over-limit adjustment method is constructed based on the confrontation process of neural network. Finally, by simulating the IEEE 30 system, the proposed method can automatically generate a power flow over-limit adjustment strategy for the over-limit power flow, so that it can be quickly adjusted to the power flow state that meets the N-1 check, verifying the effectiveness of the method proposed in the present invention. The specific contents are as follows.

[0166] (1) Framework of power flow over-limit adjustment method

[0167] The power flow over-limit adjustment framework designed by the present invention based on the confrontation process is as follows: Figure 1As shown in Figure 1, the proposed method has two stages, and its specific contents are as follows: In the first stage, it is necessary to train a power flow calculation model that can simultaneously calculate the power flow results in multiple topologies to reflect the functional relationship between the power flow input characteristics and output characteristics. Therefore, the Monte Carlo method is first used to generate power flow samples, and then the functional relationship between the power flow input and output under various N-1 topological structures is constructed based on the fully connected neural network (FCN); in the second stage, when any power flow state does not meet the N-1 check, the power flow calculation model constructed in the first stage is used to predict the power flow results and calculate the over-limit value, and the power flow adjustment strategy is generated by using the confrontation process of the neural network to adjust the node injection power of the power system so that the power flow state meets the N-1 check.

[0168] (2) Constructing the functional relationship between power flow input and output

[0169] In order to simultaneously characterize the input and output relationship of power flow in multiple N–1 scenarios, special input and output feature vectors are firstly constructed, and then a power flow calculation model based on a neural network is obtained through training.

[0170] 1. Characteristic structure of N-1 power flow calculation

[0171] In order to adjust the over-limit power flow to the power flow state that satisfies the N-1 check by adjusting the node injection power, the input feature vectors are selected as the node injection active power and the node injection reactive power. The input features are as follows:

[0172] X={P1,P2,···,P n ,Q1,Q2,···,Q n} (1)

[0173] Where P i represents the injected active power of the ith node; Q i represents the injected reactive power of the ith node; n represents the number of nodes in the power system; the dimension of the input feature vector is 2n.

[0174] In order to simultaneously calculate the power flow results under various N-1 scenarios, the voltage amplitude, phase angle, branch active power and branch reactive power under the original topology and various N-1 scenarios are selected as output features, as shown in Equations (2)-(3).

[0175] Y out ={y0,y1,y 2, ··· , y t} (2)

[0176] y i = {v i1 ,v i2 ,···,v in ,θi1 ,θ i2 ,···,θ in ,p fi1 ,p fi2 ,···,p fib ,q fi1 ,q fi2 ,···,q fib} (3)

[0177] where y i represents the power flow output characteristics under different topological structures. When i=0, it is the output characteristics under the original topological structure. When i>0, it is the output characteristics under different N-1 scenarios. t is the number of N-1 scenarios. v ij represents the voltage amplitude of the jth node in the i-th scenario; θ ij represents the phase angle of the jth node in the i-th scenario; p fij represents the branch active power of the jth branch in the i-th scenario; q fij represents the branch reactive power of the j-th branch in the i-th scenario; b represents the number of branches in the power system.

[0178] 2. Power flow calculation model based on fully connected neural network

[0179] FCN is a basic neural network structure. It consists of multiple neuron layers, each neuron is connected to each neuron in the previous layer, which means that each neuron receives input from all neurons in the previous layer. In addition, each neuron has a nonlinear activation function, which enables FCN to handle complex nonlinear relationships. The mathematical expression of FCN can be expressed as formulas (4) and (5).

[0180]

[0181]

[0182] Where X is the power flow sample input into the input layer; X i-1 Represents the input features entering the i-th neuron layer; is the feedforward propagation function of the i-th layer; k represents the number of layers of the neural network; s is the activation function; W i and b i denote the weight and bias of the i-th layer respectively.

[0183] After establishing the structure of the neural network, training is required to determine the internal parameters of the neural network. The training of a fully connected neural network is divided into two processes: forward propagation and backward propagation. Forward propagation is the process of calculating the loss function value after the data is input to the output. During the forward propagation process, the input data is propagated through each layer of the neural network until the predicted output is obtained. Backward propagation is an optimization process that uses the gradient descent method to reduce the loss function value generated by the forward propagation, thereby optimizing and updating the parameters. The mathematical expressions of the weights and biases of FCN can be expressed as (6)-(7).

[0184]

[0185]

[0186] Where L is the loss function of the neural network; η represents the learning rate.

[0187] Finally, the trained neural network can characterize the functional relationship between power flow input and output, that is, the relationship between node injection power and voltage amplitude, phase angle, branch active power, and branch reactive power in N and N-1 cases, laying the foundation for the power flow over-limit adjustment method based on the confrontation process.

[0188] (3) Specific process of the intelligent adjustment method for power flow over-limit

[0189] The power flow over-limit adjustment method proposed in the present invention is based on the confrontation process of the neural network. The training of the neural network is to optimize the internal parameters of the network, and the confrontation process is to further adjust the input characteristics according to the determined internal parameters of the neural network so that the input characteristics can achieve the expected purpose. Therefore, the confrontation process of the neural network is introduced into the power flow adjustment, and the node injection power can be continuously adjusted using the gradient iteration strategy, so that the over-limit power flow is finally adjusted to the power flow state that meets the N-1 check. The flowchart of the power flow over-limit adjustment method based on the confrontation process is shown in FIG. Figure 2 The specific process is as follows.

[0190] 1) Step 1: Check the power flow over-limit status

[0191] First, when a power flow state that does not satisfy the N-1 check is detected, its input characteristics and output characteristics are constructed according to equations (3) and (4)-(5), and the input of the over-limit power flow sample is defined as X i Since the confrontation process involves iterations, the initial i is set to 0, and i increases by 1 each time iterates. Then the initial power flow input is X0.

[0192] 2) Step 2: Analyze the power flow using the power flow calculation model

[0193] The over-limit power flow sample is sent to the trained FCN-based power flow calculation model, and then according to the functional relationship between the input and output of the power flow, that is, equations (6)-(7), the output of the power flow sample is predicted to be Y outi .

[0194] 3) Step 3: Use the power flow inequality constraint to calculate the limit value and determine whether to terminate the confrontation process

[0195] The power flow inequality constraint in formula (3) is used to check whether the power flow exceeds the limit, including the branch power flow constraint and voltage amplitude constraint in the N-1 scenario, and the over-limit value vector Δy output by the power flow sample can be calculated. i The calculation process is shown in equations (8)-(9).

[0196] Δy i ={Δy[0],Δy[1],…,Δy[m]} (8)

[0197]

[0198] Where Δy[i] represents the over-limit value corresponding to the power flow output value with index i; m represents the total number of output features; Y outmax [i] represents the upper limit value of the output feature with index i; Y outmin [i] represents the lower limit value corresponding to the output feature with index i.

[0199] If the output element Δy i If Δy is positive, the output value corresponding to the position is higher than the specified upper limit; if it is negative, it is lower than the specified lower limit; if it is equal to 0, it does not exceed the limit. i If the element values ​​are not all 0, it means that the power flow does not meet the N-1 check and will continue to step 4 for adjustment. i If the values ​​of the elements in are all 0, it means that the power flow mode meets the N-1 check, and no further iteration is required, and the confrontation process is stopped. Finally, the adjustment strategy is found through formula (10).

[0200] ΔX=X e -X0 (10)

[0201] Where e represents the final number of iterations of the adversarial process; X e It is the final power flow input that satisfies N-1 verification after adjustment through the confrontation process.

[0202] 4) Step 4: Calculate input adjustment using adversarial processing techniques

[0203] According to the adversarial processing technology, the gradient of the power flow calculation model at the current input reflects how to adjust the power flow sample input so that the power flow state can be adjusted in the direction of satisfying the N-1 check. The adjustment amount of the power injected by the calculation node is shown in formula (11). The gradient is the Jacobian matrix J. The Jacobian matrix for each iteration is calculated as shown in formula (12). The Jacobian matrix can link the over-limit value of the power flow sample with the input adjustment amount, thereby calculating the input adjustment amount corresponding to the over-limit value of the power flow under the current iteration.

[0204] Δx i = J·Δy i (11)

[0205]

[0206] where a i is the output of hidden layer i, which is a function of input X; is the derivative of the output layer activation function; is the derivative of the activation function of hidden layer i with respect to hidden layer j; is the hidden layer weight of the input layer.

[0207] 5) Step 5: Update the flow input by adding input adjustment

[0208] Based on the gradient descent method, the power flow input adjustment calculated in step 4 is added to the sample input of the previous iteration as the power flow input of this iteration, as shown in formula (13). Then return to step 3 to predict the power flow output corresponding to the updated power flow input.

[0209] X i+1 =X i -Δx i+1 (13)

[0210] Through the above-mentioned power flow over-limit adjustment method proposed by the present invention, the over-limit power flow can be intelligently adjusted to meet the N-1 check, and there is no need to manually adjust the node injection power.

[0211] Embodiment 22:

[0212] See also Figures 1 to 4 , an intelligent adjustment method for power flow exceeding limit based on confrontation process, including the following contents:

[0213] (1) Introduction to the simulation process

[0214] In this embodiment, the IEEE-30 system is used for simulation. In this simulation, 10,000 power flow samples are generated by the Monte Carlo sampling method for neural network training. A fully connected neural network is used to construct the functional relationship between the input features and output features of the power flow samples, that is, FCN is used as a power flow analysis model. It consists of 3 hidden layers, each with 1000 neurons. The activation function is ReLU (Rectified Linear Unit), and the Adam optimizer in TensorFlow 2.0 is used, and the learning rate is set to 0.001. The loss function is set to the mean absolute error (MAE). When training the neural network, the epoch is set to 2000 and the batch is 64.

[0215] (2) Simulation results and analysis

[0216] The generated power flow samples are divided into a training set and a test set. The training set is used to train the neural network. Then, the trained neural network model is used to perform power flow analysis on the test set, and it is identified that power flow samples 23 and 174 do not meet the N-1 check. Their over-limit situations are as follows: Figure 3 shown.

[0217] Figure 3 The horizontal axis in represents different output features in different scenarios, and the vertical axis represents the over-limit value corresponding to the output feature. The light blue area indicates that the sample exceeds the limit in scenario N, and N-1_i represents the N-1 scenario when the i-th line is disconnected. The over-limit situation of power flow sample 23 is Figure 3 (a) shows that when the second line is disconnected, the flow sample 23 does not meet the N-1 check; the flow sample 174 exceeds the limit. Figure 3 As shown in (b), when the first line is disconnected, the power flow sample 174 does not satisfy the N check and the N-1 check.

[0218] Table 1 Over-limit power flow adjustment strategy

[0219]

[0220] The present invention generates corresponding power flow adjustment strategies for the above-mentioned over-limit power flow samples. Table 1 lists the adjustment strategies corresponding to the above-mentioned two over-limit power flows. Through the above-mentioned adjustment strategies, after adjusting the node injection power, the power flow state satisfies the N-1 check. In order to visualize the effect of each iteration of the confrontation process, the change of the over-limit indicator of each iteration is shown in Figure 4 In the display.

[0221] from Figure 4It can be observed that as the number of iterations increases, the over-limit index gradually decreases and eventually approaches 0, indicating that after the confrontation process, the over-limit power flow state is adjusted to the power flow state that satisfies the N-1 check, that is, each adjustment of the node injection power during the confrontation process can effectively alleviate the power flow over-limit problem. In addition, the use of this power flow over-limit adjustment method will be much faster than the current manual adjustment based on experience. Generally, manual adjustment requires repeated attempts, and it is time-consuming and laborious to find a suitable adjustment solution, while it only takes a few seconds to generate an adjustment strategy using the method proposed in the present invention.

[0222] In summary, the present invention proposes an intelligent adjustment method for over-limit power flow based on an adversarial process. First, the functional relationship between power flow input and output under various N-1 scenarios is constructed based on FCN. Then, a power flow over-limit adjustment strategy is constructed based on the adversarial process of the neural network. The over-limit power flow can be adjusted to a power flow state that satisfies the N-1 check by using this strategy. The effectiveness of the method proposed in the present invention is verified through example simulation analysis. Therefore, the present invention can provide technical support for power system power flow scheduling problems.

Claims

1. A method for intelligent adjustment of power flow over-limit based on a confrontation process, characterized in that: The following steps are involved: 1) Construct the power flow input and output features that represent various N-1 scenarios and obtain the training set using the Monte Carlo method; 2) Construct a power flow calculation model based on a fully connected neural network to characterize the functional relationship between power flow input and output; 3) Using the training set to train the power flow calculation model to obtain the trained power flow calculation model; 4) Obtain the power flow state to be adjusted that does not meet the N-1 check as an over-limit power flow sample; 5) Input the over-limit power flow sample into the power flow calculation model, calculate the output and over-limit value of the power flow sample, and determine whether the N-1 check is satisfied. If not, proceed to step 6); 6) Based on the gradient descent method and the iterative strategy, the input adjustment amount is calculated, and after adjusting the over-limit power flow sample, the process returns to step 5) until the over-limit power flow sample satisfies the N-1 check, and the confrontation process ends.

2. According to claim 1, a method for adjusting power flow exceeding limit based on confrontation process is characterized in that: The power flow input feature X that characterizes multiple N-1 scenarios is as follows: X={P1,P2,···,P n ,Q1,Q2,···,Q n } (1) Where P j represents the injected active power of the jth node; Q j represents the injected reactive power of the jth node; j = 1, 2, ..., n, n represents the number of nodes in the power system; The power flow output characteristics representing various N-1 scenarios are as follows: Y out ={y0,y1,y2,···,y t } (2) y i ={v i1 ,v i2 ,···,v in ,i i1 ,i i2 ,···,θ in , p fi1 ,p fi2 ,···,p fib ,q fi1 ,q fi2 ,···,q fib } (3) Where Y out To characterize the power flow output characteristics of various N-1 scenarios, y i represents the power flow output characteristics under different topological structures, y0 is the output characteristic under the original topological structure, when i>0, y i is the power flow output characteristics under different N-1 scenarios; i = 0, 1, 2..., t, t is the number of N-1 scenarios; v ij represents the voltage amplitude of the jth node in the i-th scenario; θ ij represents the phase angle of the jth node in the i-th scenario; p fia represents the active power of the a-th branch in the i-th scenario; q fia represents the branch reactive power of the a-th branch in the i-th scenario; a=1,2...,b, b represents the number of branches in the power system.

3. According to claim 1, a method for adjusting power flow exceeding limit based on confrontation process is characterized in that: The output of the power flow calculation model is as follows: Where Y out is to characterize the power flow output characteristics of various N-1 scenarios; X is the power flow sample input into the input layer; X r-1 Represents the input features entering the rth neuron layer; is the feedforward propagation function of the rth layer; r = 1, 2, ..., k, k represents the number of layers of the neural network; s is the activation function; W r , b r They represent the weight and bias of the r-th layer respectively.

4. According to claim 1, a method for intelligently adjusting power flow over-limit based on a confrontation process, characterized in that: The weights and biases of the power flow calculation model are as follows: Where W r , b r Respectively represent the updated weight and bias of the rth layer; W r ', br' respectively represent the weight and bias of the rth layer before updating; is the feedforward propagation function of the rth layer; r = 1, 2, ..., k, k represents the number of layers of the neural network; L is the loss function of the neural network; η represents the learning rate.

5. According to claim 1, a method for intelligently adjusting power flow over-limit based on a confrontation process, characterized in that: The calculation formula of the over-limit value is as follows: Δy r ={Δy[0],Δy[1],…,Δy[m]} (8) In the formula, Δy r is the over-limit value vector of the power flow sample output calculated at the rth iteration; Δy[x] represents the over-limit value corresponding to the power flow output value indexed as x; x=1,2,...,m, where m represents the total number of output features; Y out [x] represents the output feature with index x; Y outmax [x] represents the upper limit value of the output feature indexed by x; Y outmin [x] represents the lower limit of the output feature indexed by x.

6. The method for adjusting power flow exceeding limit based on confrontation process according to claim 1 is characterized in that: The power flow input adjustment amount is calculated through the confrontation process of the neural network.

7. The method for adjusting power flow exceeding limit based on confrontation process according to claim 1 is characterized in that: The method of judging that the over-limit power flow samples satisfy the N-1 check includes that the over-limit values ​​are all 0.

8. The method for adjusting power flow exceeding limit based on confrontation process according to claim 1 is characterized in that: The calculation formula of the power flow input adjustment amount is as follows: Δx r =J·Δy r (10) In the formula, Δx r is the input adjustment of the rth layer of the neural network; Δy r is the over-limit value vector output by the power flow sample of the r-th layer neural network; Among them, the Jacobian matrix J is as follows: In the formula, a r is the output of the rth hidden layer; r = 1, 2, ..., k, k represents the number of layers of the neural network; Y out is the output feature; X is the power flow sample input into the input layer; is the derivative of the output layer activation function; is the derivative of the activation function of the r1th hidden layer with respect to the r2th hidden layer; z represents the weighted input of the input layer; is the hidden layer weight of the input layer.

9. The method for adjusting power flow exceeding limit based on confrontation process according to claim 1 is characterized in that: After adjusting the over-limit power flow samples, the input of the over-limit power flow samples is as follows: X r+1 =X r -Δx r+1 (12) Where, X r , X r+1 are the power flow inputs of the rth and r+1th layers respectively; Δx r+1 is the input adjustment amount of the r+1th layer.

10. The method for intelligently adjusting power flow exceeding the limit based on the confrontation process according to claim 1 is characterized in that: The final adjustment of the over-limit power flow sample is as follows: ΔX=X e -X0 (13) Where, e represents the final number of iterations of the adversarial process; X e It is the final power flow input that satisfies the N-1 check after adjustment by the confrontation process; X0 represents the original input of the over-limit power flow sample without the adjustment process; ΔX represents the power flow adjustment amount finally found by the adjustment method.

Images