A method for constructing transient attraction domain of interconnected systems based on residual neural network

Abstract

The present invention discloses a method for constructing a transient attraction domain of a network-building, grid-connected and interconnected system based on a residual neural network, comprising: S1, establishing a network-building, grid-connected and interconnected system; S2, constructing a Lyapunov function of the network-building, grid-connected and interconnected system based on a residual neural network; S3, constructing a transient stable attraction domain of the network-building, grid-connected and interconnected system according to the Lyapunov function. The present application can establish the Lyapunov function of a heterogeneous system and thereby construct the transient stable attraction domain of the system.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method for constructing a transient attraction domain of a grid-connected interconnected system based on a residual neural network. Background Art

[0002] As the proportion of wind, solar and other renewable energy power generation continues to increase, the power system is gradually dominated by power electronic equipment, showing characteristics such as low inertia and low voltage support capacity, and the safe and stable operation of the power system faces challenges. Under transient faults in the power system, the stable operation area of ​​the system needs further analysis.

[0003] At present, transient stability assessment (TSA) of power systems mainly relies on the following methods: time domain simulation, equal area method, Lyapunov function-based method and artificial intelligence-based method. Each method has specific application scenarios and limitations. Time domain simulation can accurately calculate the transient operation trajectory of the system, but this method has high computational cost, long time consumption and is highly dependent on model accuracy. The Lyapunov function-based method constructs a non-negative scalar function that takes the minimum value when there is no disturbance and increases when the system state deviates from the equilibrium point; if the derivative of the function remains negative in the defined area, it can be proved that the system is Lyapunov stable. The Lyapunov function-based method provides a strong theoretical foundation and quantitative stability domain characterization method, which is suitable for complex system analysis. However, constructing the Lyapunov function of a complex power system is often very challenging and requires in-depth theoretical knowledge and complex mathematical derivation. The artificial intelligence-based method can characterize the dynamics of complex power system models by learning a large amount of measurement data, but they face generalization and interpretability problems. The equal-area method is intuitive and easy to implement, but it is mainly applicable to simple systems, may ignore interactions in complex systems, and has limited accuracy.

[0004] With the continuous development of AI technology, its application in the power industry has become more and more common. With its powerful nonlinear fitting ability and data processing ability, neural networks can construct a Lyapunov function that meets the conditions for a stable system in a short time, but there are problems such as slow function training speed and small estimation of the attraction domain, and the transient synchronization stability problem of the VSG and PLL parallel power supply system is not involved. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the method for constructing a transient attraction domain of a network-connected interconnected system based on a residual neural network provided by the present invention solves the problems of slow training process, unstable training and small stable domain estimation area existing in the prior art.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for constructing a transient attraction domain of a network-connected interconnected system based on a residual neural network, comprising:

[0007] S1. Establish a network connection and interconnection system;

[0008] S2. Based on the residual neural network, the Lyapunov function of the grid-connected interconnection system is constructed;

[0009] S3. Construct the transient stability attraction domain of the grid-connected interconnected system based on the Lyapunov function.

[0010] Further: the grid-connected interconnection system in S1 includes a common grid connection point PCC, and the common grid connection point PCC is respectively connected to one end of the equivalent line inductance L2, one end of the equivalent line resistance R1, one end of the equivalent line resistance R3 and the first input end of the current-controlled converter;

[0011] The other end of the equivalent line inductor L2 is connected to one end of the equivalent line resistor R2, and the other end of the equivalent line resistor R2 is connected to the power grid;

[0012] The other end of the equivalent line resistor R1 is connected to one end of the equivalent line inductor L1, and the other end of the equivalent line inductor L1 is respectively connected to the grounding capacitor C and one end of the equivalent line inductor L4; the other end of the equivalent line inductor L4 is respectively connected to the first input end of the virtual synchronous generator and the voltage stabilizing circuit; the output end of the virtual synchronous generator is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage of magnitude U dc The second input terminal of the virtual synchronous generator receives the voltage U on the capacitor C. abc ;

[0013] The other end of the equivalent line resistor R3 is connected to one end of the equivalent line inductor L3, the other end of the equivalent line inductor L3 is connected to the second input end of the current-controlled converter, the output end of the current-controlled converter is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage of magnitude U dc The third input terminal of the virtual synchronous generator receives a DC voltage of U dc of DC voltage.

[0014] Further: the expression of the active ring rotor motion equation of the virtual synchronous generator is:

[0015]

[0016]

[0017] in, i 1 is the power angle of the virtual synchronous generator, and Represent the virtual inertia and active loop damping coefficient respectively, is the electromagnetic torque, is the mechanical torque, is the reference active power, is the output power of the current-controlled converter, D p is the active loop damping coefficient of the virtual synchronous generator, D is the equivalent damping coefficient, is the grid reference angular frequency;

[0018] The dynamic equation of the current-controlled converter is:

[0019]

[0020] in, U pccq The voltage at the grid point abc—dq The value obtained after transformation is i 2 is the output power angle of the current-controlled converter, K 1. K 2 are the proportional coefficient and the integral coefficient respectively.

[0021] Further: The expression of the active loop equation of the virtual synchronous generator after incorporating the current-controlled converter is:

[0022]

[0023] in, , , , Y 1 is the equivalent admittance between the current-controlled converter and the virtual synchronous generator, Y 2 is the equivalent admittance between the current-controlled converter and the grid, i a , i b , i c are the admittance angles of the admittance, oh 1 is the angular frequency of the virtual synchronous generator, oh 2 is the angular frequency of the current-controlled converter, E g is the effective value of the virtual synchronous machine phase voltage, V g is the effective value of the equivalent voltage of the power grid, I pq is the effective value of the output current of the current-controlled converter, i pq yes Ipq The phase angle.

[0024] Further: S2 includes:

[0025] S21. Describe the nonlinear system based on the dynamic equations of the grid-connected interconnected system;

[0026] S22, find the equilibrium point that makes the value of the nonlinear system 0, and move the equilibrium point to the zero point;

[0027] S23. Based on the moved equilibrium point and residual neural network, training is performed to construct the Lyapunov function of the grid-connected interconnected system.

[0028] Further: In said S23, the Lyapunov function of the grid-connected interconnected system is constructed V ( x ) is:

[0029]

[0030] Among them, tanh(.) represents the activation function, W 1, W 2 are weight coefficients, b 1, b 2 are bias coefficients, W 3 is a constant, representing the weight of the direct connection layer, r 1， r 2 is the activation function, Represents the input of the residual neural network; where the input include:

[0031]

[0032] in, oh is the grid angular velocity, i 1 s , i 2 s They are the steady-state power angles of the virtual synchronous generator and the current-controlled converter, respectively.

[0033] Further: Lyapunov function of grid-connected interconnected systems V ( x ) is constrained as follows:

[0034]

[0035] in, represents the Lyapunov function V (.) The derivative with respect to time, D For inspection area.

[0036] Further: S3 includes:

[0037] S31, model the grid-connected interconnection system and construct the system dynamics equations;

[0038] S32, randomly generate several groups of sample points as training sets, and substitute them into the residual neural network to obtain the value of the corresponding Lyapunov function;

[0039] S33. Calculate the risk function according to the value of the Lyapunov function;

[0040] S34, optimizing the risk function by gradient descent method, and calculating the output value of the residual neural network;

[0041] S35. Determine whether the output value of the residual neural network satisfies the constraint function:

[0042] If yes, proceed to step S36;

[0043] If not, calculate the counterexample samples that do not meet the defined interval generation, add the counterexample samples to the training set, and return to S32;

[0044] S36, constructing a transient attraction domain of the grid-connected interconnected system according to the Lyapunov function corresponding to the output value of the residual neural network that satisfies the constraint function.

[0045] Further: In S33, the risk function The expression is:

[0046]

[0047] in, i Indicates i training sample points, N is the number of training samples, x si are randomly generated sample points, l 1. l 2 and l 3 are all weight coefficients greater than zero and can be adjusted.

[0048] Further, in S35, the expression of the constraint function is:

[0049]

[0050] in, , , , , , x in is the input of the neural network, w 1. w2 and w 3 are the weight coefficients of the first layer, second layer and direct connection layer of the residual neural network, b and b 2 are the bias coefficients of the first and second layers of the residual neural network, d 1 and d 2 are the outputs of the first and second layers of the residual neural network, d 3 Yes x in After the output of the direct connection layer, d is the final output of the residual neural network.

[0051] The beneficial effects of the present invention are:

[0052] 1. The bypass structure based on residual neural network can improve the training optimization process of the model, enhance the learning speed and stability of the model, and construct a transient attraction domain;

[0053] 2. Compared with traditional feedforward neural networks, the use of residual neural networks can effectively solve the problems of gradient disappearance and gradient explosion, and can reduce the risk of model overfitting;

[0054] 3. Be able to establish the Lyapunov function of heterogeneous systems and use it to construct the transient stable attraction domain of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the transient attraction domain construction method of the grid-connected interconnected system based on residual neural network;

[0056] Figure 2 To construct the topology of the grid-connected interconnection system;

[0057] Figure 3 It is the control block diagram of the active power control loop of the virtual synchronous generator;

[0058] Figure 4 This is the control block diagram of the current controlled converter.

[0059] Figure 5 This is an example diagram of a residual neural network. DETAILED DESCRIPTION

[0060] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0061] like Figure 1As shown, in one embodiment of the present invention, a method for constructing a transient attraction domain of a network-connected interconnected system based on a residual neural network is provided, comprising:

[0062] S1. Establish a network connection and interconnection system;

[0063] S2. Based on the residual neural network, the Lyapunov function of the grid-connected interconnection system is constructed;

[0064] S3. Construct the transient stability attraction domain of the grid-connected interconnected system based on the Lyapunov function.

[0065] like Figure 2 As shown, in this embodiment, the grid-connected interconnection system in S1 includes a common grid connection point PCC, and the common grid connection point PCC is respectively connected to one end of the equivalent line inductance L2, one end of the equivalent line resistance R1, one end of the equivalent line resistance R3 and the first input end of the current-controlled converter;

[0066] The other end of the equivalent line inductor L2 is connected to one end of the equivalent line resistor R2, and the other end of the equivalent line resistor R2 is connected to the power grid;

[0067] The other end of the equivalent line resistor R1 is connected to one end of the equivalent line inductor L1, and the other end of the equivalent line inductor L1 is respectively connected to the grounding capacitor C and one end of the equivalent line inductor L4; the other end of the equivalent line inductor L4 is respectively connected to the first input end of the virtual synchronous generator and the voltage stabilizing circuit; the output end of the virtual synchronous generator is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage of magnitude U dc The second input terminal of the virtual synchronous generator receives the voltage U on the capacitor C. abc ;

[0068] The other end of the equivalent line resistor R3 is connected to one end of the equivalent line inductor L3, the other end of the equivalent line inductor L3 is connected to the second input end of the current-controlled converter, the output end of the current-controlled converter is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage of magnitude U dc The third input terminal of the virtual synchronous generator receives a DC voltage of U dc of DC voltage.

[0069] like Figure 3 As shown, it is the control block diagram of the active power control loop of VSG (virtual synchronous generator):

[0070] in, i 1 is the power angle of VSG, oh 1 is the angular frequency of the VSG. i 1 is defined by the following formula:

[0071] (1)

[0072] in, oh 0 is the grid reference angular frequency, oh 1 is the angular frequency of the virtual synchronous generator, and the active ring rotor motion equation is shown in formula (2):

[0073] (2)

[0074] in, and represent the virtual inertia and active loop damping coefficient respectively, is the electromagnetic torque, is the mechanical torque, which can be obtained by the following formula:

[0075] (3)

[0076] Mechanical power It can be obtained by the following formula:

[0077] (4)

[0078] is the reference active power. Combining equations (2), (3) and (4), we can get:

[0079] (5)

[0080] in, , D It can be obtained by the following formula:

[0081] (6)

[0082] like Figure 4 The control block diagram of PLL (current controlled converter) is shown in the figure. i 2 is the output power angle of PLL, oh 2 is the angular frequency of the PLL, where the power angle is defined by:

[0083] (7)

[0084] The dynamic equation of the PLL is:

[0085] (8)

[0086] in K 1. K 2 are the proportional coefficient and the integral coefficient respectively. Taking the second derivative of both sides with respect to time, we get:

[0087] (9)

[0088] Grid connection voltage pcc It can be obtained from formula (10):

[0089] (10)

[0090] in, V· g is the grid voltage, Y 2 is the equivalent admittance between PLL and the power grid, g is the VSG voltage, Y 1 is the equivalent admittance between PLL and VSG, pq is the effective value of the current of the PLL equivalent current source, and we can get:

[0091] (11)

[0092] make , , , i a , i b , i c are the admittance angles of the admittance, abc-dq Transformation yields:

[0093] (12)

[0094] in, i pq yes I· pq Substituting (12) into (9) yields

[0095] (13)

[0096] The output power of the current-controlled converter is now P e Output current It can be expressed by the following formula:

[0097] (14)

[0098] Substituting (10) into (14), we obtain:

[0099] (15)

[0100] but P e It can be obtained by the following formula:

[0101] (16)

[0102] Among them, Re(.) means taking the real part of the operation result, I· e * express I· e The conjugate of

[0103] (17)

[0104] Substituting (17) into (5) we can get the active loop equation of VSG after incorporating PLL:

[0105] (18)

[0106] in, E g is the effective value of the virtual synchronous machine phase voltage, V g is the effective value of the grid equivalent voltage, I pq is the effective value of the output current of the current-controlled converter, i pq yes I pq The phase angle of

[0107] After sorting out, the expression of the active loop equation of the virtual synchronous generator after incorporating the current-controlled converter is obtained as follows:

[0108] (19)

[0109] In this embodiment, S2 includes:

[0110] S21. Describe the nonlinear system based on the dynamic equations of the grid-connected interconnected system;

[0111] S22, find the equilibrium point that makes the value of the nonlinear system 0, and move the equilibrium point to the zero point;

[0112] The balance point moves to zero by translating the horizontal coordinate of Lyapunov (V(x)). In order to facilitate the construction of Lyapunov function, the input of residual neural network is the physical quantity after translation, that is, the amount after subtracting the balance point from each physical quantity is used as input;

[0113] S23. Based on the moved equilibrium point and residual neural network, training is performed to construct the Lyapunov function of the grid-connected interconnected system.

[0114] In S21, the nonlinear system can be expressed as f (.) description, where Determined by the system's dynamic equations, assumingx e is an equilibrium point of the system, that is f(x e ) =0, x ( t 0) is the initial state of the system after the disturbance. , so that when , the state of the system will converge to the equilibrium point, that is, , then the equilibrium point is asymptotically stable.

[0115] Assume that the equilibrium point of the system x e =0, that is f (0) = 0. According to Lyapunov's second method, if we can find a scalar function V ( x ), and its derivative with respect to time If equation (19) is satisfied, the system is asymptotically stable.

[0116] Among them, the constraints of the scalar function V(.) are:

[0117]

[0118] in, represents the Lyapunov function V (.) The derivative with respect to time, D For inspection area.

[0119] In order to construct the energy function, the equilibrium point is moved to zero as the input of the residual neural network;

[0120] (twenty one)

[0121] in, oh is the grid angular velocity, i 1 s , i 2 s They are the steady-state power angles of VSG and PLL, and after normalization, they are obtained

[0122] (twenty two)

[0123] The two-layer residual neural network structure used in this embodiment is composed of Figure 5 As shown, the output of this structure can be expressed by the formula:

[0124] (twenty three)

[0125] in, W 1, W2 is the weight coefficient, b 1, b 2 is the bias coefficient, W 3 is a constant, representing the weight of the direct connection layer, r 1， r 2 is the activation function. Using residual neural network can effectively solve the problem of gradient disappearance and gradient explosion, and can reduce the risk of model overfitting, making the network easier to train.

[0126] The Lyapunov function can be obtained by the output of the residual neural network. In this embodiment, tanh is selected as the activation function, which is expressed by formula (24):

[0127] (twenty four)

[0128] As a preferred embodiment of this invention, in order to effectively train the Lyapunov function that satisfies the definition of formula (20), it is necessary to construct a risk function to measure the degree to which the Lyapunov function does not satisfy the definition through risk. The risk function is defined as follows:

[0129] (25)

[0130] Where N is the number of training samples, x si are randomly generated sample points, l 1, l 2, l 3 is a weight coefficient greater than zero and adjustable; if the output of the residual neural network structure does not satisfy the defined Lyapunov function, the value of the risk function will be relatively large.

[0131] As a preferred embodiment of this embodiment, in order to obtain a Lyapunov function with a larger stability domain as much as possible during the training process, a loss term can be added to the risk function, as shown in the following formula:

[0132] (26)

[0133] in, l 4 is an adjustable weight coefficient greater than zero. l 1, l 2, l 3, l The weight of 4 can control the gradient descent to focus on optimizing a certain loss function term. The definition of loss function (25) is the desired function V(x) The value is positive, the derivative of the function L_V(x) The value of is negative, V( 0 ) The value of is 0, the state variable x The modulus value is close to the corresponding V (x )value, w,b is the parameter of the neural network. The risk function value can be reduced in the iterative process through the gradient descent method, accompanied by w,b The parameters are updated and finally the Lyapunov function that satisfies the definition is obtained through training.

[0134] For the calculation of function derivatives, we know ,in It can be found by the chain rule. It can be obtained by the following formula:

[0135] (27)

[0136] in d,y 1 ,d 1 ,d 2 ,d 3 is:

[0137] (28)

[0138] (29)

[0139] (30)

[0140] (31)

[0141] (32)

[0142] Substituting equations (28)-(32) into equation (27), we can obtain:

[0143] (33)

[0144] in, x in is the input of the neural network, w 1. w 2 and w 3 are the weight coefficients of the first layer, second layer and direct connection layer of the residual neural network, b and b 2 are the bias coefficients of the first and second layers of the residual neural network, d 1 and d 2 are the outputs of the first and second layers of the residual neural network, d 3 Yes x in After the output of the direct connection layer, d is the final output of the residual neural network;

[0145] The function output by the above residual neural network structure only satisfies the constraints of positive definite function and negative definite derivative in the training sample. To determine whether the above conditions are met in the continuous region, it is necessary to solve the inequality in the continuous region. The logical inequality is defined as shown in equation (34). If the logical inequality has a solution, the Lyapunov function output by the model is not a qualified Lie function.

[0146] (34)

[0147] in d、r is much smaller than R Hyperparameters are used to avoid instability problems in numerical solutions. is the area we want to focus on. This paper uses the STM solver to solve the nonlinear constraint problem described in equation (34). If the Lyapunov function is unqualified, the solver will return the state variable interval that does not meet the definition of the Lyapunov function, which is expressed by equation (34):

[0148] (35)

[0149] Formula (35) shows that the function of the model output is , The two-dimensional rectangular area formed by the interval does not satisfy the definition of formula (34). In order to speed up the training process of the residual neural network model, after each falsification, a certain number of sample points can be randomly sampled on the above counterexample interval and added to the training set.

[0150] In this embodiment, S3 includes:

[0151] S301. Estimate the maximum stability region based on the Lyapunov function;

[0152] (36)

[0153] in yes D The boundary of can be expressed as follows

[0154] (37)

[0155] S302. Solve the maximum stability domain estimate by the Granger multiplier method to obtain the energy plane corresponding to the Lyapunov function R d ;

[0156] (38)

[0157] Construct the function using the Lagrange multiplier method:

[0158] (39)

[0159] in m is a Lagrangian operator, yes x and m Calculate the partial derivatives respectively and set them to zero, and obtain the constraints of the equation system as shown in formula (40):

[0160] (40)

[0161] S303, obtain a series of satisfying conditions in S31 and S32 x , denoted as ,according to x The new value of is obtained to obtain the transient stability safety domain of the grid-connected interconnected system;

[0162] (41)

[0163] Specifically, in this embodiment, S3 includes:

[0164] S31, model the grid-connected interconnection system and construct the system dynamics equations;

[0165] S32, randomly generate several groups of sample points as training sets, and substitute them into the residual neural network to obtain the value of the corresponding Lyapunov function;

[0166] S33. Calculate the risk function according to the value of the Lyapunov function;

[0167] S34, optimizing the risk function by gradient descent method, and calculating the output value of the residual neural network;

[0168] S35. Determine whether the output value of the residual neural network satisfies the constraint function:

[0169] If yes, proceed to step S36;

[0170] If not, calculate the counterexample samples that do not meet the defined interval generation, add the counterexample samples to the training set, and return to S32;

[0171] S36, constructing a transient attraction domain of the grid-connected interconnected system according to the Lyapunov function corresponding to the output value of the residual neural network that satisfies the constraint function.

[0172] In S43, the risk function The expression of is shown in formula (25). In S45, the constraint function is shown in formula (28)-formula (33).

[0173] The beneficial effects of the present invention are: 1. It can establish the Lyapunov function of heterogeneous systems and use it to construct the transient stable attraction domain of the system; 2. Compared with the traditional feedforward neural network, the use of residual neural network can effectively solve the gradient disappearance and gradient explosion problems, and can reduce the risk of model overfitting, making the network easier to train.

[0174] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for constructing transient attraction domain of grid-connected interconnected systems based on residual neural networks, characterized in that: include: S1. Establish a network connection and interconnection system; The grid-connected interconnection system includes a public grid connection point PCC, which is respectively connected to one end of an equivalent line inductance L2, one end of an equivalent line resistance R1, one end of an equivalent line resistance R3 and a first input end of a current-controlled converter; the other end of the equivalent line inductance L2 is connected to one end of the equivalent line resistance R2, and the other end of the equivalent line resistance R2 is connected to a power grid; the other end of the equivalent line resistance R1 is connected to one end of the equivalent line inductance L1, and the other end of the equivalent line inductance L1 is respectively connected to a grounding capacitor C and one end of an equivalent line inductance L4; the other end of the equivalent line inductance L4 is respectively connected to a first input end of a virtual synchronous generator and a voltage stabilizing circuit; the output end of the virtual synchronous generator is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage with a magnitude of U dc The second input terminal of the virtual synchronous generator receives the voltage U on the capacitor C. abc The other end of the equivalent line resistor R3 is connected to one end of the equivalent line inductor L3, the other end of the equivalent line inductor L3 is connected to the second input end of the current-controlled converter, the output end of the current-controlled converter is connected to the voltage stabilizing circuit, and the voltage stabilizing circuit receives a voltage of U dc The third input terminal of the virtual synchronous generator receives a DC voltage of U dc DC voltage; The expression of the active ring rotor motion equation of the virtual synchronous generator is: in, θ 1 is the power angle of the virtual synchronous generator, and represent the virtual inertia and active loop damping coefficient respectively, is the electromagnetic torque, is the mechanical torque, is the reference active power, is the output power of the current-controlled converter, D p is the active loop damping coefficient of the virtual synchronous generator, D is the equivalent damping coefficient, is the grid reference angular frequency; The dynamic equation of the current controlled converter is: in, U pccq The voltage at the grid point abc—dq The value obtained after transformation is θ 2 is the output power angle of the current-controlled converter, K 1. K 2 are the proportional coefficient and the integral coefficient respectively; The expression of the active loop equation of the virtual synchronous generator after incorporating the current-controlled converter is: in, , , , Y 1 is the equivalent admittance between the current-controlled converter and the virtual synchronous generator, Y 2 is the equivalent admittance between the current-controlled converter and the grid, θ a , θ b , θ c are the admittance angles of the admittance, ω 1 is the angular frequency of the virtual synchronous generator, ω 2 is the angular frequency of the current-controlled converter, E g is the effective value of the virtual synchronous machine phase voltage, V g is the effective value of the grid equivalent voltage, I pq is the effective value of the output current of the current-controlled converter, θ pq yes I pq The phase angle of S2. Based on the residual neural network, a Lyapunov function of the grid-connected interconnection system is constructed, which includes: S21. Describe the nonlinear system based on the dynamic equations of the grid-connected interconnected system; S22, find the equilibrium point that makes the value of the nonlinear system 0, and move the equilibrium point to the zero point; S23, based on the moved equilibrium point and residual neural network training, construct the Lyapunov function of the grid-connected interconnection system; Lyapunov function V ( x ) is: Among them, tanh(.) represents the activation function, W 1, W 2 are weight coefficients, b 1, b 2 are bias coefficients, W 3 is a constant, representing the weight of the direct connection layer, ρ 1， ρ 2 is the activation function, Represents the input of the residual neural network; where the input include: in, ωg is the grid angular velocity, θ 1 s , θ 2 s They are the steady-state power angles of the virtual synchronous generator and the current-controlled converter, respectively; S3. Construct the transient stability attraction domain of the grid-connected interconnected system based on the Lyapunov function.

2. The method for constructing transient attraction domain of a network-connected interconnected system based on residual neural network according to claim 1 is characterized in that: Lyapunov function of grid-connected interconnected system V ( x ) is constrained as follows: in, represents the Lyapunov function V (.) The derivative with respect to time, D For inspection area.

3. The method for constructing transient attraction domain of a network-connected interconnected system based on residual neural network according to claim 2 is characterized in that: The S3 includes: S31, model the grid-connected interconnection system and construct the system dynamics equations; S32, randomly generate several groups of sample points as training sets, and substitute them into the residual neural network to obtain the value of the corresponding Lyapunov function; S33. Calculate the risk function according to the value of the Lyapunov function; S34, optimizing the risk function by gradient descent method, and calculating the output value of the residual neural network; S35. Determine whether the output value of the residual neural network satisfies the constraint function: If yes, proceed to step S36; If not, calculate the counterexample samples that do not meet the defined interval generation, add the counterexample samples to the training set, and return to S32; S36, constructing a transient attraction domain of the grid-connected interconnected system according to the Lyapunov function corresponding to the output value of the residual neural network that satisfies the constraint function.

4. The method for constructing transient attraction domain of a network-connected interconnected system based on residual neural network according to claim 3 is characterized in that: In S33, the risk function The expression is: in, i Indicates i training sample points, N is the number of training samples, x si are randomly generated sample points, l 1. l 2 and l 3 are all weight coefficients greater than zero and can be adjusted.

5. The method for constructing transient attraction domain of a network-connected interconnected system based on residual neural network according to claim 4 is characterized in that: In S35, the expression of the constraint function is: in, , , , , , x in is the input of the neural network, w 1. w 2 and w 3 are the weight coefficients of the first layer, second layer and direct connection layer of the residual neural network, b and b 2 are the bias coefficients of the first and second layers of the residual neural network, d 1 and d 2 are the outputs of the first and second layers of the residual neural network, d 3 Yes x in After the output of the direct connection layer, d is the final output of the residual neural network.

Images