A fan control parameter optimization acceleration method of a least square generative adversarial network

Abstract

The application provides a fan control parameter optimization acceleration method of a least square generative adversarial network, which is based on the least square generative adversarial network to optimize the parameters of a doubly-fed induction fan controller; a coordinate rotation optimization method is used in the iteration process; the generative adversarial network used in the image recognition field is used in the fan parameter optimization field; the generative adversarial network is used to reduce the iteration times to cope with the situation that the calculation time of the parameter optimization method is too long; and the least square loss function is introduced to optimize the generative adversarial network. The fan control parameter optimization acceleration method of the least square generative adversarial network can solve the optimization problem of the parameters of the doubly-fed induction fan controller, realize the function of quickly and accurately obtaining the optimal parameters, reduce the required optimization time, improve the operation speed of the optimization method, and improve the operation efficiency of the doubly-fed induction fan.

Description

Technical Field

[0001] This invention belongs to the fields of power system controller parameter optimization, wind turbine control parameter optimization, artificial intelligence, machine learning, image processing, generative adversarial networks and automatic control. It involves coordinate rotation optimization methods, deep learning and generative adversarial network methods, and is applicable to the optimization and acceleration of controller parameters of doubly fed induction wind turbines in power systems. Background Technology

[0002] Existing control methods based on proportional-integral-derivative controllers have the disadvantage of difficulty in precisely adjusting control parameters.

[0003] In addition, wind turbine parameter optimization requires a large amount of calculation, and the calculation time of existing parameter optimization methods is too long.

[0004] Furthermore, the training process of the original generative adversarial network is unstable and prone to problems such as pattern collapse and training non-convergence.

[0005] Therefore, a coordinate rotation optimization method is introduced to automatically adjust parameters to solve the problem of difficulty in accurately adjusting control parameters; a generative adversarial network acceleration method is proposed to solve the problem of excessive computation time of parameter optimization methods; and a least squares loss function is introduced to solve the problems of pattern collapse and training non-convergence in generative adversarial networks. Summary of the Invention

[0006] A least-squares generative adversarial network (GAN) method for accelerating wind turbine control parameter optimization combines a least-squares GAN model with a coordinate rotation optimization method for parameter optimization of doubly-fed induction generator (DFIG) wind turbine controllers. This improves the operating efficiency of DFIG wind turbines, reduces the required optimization time, and increases the speed of the optimization method. The steps in its application are as follows:

[0007] Step (1): Initialize the doubly fed induction fan control model;

[0008] The system equations for a wind turbine are:

[0009]

[0010] Among them, T m,DFIG T represents the mechanical torque of the wind turbine. e,DFIG ω represents the electromagnetic torque of the wind turbine, J represents the system parameter of the wind turbine generator, and ω represents the electromagnetic torque of the wind turbine generator. r Let dt be the rotor speed, and dt be the differential of the variable with respect to time t.

[0011] Wind turbine mechanical power P m and electromagnetic power P e The relationship between them is:

[0012]

[0013] Where, ω m It is the mechanical angular velocity;

[0014] Fan mechanical power P m for:

[0015]

[0016] Where R is the radius of the wind turbine rotor, ρ is the air density, and v wind Where π is the wind speed and π is the mathematical constant pi.

[0017] Power coefficient C p for:

[0018]

[0019] Where, β DFIG pitch angle

[0020] The tip speed ratio λ is:

[0021]

[0022] Where, ω m,DFIG Wind turbine blade speed

[0023] intermediate variable λ i for:

[0024]

[0025] The model of the doubly-fed induction fan is as follows:

[0026]

[0027] Among them, e' qs and e' ds These represent the equivalent internal q-axis and d-axis voltages, respectively. ds and v dr These represent the direct-axis voltage of the stator winding and the direct-axis voltage of the rotor winding, respectively. qs and v qr These represent the cross-axis voltage of the stator winding and the cross-axis voltage of the rotor winding, respectively. b It is the reference electric angular velocity, ω s It is the synchronous angular velocity, ω r It is the rotor angular velocity, L ss L is the stator winding inductance. rr L is the rotor winding inductance. m For the mutual inductance between the stator winding and the rotor winding, i ds and i qs R represents the direct-axis and quadrature-axis currents of the stator windings, respectively. s R represents the stator resistance.r Let R1 represent the rotor resistance, ∫dt be the integral of the variable with respect to time, and the intermediate derivation variable R1 for the wind turbine model be:

[0028] R1 = R s +R2 (8)

[0029] The intermediate derivation variable R2 is:

[0030]

[0031] intermediate derivation variable T r for:

[0032]

[0033] Intermediate derivation variable L s for:

[0034]

[0035] To achieve optimal performance, four proportional-integral loops are employed; in the control loop, i qr and i dr The controller outputs v, which is the input reference value. qr1 and v dr1 , and the compensation voltage v qr2 and v dr2 Combined to form the output value of the rotor voltage; compensation voltage v qr2 and v dr2 for:

[0036]

[0037] The objective function for accelerating the optimization of wind turbine controller parameters is:

[0038]

[0039] Where, k Pi and k Ii These are the proportional and integral parameters of the proportional-integral controller, k. Pimin and k Iimin Let k be the minimum value of the proportional parameter and the integral parameter, respectively. Pimax and k Iimax v represents the maximum values ​​of the proportional parameter and the integral parameter, respectively. windmax and v windmin V represents the maximum and minimum wind speeds, respectively. s V is the system voltage. smax and V smin Q represents the maximum and minimum values ​​of the system voltage, respectively. s Q represents reactive power. smax and Qsmin These represent the maximum and minimum reactive power values, respectively, and T1 is the sampling time for one operation of the wind power system. To integrate the objective function over a single sampling time, w1 is the weighted value of the rotor angular velocity error, w2 is the weighted value of the reactive power deviation, Σ is the summation sign, and |·| represents the absolute value of the error. In a doubly-fed induction fan system, frequency stability is affected by the rotor angular velocity error. Voltage stability is affected by reactive power deviation. The goal of accelerating parameter optimization for doubly-fed induction fan controllers is to reduce rotor speed error. Minimize and make reactive power error Minimum, that is, minimum fitness value of objective function f(x), ω r and These are the precise and reference values ​​of the rotor speed, respectively. s and These are the precise and reference values ​​for reactive power, respectively.

[0040] Step (2): Initialize the parameters of the doubly fed induction fan and randomly generate the controller initialization position X0;

[0041] Step (3): The least squares generative adversarial network wind turbine control parameter optimization acceleration method determines whether the current iteration algebra k satisfies k < N1, where N1 is the initial iteration number. If the condition is met, the coordinate rotation optimization method is used for iteration. If the condition is not met, the least squares generative adversarial network model training is entered and the acceleration process is executed.

[0042] Step (4): The coordinate rotation optimization method begins to iterate, calculates the fitness function value of each initial position X0, and searches for the global optimal solution;

[0043] In each iteration, the selected optimal solution guides the variables to explore the region with the minimum fitness function value, finding a solution closer to the global optimum. The coordinate rotation optimization method transforms a multidimensional problem into a series of one-dimensional problems, alternately converting a multivariate optimization problem into a single-variable optimization problem. The search process involves alternating searches along the coordinate directions, allowing only one variable to change at a time while keeping the others constant. When the scaling parameter k... Pi When it changes, the integral parameter k Ii When the integration parameter k remains unchanged Ii When it changes, the proportional parameter k Pi Remain unchanged;

[0044] The update strategy for the coordinate rotation optimization method is as follows:

[0045] Step (4.1) Given the objective function f(x), randomly change the initialization position X0 as the starting position of the first iteration;

[0046] Step (4.2) sets the search direction to n coordinate directions, and searches along one coordinate direction at a time. The search direction is represented as follows:

[0047]

[0048] Among them, e1, e2, ..., e n For the n coordinate directions set, [·] T It is the transpose symbol;

[0049] The formula for iterative calculation using the optimal step size in step (4.3) is as follows:

[0050]

[0051] Where k is the iteration number, k = 1, 2, ..., N1, N1 is the initial iteration number, and i is the index of the current one-dimensional search loop. This represents the position of the i-th search in the k-th iteration. This represents the position of the search in the (i-1)th iteration of the k-th loop. e is the step size of the i-th search in the k-th iteration. i This represents the search direction for the i-th search.

[0052] Step (4.4) uses a one-dimensional search to find the optimal step size.

[0053]

[0054] Where f(·) is the iterative formula in step (4.4), and the function Used to find the minimum value of the function f(·), and the corresponding minimum point. To find the optimal step size;

[0055] If i < n in step (4.5), i = i + 1, return to step (4.3) and perform iterative calculation again; if i = n, proceed to the next step.

[0056] In step (4.6), if k < N1, k = k + 1, i = 0, return to step (4.3) for the next iteration; if k = N1, end the iteration.

[0057] The position information is updated by using a coordinate rotation optimization method. The position X1 and the corresponding fitness function value f(X1) are recorded during the iteration process and used as the input value for the next step of the least squares generative adversarial network.

[0058] Step (5): If the algebra of the current iteration does not satisfy k < N1 or k > N2, where N2 is the iteration number of the least squares generative adversarial network, then enter the least squares generative adversarial network, learn the data output in step (4), and start the acceleration process.

[0059] Step (6): Train the generative adversarial network using the input positions and corresponding fitness function values ​​obtained from the iterations in steps (4) and (5); the learning process of the generative adversarial network involves simultaneously training the generator and the discriminator, and the minimax objective function of the generative adversarial network. for:

[0060]

[0061] Where G is the generator, whose goal is to learn the input data and generate the output data; D is the discriminator, which determines whether the received data comes from real data or the generator G; x is the real data, z is the input of the generator, G(z) is the fake data generated by the generator; D(x) is the discriminator's probability assessment of the real data, and D(G(z)) is the discriminator's probability assessment of the fake data generated by the generator. The expected value of the true data distribution. The expected value of the fake data generated by the generator; logD(x) and log(1-D(G(z))) are the loss functions of the network, and log(·) is the logarithmic function to the base 10;

[0062] Standard generative adversarial networks (GANs) suffer from two drawbacks: low-quality generated data and unstable training processes. This paper improves the method by replacing the cross-entropy loss in the standard GAN's objective function with least squares loss, resulting in the objective function of a least squares GAN. and for:

[0063]

[0064] Where a and b are the labels of the fake data and the real data, respectively, and c represents the fake data value that the generator wants the discriminator to believe;

[0065] Step (7): The least squares generative adversarial network completes training through N2 iterations. The position X1 and the corresponding fitness function value f(X1) are input to the trained network to obtain the predicted optimal position X2 and the optimal fitness function value f(X2).

[0066] Step (8): Since the prediction accuracy of the trained least squares generative adversarial network is not 100%, in order to improve the accuracy of parameter optimization, the coordinate rotation method is used for optimization in the last N3 iterations. Therefore, the coordinate rotation optimization method only performs N1+N3 iterations, which will greatly reduce the running time of the method and achieve the purpose of accelerating the method using the least squares generative adversarial network. Compared with the method without acceleration, the number of iterations of the coordinate rotation optimization method is reduced during the iteration process.

[0067]

[0068] Step (9): If the algebra of the current iteration satisfies k>N1+N2, calculate the coordinate rotation optimization method, and the calculation method is the same as in step (4); to make the results more accurate, update the method termination judgment criteria. Based on the judgment of the number of iterations, set the termination precision ε. If the number of iterations k > N1 + N2 + N3, stop the iteration and output the optimal position X3;

[0069] Step (10): The optimal position output by the coordinate rotation optimization method and the least squares generative adversarial network represents the controller parameters of the doubly fed induction fan. The obtained parameters are then input into the doubly fed induction fan to obtain higher fan control performance. Attached Figure Description

[0070] Figure 1 This is a complete flowchart of the method of the present invention. Detailed Implementation

[0071] This invention proposes an accelerated method for optimizing wind turbine control parameters using least-squares generative adversarial networks, which is described in detail below with reference to the accompanying drawings:

[0072] Figure 1 This is a complete flowchart of the control process of the method of this invention. First, the parameters of the doubly fed induction fan are initialized, and the initial position of the controller is randomly generated. Then, it is determined whether the current iteration satisfies the condition k≤N1. If the condition is met, the coordinate rotation optimization method is used for iteration; otherwise, the generative adversarial network (GAN) model is entered. After entering the coordinate rotation optimization method iteration process, the optimal step size is calculated and the search is performed along the set direction, updating the position and fitness function values. Subsequently, after the iteration number k≮N1, the GAN acceleration process is entered. The updated position and fitness function values ​​are input into the GAN, and the GAN is trained using the input data. The trained GAN then outputs the predicted position and fitness function values. Then, after the iteration number k≮N1+N2, the coordinate rotation optimization method is used again to update the position and fitness function values, and the termination precision ε is set. The iteration is stopped and the optimal position X3 and the optimal fitness function value f(X3) are output. Finally, the optimal position output by the coordinate rotation optimization method and the least squares generative adversarial network represents the controller parameters of the doubly-fed induction fan, and the obtained parameters are used to obtain higher fan control performance.

[0073] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or new process transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for accelerating the optimization of wind turbine control parameters using least-squares generative adversarial networks, characterized in that, This paper combines a least-squares generative adversarial network model with a coordinate rotation optimization method for parameter optimization of doubly-fed induction generator (DFIG) wind turbine controllers. This improves the operating efficiency of DFIG wind turbines, reduces the required optimization time, and increases the speed of the optimization method. The steps involved in its application are as follows: Step (1): Initialize the doubly fed induction fan control model; The system equations for a wind turbine are: Among them, T m,DFIG T represents the mechanical torque of the wind turbine. e,DFIG ω represents the electromagnetic torque of the wind turbine, J represents the system parameter of the wind turbine generator, and ω represents the electromagnetic torque of the wind turbine generator. r Let dt be the rotor speed, and dt be the differential of the variable with respect to time t. Wind turbine mechanical power P m and electromagnetic power P e The relationship between them is: Where, ω m It is the mechanical angular velocity; Fan mechanical power P m for: Where R is the radius of the wind turbine rotor, ρ is the air density, and v wind π is the wind speed, and π is the mathematical constant pi. Power factor C p for: Where, β DFIG pitch angle The tip speed ratio λ is: Where, ω m,DFIG Wind turbine blade speed intermediate variable λ i for: The model of the doubly-fed induction fan is as follows: Among them, e' qs and e' ds These represent the equivalent internal q-axis and d-axis voltages, respectively. ds and v dr These represent the direct-axis voltage of the stator winding and the direct-axis voltage of the rotor winding, respectively. qs and v qr These represent the cross-axis voltage of the stator winding and the cross-axis voltage of the rotor winding, respectively. b It is the reference electric angular velocity, ω s It is the synchronous angular velocity, ω r It is the rotor angular velocity, L ss L is the stator winding inductance. rr L is the rotor winding inductance. m For the mutual inductance between the stator winding and the rotor winding, i ds and i qs R represents the direct-axis and quadrature-axis currents of the stator windings, respectively. s R represents the stator resistance. r Let R1 represent the rotor resistance, ∫dt be the integral of the variable with respect to time, and the intermediate derivation variable R1 for the wind turbine model be: R1=R s +R2 (8) The intermediate derivation variable R2 is: intermediate derivation variable T r for: Intermediate derivation variable L′ s for: To achieve optimal performance, four proportional-integral loops are employed; in the control loop, i qr and i dr The controller outputs v, which is the input reference value. qr1 and v dr1 , and the compensation voltage v qr2 and v dr2 Combined to form the output value of the rotor voltage; compensation voltage v qr2 and v dr2 for: The objective function for accelerating the optimization of wind turbine controller parameters is: Where, k Pi and k Ii These are the proportional and integral parameters of the proportional-integral controller, k. Pimin and k Iimin Let k be the minimum value of the proportional parameter and the integral parameter, respectively. Pimax and k Iimax v represents the maximum values ​​of the proportional parameter and the integral parameter, respectively. windmax and v windmin V represents the maximum and minimum wind speeds, respectively. s V is the system voltage. smax and V smin Q represents the maximum and minimum values ​​of the system voltage, respectively. s Q represents reactive power. smax and Q smin These represent the maximum and minimum reactive power values, respectively, and T1 is the sampling time for one operation of the wind power system. To integrate the objective function over a single sampling time, w1 is the weighted value of the rotor angular velocity error, w2 is the weighted value of the reactive power deviation, ∑ is the summation sign, and |·| represents the absolute value of the error. In a doubly-fed induction fan system, frequency stability is affected by the rotor angular velocity error. Voltage stability is affected by reactive power deviation. The goal of accelerating parameter optimization for doubly-fed induction fan controllers is to reduce rotor speed error. Minimize and make reactive power error Minimum, that is, minimum fitness value of objective function f(x), ω r and These are the precise and reference values ​​of the rotor speed, respectively. s and These are the precise and reference values ​​for reactive power, respectively. Step (2): Initialize the parameters of the doubly fed induction fan and randomly generate the controller initialization position X0; Step (3): The least squares generative adversarial network wind turbine control parameter optimization acceleration method determines whether the current iteration algebra k satisfies k < N1, where N1 is the initial iteration number. If the condition is met, the coordinate rotation optimization method is used for iteration. If the condition is not met, the least squares generative adversarial network model training is entered and the acceleration process is executed. Step (4): The coordinate rotation optimization method begins to iterate, calculates the fitness function value of each initial position X0, and searches for the global optimal solution; In each iteration, the selected optimal solution guides the variables to explore the region with the minimum fitness function value, finding a solution closer to the global optimum. The coordinate rotation optimization method transforms a multidimensional problem into a series of one-dimensional problems, alternately converting a multivariate optimization problem into a single-variable optimization problem. The search process involves alternating searches along the coordinate directions, allowing only one variable to change at a time while keeping the others constant. When the scaling parameter k... Pi When it changes, the integral parameter k Ii When the integration parameter k remains unchanged Ii When it changes, the proportional parameter k Pi Remain unchanged; The update strategy for the coordinate rotation optimization method is as follows: Step (4.1) Given the objective function f(x), randomly change the initialization position X0 as the starting position of the first iteration; Step (4.2) sets the search direction to n coordinate directions, and searches along one coordinate direction at a time. The search direction is represented as follows: Among them, e1, e2, ..., e n For the n coordinate directions set, [·] T It is the transpose symbol; The formula for iterative calculation using the optimal step size in step (4.3) is as follows: Where k is the iteration number, k = 1, 2, ..., N1, N1 is the initial iteration number, and i is the index of the current one-dimensional search loop. This represents the position of the i-th search in the k-th iteration. This represents the position of the search in the (i-1)th iteration of the k-th loop. e is the step size of the i-th search in the k-th iteration. i This represents the search direction for the i-th search. Step (4.4) uses a one-dimensional search to find the optimal step size. Where f(·) is the iterative formula in step (4.4), and the function Used to find the minimum value of the function f(·), and the corresponding minimum point. To find the optimal step size; If i < n in step (4.5), i = i + 1, return to step (4.3) and perform iterative calculation again; if i = n, proceed to the next step. In step (4.6), if k < N1, k = k + 1, i = 0, return to step (4.3) for the next iteration; if k = N1, end the iteration. The position information is updated by using a coordinate rotation optimization method. The position X1 and the corresponding fitness function value f(X1) are recorded during the iteration process and used as the input value for the next step of the least squares generative adversarial network. Step (5): If the algebra of the current iteration does not satisfy k < N1 or k > N2, where N2 is the iteration number of the least squares generative adversarial network, then enter the least squares generative adversarial network, learn the data output in step (4), and start the acceleration process. Step (6): Train the generative adversarial network using the input positions and corresponding fitness function values ​​obtained from the iterations in steps (4) and (5); the learning process of the generative adversarial network involves simultaneously training the generator and the discriminator, and the minimax objective function of the generative adversarial network. for: Where G is the generator, whose goal is to learn the input data and generate the output data; D is the discriminator, which determines whether the received data comes from real data or the generator G; x is the real data, z is the input of the generator, G(z) is the fake data generated by the generator; D(x) is the discriminator's probability assessment of the real data, and D(G(z)) is the discriminator's probability assessment of the fake data generated by the generator. The expected value of the true data distribution. The expected value of the fake data generated by the generator; logD(x) and log(1-D(G(z))) are the loss functions of the network, and log(·) is the logarithmic function to the base 10; Standard generative adversarial networks (GANs) suffer from two drawbacks: low-quality generated data and unstable training processes. This paper improves the method by replacing the cross-entropy loss in the standard GAN's objective function with least squares loss, resulting in the objective function of a least squares GAN. and for: Where a and b are the labels of the fake data and the real data, respectively, and c represents the fake data value that the generator wants the discriminator to believe; Step (7): The least squares generative adversarial network completes training through N2 iterations. The position X1 and the corresponding fitness function value f(X1) are input to the trained network to obtain the predicted optimal position X2 and the optimal fitness function value f(X2). Step (8): Since the prediction accuracy of the trained least squares generative adversarial network is not 100%, in order to improve the accuracy of parameter optimization, the coordinate rotation method is used for optimization in the last N3 iterations. Therefore, the coordinate rotation optimization method only performs N1+N3 iterations, which will greatly reduce the running time of the method and achieve the purpose of accelerating the method using the least squares generative adversarial network. Compared with the method without acceleration, the number of iterations of the coordinate rotation optimization method is reduced during the iteration process. Step (9): If the algebra of the current iteration satisfies k>N1+N2, calculate the coordinate rotation optimization method, and the calculation method is the same as in step (4); to make the results more accurate, update the method termination judgment criteria. Based on the judgment of the number of iterations, set the termination precision ε. If the number of iterations k > N1 + N2 + N3, stop the iteration and output the optimal position X3; Step (10): The optimal position output by the coordinate rotation optimization method and the least squares generative adversarial network represents the controller parameters of the doubly fed induction fan. The obtained parameters are then input into the doubly fed induction fan to obtain higher fan control performance.

Images