Simulation setting method for variable-pitch PID controller based on artificial neural network

Abstract

The invention provides a variable-pitch PID (Proportion Integration Differentiation) controller simulation setting method based on an artificial neural network. The variable-pitch PID controller simulation setting method comprises the following steps: solving real-time simulation power of a wind wheel by using a joint simulation variable-pitch control model; the wind wheel rated power, the wind wheel real-time simulation power and the difference between the wind wheel rated power and the wind wheel real-time simulation power serve as input of an artificial neural network, and parameters of a variable-pitch PID controller are calculated through an artificial neural network PID control algorithm; generating a variable pitch control signal according to the variable pitch PID control parameter; weight coefficient learning and gradient moving average are adopted in the variable pitch control signal generation process, loop iteration is carried out in the artificial neural network according to the RMSProp algorithm, and simulation online setting is carried out. According to the method, the control signal can be further optimized, the rotating speed fluctuation during the operation of the generator is reduced, repeated debugging and verification do not need to be carried out on a real machine of the wind generating set, and the variable-pitch PID controller parameters combining the specific application scene and the characteristics of the wind generating set can be obtained through a simulation mode.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and particularly to a simulation tuning method for a pitch PID controller based on an artificial neural network. Background Art

[0002] The pitch PID controller (Proportional-Integral-Derivative controller) of a wind turbine can automatically adjust the blade angle according to the change of wind speed, so that the wind turbine always operates at the optimal power output point. For example, when the wind speed is too high or too low, the pitch PID controller can limit the output power or speed of the wind turbine by adjusting the blade angle, thereby avoiding overload or damage. For the commonly used pitch PID controller at present, its control accuracy depends to a large extent on the settings of its proportional coefficient, integral coefficient, and derivative coefficient. However, due to the non-linear characteristics of the wind turbine, it is very difficult to tune the PID parameters of the pitch controller, resulting in the wind turbine often being unable to adjust in time when affected by external disturbances and operating in an unstable state.

[0003] In view of the above technical problems, a parameter self-tuning method for a torque-pitch controller of a megawatt-class asynchronous doubly-fed wind turbine with a publication number of CN103184972A is provided in the prior art. The method includes the following steps: after the unit starts, the yaw system first works to a specified position; after the self-check program passes, it will enter the free rotation process; when the generator speed reaches the switching speed, the pitch system fully opens the blade angle and the generator speed continues to rise; when the generator speed reaches the grid connection speed, the torque controller controls the generator speed to rise at a certain slope; when the generator speed reaches the rated speed, the converter controller keeps the generator torque constant, and the pitch controller controls the generator speed to stabilize at the rated speed. The pitch and torque controllers are composed of a BP neural network PID parameter self-tuning controller with an integral separation algorithm. The controller can automatically adjust the parameters according to the rules and optimize the parameters through online training according to the convergence rules.

[0004] However, when training a BP neural network using a fixed learning rate η, the above technical solutions have many drawbacks. Under dynamic wind conditions, it is difficult to balance the rapid response and stability requirements of the pitch control system;; When encountering turbulence, the pitch angle needs to be finely adjusted at a high frequency to suppress the fluctuation of the impeller speed. A fixed η may lead to a lag in the update of PID parameters (when η is low, Kp is adjusted too slowly to limit overspeed in time) or excessive oscillation (when η is high, Kd frequently crosses the resonance suppression threshold, exacerbating the fatigue load of the blade), losing balance in training requirements. Moreover, the above technical solutions are only applicable to megawatt-class variable-speed variable-pitch wind turbines, and real-time online tuning is required when the wind turbine is starting up. However, since the adjustment of PID controller parameters usually requires repeated experiments and optimization according to specific application scenarios and the characteristics of wind turbines, real-time online tuning through actual machine operation is time-consuming and costly, and may damage the wind turbine generator set in case of a fault during the tuning process. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention proposes a simulation tuning method for a pitch PID controller based on an artificial neural network, which does not require repeated experimental verification on the actual machine of the wind turbine generator set, and can obtain the parameters of the pitch PID controller combined with specific application scenarios and the characteristics of the wind turbine through simulation.

[0006] The technical solution adopted by the present invention is as follows:

[0007] In the first aspect, a simulation tuning method for a pitch PID controller based on an artificial neural network is provided, including: using a co-simulation pitch control model to obtain the real-time simulation power of the wind turbine according to the wind turbine speed and the generator resistance torque; taking the rated power of the wind turbine, the real-time simulation power of the wind turbine, and the difference between the rated power of the wind turbine and the real-time simulation power of the wind turbine as the inputs of the artificial neural network, and using the artificial neural network PID control algorithm to calculate the parameters of the pitch PID controller; generating a pitch control signal according to the pitch PID control parameters; adopting weight coefficient learning and gradient moving average for the process of generating the pitch control signal, and performing simulation online tuning through cyclic iteration in the BP neural network according to the RMSProp algorithm.

[0008] Further, the co-simulation pitch control model includes a random wind speed generation module, an aerodynamic load simulation interface, a wind turbine generator set dynamics model, and a pitch control model;

[0009] The random wind speed generation module is signal-connected to the aerodynamic load simulation interface, and is used to generate a random wind speed and input it into the aerodynamic load simulation interface, and then input it into the pitch control model through the aerodynamic load simulation interface and the wind turbine generator set dynamics model;

[0010] The aerodynamic load simulation interface is signal-connected to the wind turbine overall dynamics model and is used to calculate the aerodynamic load based on the random wind speed;

[0011] The wind turbine overall dynamics model is respectively signal-connected to the aerodynamic load simulation interface and the pitch control model, and is used to apply the aerodynamic load to the wind turbine overall dynamics model, obtain the wind turbine rotational speed and the generator drag torque, and calculate the wind turbine power; it is used to input the wind turbine power and the wind turbine rotational speed into the pitch control model; it is also used to feedback the pitch angle to the aerodynamic load simulation interface;

[0012] The pitch control model is used to calculate the pitch control signal according to the wind turbine power and the wind turbine rotational speed, and the pitch control signal includes the angular displacement, angular velocity, and angular acceleration signals of the pitch angle.

[0013] Furthermore, the pitch control model determines whether to start pitching through the wind turbine rotational speed and the random wind speed. An angular velocity limiting module and a filter are set at the output end of the pitch control signal, and the actuator is a first-order inertial system.

[0014] Furthermore, the co-simulation pitch control model is jointly constructed by TurbSim, Aerodyn, Simpack, and Simulink software.

[0015] Furthermore, the working process of the co-simulation pitch control model includes the following steps: the random wind speed generation module generates the random wind speed, and the aerodynamic load simulation interface calculates the aerodynamic load according to the random wind speed; the aerodynamic load is applied to the wind turbine overall dynamics model, the wind turbine rotational speed and the generator drag torque are obtained, and the wind turbine power is calculated according to the wind turbine rotational speed and the generator drag torque; the pitch control model calculates the pitch signal according to the wind turbine power and feedbacks the pitch signal to the wind turbine overall dynamics model. At the same time, the wind turbine overall dynamics model feedbacks the pitch angle and the wind turbine rotational speed to the aerodynamic load simulation interface for closed-loop control.

[0016] Furthermore, the artificial neural network includes a backpropagation neural network, a radial basis function network, a recurrent neural network, or a convolutional neural network.

[0017] Furthermore, when the artificial neural network is a backpropagation neural network, the artificial neural network PID control algorithm includes:

[0018] Determine the overall structure of the BP neural network, including specifying the specific number of nodes in the input layer and the hidden layer, and setting initial weighting coefficients for each layer; select the learning rate and the inertia coefficient; perform data sampling to obtain a given reference signal and the feedback signal of the system, and calculate the error based on the reference signal and the feedback signal; determine the input quantity and provide the input quantity to the BP neural network for processing; after completing the calculation of the control output, perform the learning and training of the BP neural network, use it to adjust the weighting coefficients of the output layer and the hidden layer, and perform the adaptive adjustment of the parameters of the pitch PID controller.

[0019] Furthermore, the process of generating the pitch control signal adopts weight coefficient learning and gradient moving average, and optimizes the adaptive learning rate according to the RMSProp algorithm when performing cyclic iteration in the BP neural network.

[0020] In a second aspect, an electronic device is provided, including: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the pitch PID controller simulation tuning method based on an artificial neural network described in the first aspect.

[0021] In a third aspect, a computer program product is provided, including a computer program / instructions, which when executed by a processor, implements the steps of the pitch PID controller simulation tuning method based on an artificial neural network described in the first aspect.

[0022] As can be seen from the above technical solutions, the beneficial technical effects of the present invention are as follows:

[0023] 1. A combined simulation pitch control model is constructed, which can complete the simulation online tuning of the pitch PID controller without performing the actual start-up operation of the wind turbine generator set.

[0024] 2. Combining the RMSProp algorithm and the BP neural network, the pitch PID controller can adjust the learning rate of the neural network according to the moving average of the gradients of each weight and threshold, dynamically adjust the learning rates of different weights and thresholds, and maintain a small oscillation during the response process; further optimize the control signal and reduce the rotational speed fluctuation during the operation of the generator. Description of the Drawings

[0025] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for use in the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0026] Figure 1 Schematic diagram of the co - simulation pitch control model architecture for the embodiments of the present invention;

[0027] Figure 2 Architecture of the pitch control model for the embodiments of the present invention Figure 1 ;

[0028] Figure 3 Architecture of the pitch control model for the embodiments of the present invention Figure 2 ;

[0029] Figure 4 Schematic diagram of the horizontal - direction random wind speed generated for the embodiments of the present invention;

[0030] Figure 5 Schematic diagram of the generator drag torque for the embodiments of the present invention;

[0031] Figure 6 Schematic diagram of the cyclic iteration using the RMSProp algorithm in the BP neural network for the embodiments of the present invention;

[0032] Figure 7 Comparison chart of the learning effects of each algorithm for the embodiments of the present invention;

[0033] Figure 8 Parameter variation of the pitch - controlled PID controller controlled by the BP - RMSProp algorithm during the simulation process for the embodiments of the present invention;

[0034] Figure 9 Comparison of the control performance effects of each algorithm for the embodiments of the present invention Figure 1 ;

[0035] Figure 10 Comparison of the control performance effects of each algorithm for the embodiments of the present invention Figure 2 ;

[0036] Figure 11 Schematic diagram of the online tuning method process of the pitch - controlled PID controller of the wind turbine based on the artificial neural network for the embodiments of the present invention. Detailed implementation manners

[0037] Hereinafter, embodiments of the technical solutions of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, so they are only examples and cannot be used to limit the protection scope of the present invention.

[0038] It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meanings understood by those skilled in the art to which the present invention belongs.

[0039] Embodiment

[0040] This embodiment provides a simulation online tuning method for the pitch PID controller of a wind turbine based on an artificial neural network, providing a reference for the efficient pitch and yaw control of the wind turbine.

[0041] To achieve the online tuning of the pitch PID controller of the wind turbine and enable the PID controller to obtain better control results, it is necessary to adjust the three control actions of proportional, integral, and differential to form a relationship of mutual cooperation and mutual restraint in the control quantity. The arbitrary non-linear expression ability of the artificial neural network can be used to achieve the optimal combination of PID control through learning the system performance. Selecting a suitable artificial neural network can establish a self-learning PID controller with parameters k p 、k i 、k d The types of artificial neural networks can include backpropagation neural network (BP), radial basis function network (RBF), recurrent neural network (RNN), and convolutional neural network (CNN). In this embodiment, the BP neural network is used as an example for illustration.

[0042] For the simulation tuning of the pitch PID controller, the pitch PID controller based on the BP neural network first constructs a Simat co-simulation pitch control model. Then, the wind turbine power output by the Simat co-simulation pitch control model is used as the input of the BP neural network, and the control parameters of the PID controller are output through the BP neural network, including the following content:

[0043] As Figure 1 shown, the Simat co-simulation pitch control model includes a random wind speed generation module, an aerodynamic load simulation interface, a wind turbine overall dynamics model, and a pitch control model. In a specific implementation, it is constructed by coordinating the four software of TurbSim, Aerodyn, Simpack, and Simulink. Simat is a co-simulation interface provided in the form of a module.

[0044] The random wind speed generation module is signal-connected to the aerodynamic load simulation interface, used to generate a random wind speed and input it into the aerodynamic load simulation interface, and then input it into the pitch control model through the aerodynamic load simulation interface and the wind turbine overall dynamics model. The random wind speed generation module is implemented by TurbSim software.

[0045] The aerodynamic load simulation interface is signal-connected to the wind turbine overall dynamics model, used to calculate the aerodynamic load according to the random wind speed. The aerodynamic load simulation interface is implemented by Aerodyn software.

[0046] The overall dynamic model of the wind turbine is respectively connected to the aerodynamic load simulation interface and the pitch control model signal, used to apply the aerodynamic load to the overall dynamic model of the wind turbine, obtain the wind turbine speed and the generator drag torque, and calculate the wind turbine power; used to input the wind turbine power and the wind turbine speed into the pitch control model; and also used to feedback the pitch angle to the aerodynamic load simulation interface. The overall dynamic model of the wind turbine is implemented by Simpack software.

[0047] The pitch control model is used to calculate the pitch control signal according to the wind turbine power and the wind turbine speed. Specifically, it determines whether to start pitching by the wind turbine speed and the random wind speed. The wind turbine power is used as the input of the artificial neural network to calculate the PID controller parameters. The pitch control signal includes the angular displacement, angular velocity, and angular acceleration signals of the pitch angle. The pitch control model is also used to feedback the pitch control signal to the overall dynamic model of the wind turbine. The pitch control model is implemented by Simulink software.

[0048] To establish the above-mentioned Simat co-simulation pitch control model, a series of co-simulation interface variables are set in this embodiment, as follows:

[0049] The actuator of the pitch angle is defined as an inertial system in Simulink software, and the pitch angle is adjusted by directly rotating the blade through Move_Marker in Simpack software. Specifically, Move_Marker is set as a rigid body hinged on the pitch bearing, retaining one degree of freedom of rotation around the z-axis of the blade coordinate. Its position is determined by the beta angular displacement, angular velocity, and angular acceleration signals output by Simulink software. The angular displacement is directly output by the controller, and the angular velocity and angular acceleration are obtained by superimposing the angular velocity signal with the differential module. The architecture of the pitch control model is as Figure 2 、 Figure 3 shown:

[0050] The SIMAT module is provided with 3 input ports and 3 output ports; the input ports receive the angular displacement, angular velocity, and angular acceleration signals of the pitch angle, and the output ports respectively output the wind turbine speed, the wind turbine power, and the horizontal wind load speed. Among them, an if logic module is added before the PID controller based on the BP neural network to judge whether to start the subsequent control steps according to the wind turbine power and the wind turbine speed. An angular velocity limit module and a filter are added before the pitch control signal output by the PID controller based on the BP neural network to prevent the pitch angle from changing too much in a short time and not conforming to the actual situation, or the oscillation phenomenon caused by the large fluctuation of the output signal. The actuator is completed by a first-order inertial system.

[0051] For the convenience of those skilled in the art to understand, the working process of the above-mentioned simat co-simulation pitch control model is briefly introduced below. The working process mainly includes the following steps:

[0052] Step S11: The random wind speed generation module generates a random wind speed, and the aerodynamic load simulation interface calculates the aerodynamic load according to the random wind speed

[0053] In this embodiment, the method of generating a random wind speed is illustrated by using TurbSim software: TurbSim software can generate random wind speed data. Through the average wind speed and turbulence intensity input by the user, combined with Fourier transform technology, the complex random wind speed signal is decomposed into multiple superimposed sine waves to simulate the real turbulence characteristics; in addition, the spatial correlation of the wind speed is also considered to ensure that the generated wind speed time series is more practical when simulating the performance of the wind turbine and conducting dynamic analysis, and these data can be used for more accurate load prediction and design optimization.

[0054] In some embodiments, the generated random wind speed field is based on the IEC Kaimal spectrum model (IEC KAI), conforms to the IEC61400-1 3rd Edition (1-Ed3) standard, the turbulence intensity is Class B, the standard deviation of the horizontal wind speed at the hub height is about 1.82 m / s (turbulence intensity 14%), the vertical wind speed distribution is described by the power law profile (PL), the default exponent is about 0.14, and the reference wind speed at the hub height is 10 m / s; the spatial coherence defaults to the IEC exponential decay model, the horizontal component coherence decays with distance and frequency, and the time step is 0.05 seconds; the random seed (2318573 and RANLUX algorithm) ensures the repeatability of the results, and the simulation duration is 500 seconds, meeting the requirements of the IEC standard for statistical stability; the overall design follows the IEC normal turbulence model (NTM), which is applicable to the typical working conditions for the load certification of wind turbines. Figure 4 To utilize the horizontal random wind speed with a mean value of 10 m / s generated by TurbSim.

[0055] The generated random wind speed is imported into the Aerodyn software, and combined with the force element 241 unit, the aerodynamic load can be calculated.

[0056] Step S12: Apply the aerodynamic load to the wind turbine overall dynamics model, obtain the wind turbine rotational speed and the generator resistance torque, and calculate the wind turbine power according to the wind turbine rotational speed and the generator resistance torque

[0057] In this embodiment, the wind turbine overall dynamics model is established in the Simpack software, and the expression of the model is:

[0058]

[0059] In the above formula, τ is the pitch angle actuator time constant, is the pitch angular velocity, β is the pitch angle, and P r is the wind turbine simulation power simulating real-time measurement values, and P ref is the rated power of the wind turbine, and k p is the proportional gain, and k i is the integral gain, and k d is the derivative gain, and t is time;

[0060] T gen is the generator drag torque, and f(w r ) is the torque function related to the motor rotor speed.

[0061] In this embodiment, taking Aerodyn as the simulation interface, the aerodynamic load is input as an excitation into the wind turbine overall dynamics model. The dynamics model calculates the wind turbine speed according to the dynamics differential equation provided by the Simpack software. The wind turbine speed is multiplied by the generator drag torque to obtain the wind turbine power. In a specific implementation manner, the generator drag torque is formed by adding a feedback torque force element formed by a speed-torque curve and a mass block in the Simpack software, as Figure 5 shown.

[0062] Step S13: The pitch control model calculates a pitch signal according to the wind turbine power and feeds the pitch signal back to the wind turbine overall dynamics model. At the same time, the wind turbine overall dynamics model feeds back the pitch angle and the wind turbine speed to the aerodynamic load simulation interface to form a closed-loop control.

[0063] Specifically, the pitch control model determines whether to start pitching according to the wind turbine speed and the horizontal wind speed component in the random wind speed. The pitch signal calculated according to the wind turbine power is fed back to the pitch angle of the aerodynamic load simulation interface as a control signal. The control signal changes the blade angle during the simulation process. After the blade angle changes, it will bring about a change in the aerodynamic load, and then the next round of simulation iteration is carried out according to the new aerodynamic load.

[0064] The above is the main working process of the simat co-simulation pitch control model.

[0065] Through the above simat co-simulation pitch control model, the wind turbine power can be calculated. Using the BP neural network PID control algorithm, taking the rated power of the wind turbine, the real-time simulation power of the wind turbine, and the difference between the rated power of the wind turbine and the real-time simulation power of the wind turbine as the inputs of the BP neural network, and using the BP neural network to calculate the PID controller parameters. This algorithm includes the following steps:

[0066] Step S21: Determine the overall structure of the BP neural network, including specifying the specific number of nodes in the input layer and the hidden layer, and setting the initial weighting coefficients for each layer.

[0067] Step S22: Select appropriate learning rate and inertia coefficient to ensure the learning effect and stability of the neural network. In this process of this embodiment, the initial number of iterations is set to k = 1.

[0068] Step S23: Conduct data sampling to obtain the given reference signal r(k) and the feedback signal y(k) of the system, and calculate the error e(k) based on this. The calculation formula for this error is e(k) = r(k) - y(k), which is used to reflect the deviation between the system output and the desired target.

[0069] Step S24: Determine the input quantity and provide it to the neural network for processing. This input quantity will be used to drive the neurons in the network to enable them to learn and adjust. Calculate the inputs and outputs of the neurons in each layer according to the previously set formula.

[0070] The output layer of the neural network will obtain the three adjustable parameters of the PID controller, which are the proportional gain k p , the integral gain k i and the derivative gain k d . Calculate the control output u(k) of the PID controller according to the incremental PID control formula. When using a BP neural network to optimize the PID controller parameters, the controller uses the incremental type:

[0071] u(k) = u(k - 1) + k p (error(k) - error(k - 1)) + k i error(k)

[0072] + k d (error(k) - 2error(k - 1) + error(k - 2))

[0073] In the above formula, u(k) is the control quantity at the current sampling moment, and error(k) is the difference between the desired output and the actual output at the current sampling moment.

[0074] Step S25: After completing the calculation of the control output, conduct the learning process of the BP neural network. Adjust the weighting coefficients of the output layer and the hidden layer to achieve the adaptive adjustment of the PID control parameters to adapt to the dynamically changing control requirements.

[0075] In this embodiment, the experimental simulation data shows that if the power is used as the input variable of the neural network, the fluctuation of the generator speed can be made smaller and the control accuracy of the PID controller can be higher.

[0076] Combined with the simat co-simulation pitch control model and the BP neural network PID control algorithm described above, the pitch PID controller of the wind turbine is simulated and tuned online, and the method includes the following steps:

[0077] Step S31: Use the simat co-simulation pitch control model to obtain the real-time simulation power of the wind turbine according to the wind turbine speed and the generator resistance torque

[0078] As described above, in this embodiment, Aerodyn is used as the simulation interface, and the aerodynamic load is input as an excitation into the overall dynamics model of the wind turbine. The dynamics model calculates the wind turbine speed according to the dynamics differential equation provided by the Simpack software, and multiplies the wind turbine speed by the generator resistance torque to obtain the real-time simulation power of the wind turbine.

[0079] Step S32: Use the rated power of the wind turbine, the real-time simulation power of the wind turbine, and the difference between the rated power of the wind turbine and the real-time simulation power of the wind turbine as the inputs of the artificial neural network, and use the BP neural network PID control algorithm to calculate the parameters of the pitch PID controller

[0080] Compare the real-time power simulation value of the wind turbine with the rated power P e and perform normalization to obtain the error at time t as:

[0081]

[0082] where P(t) is the sampled value of the power at time t.

[0083] For time t, use it as the input of the BP neural network in the form of a 1×3 vector:

[0084] input(t) = [r(t), p(t), e(t)]

[0085] Calculate the output of the neural network, where the activation function of the hidden layer is set as:

[0086]

[0087] The output of the output layer is:

[0088] output(t) = [K] = [k p (t), k i (t), k d (t)]

[0089] In order to make the initial values of the PID control parameters reasonable and improve the initial learning efficiency, set the activation function of the output layer as:

[0090]

[0091] where [kp0 , k i0 , k d0 is the initial value parameter of the PID after normalization, which can be set artificially. This enables the PID parameters output by the neural network at the beginning of learning to be near the global optimal value, shortening the learning time and cost.

[0092] In this embodiment, the numbers of neurons in the input layer, hidden layer, and output layer of the BP neural network are taken as 3, 8, and 3 respectively. For the convenience of operation, the weight values of the hidden layer and output layer are stored as matrices of 3×8 and 8×3 respectively. At the running time t, the weight matrices of the output and input layers are:

[0093]

[0094] According to the delta learning rule:

[0095]

[0096] where p(t) is the power sampling after normalization at time t, is element-wise multiplication.

[0097] Taking the output layer as an example, its weight learning is:

[0098] wo(t) = wo(t - 1) + η·δ o (t)·f(input(t)·wi(t - 1))

[0099] wi(t) = wi(t - 1) + η·δ i (t)·input(t)

[0100] Through the above steps, the variable pitch PID controller parameters can be output, and the parameters include the proportional gain k p , integral gain k i and derivative gain k d .

[0101] Step S33: Generate a variable pitch control signal according to the variable pitch PID control parameters output by the neural network in step 32

[0102] In this step, when generating the variable pitch control signal, the controller increment is as follows:

[0103] u(t) = u(t - 1) + k p (error(t) - error(t - 1)) + k i error(t)

[0104] + k d (error(t) - 2error(t - 1) + error(t - 2))

[0105] In the above formula, u(t) is the control quantity at the current sampling moment, and error(t) is the difference between the expected output and the actual output at the current sampling moment. During the calculation, the pitch PID control parameters output by the neural network in step 32 are multiplied by the corresponding incremental terms of the controller to obtain a preliminary output result.

[0106] Then, the preliminary output result is subjected to anti-normalization processing to generate a pitch control signal, which includes the angular displacement, angular velocity, and angular acceleration signals of the pitch angle.

[0107] Step S34: Apply weight coefficient learning and gradient moving average to the process of generating the pitch control signal in step S33, and perform cyclic iteration in the BP neural network according to the RMSProp algorithm for online simulation tuning.

[0108] In a specific implementation, the construction of the BP neural network and the implementation of the RMSProp algorithm are completed through the s-function module of Simulink. While the BP neural network generates the pitch control signal, it also generates a new round of weights and gradient averages, and then the pitch control signal is input to simpack; the weights and gradient moving averages are input to the next cycle in simulink for iterative simulation and online tuning.

[0109] During the forward propagation process of the S-function module, the input variables mainly include the following control input and output variables:

[0110] Control input: u(1), the system error e(t) at the current moment;

[0111] u(2), the system error e(t - 1) at the previous moment;

[0112] u(3), the system error e(t - 2) at the two previous moments;

[0113] u(4), the system output p(t) at the current moment;

[0114] u(5), the system output p(t - 1) at the previous moment;

[0115] u(6), the expected reference value r(t);

[0116] u(7), the previous control output output(t - 1).

[0117] Output variables: the value of the hidden layer weight matrix at the previous moment wi(t - 1);

[0118] the value of the output layer weight matrix at the previous moment wo(t - 1);

[0119] The value of the hidden layer weight matrix at the current moment \(w_i(t)\);

[0120] The value of the output layer weight matrix at the current moment \(w_o(t)\);

[0121] The cumulative hidden layer squared gradient \(E_g\) i (t) for the RMSProp algorithm;

[0122] The cumulative output layer squared gradient \(E_g\) o (t) for the RMSProp algorithm.

[0123] In each iteration loop, the average values of the weight and gradient movements are input into the S-function as iteration variables, and after calculation, learning, and correction, they are output. This process is accumulated until a relatively stable equilibrium value is reached. The data and parameter exchanges between modules are as Figure 6 shown. The pitch angle is used as a variable pitch control signal and is input into the wind turbine overall dynamics model in the Simpack software for the next cycle. The RMSProp algorithm performs iterations inside the s-func module. According to the \(E_g(t)\) output in each cycle, this value is updated and then output in each iteration. \(E_g(t)\) is used for the adaptive calculation of the learning rate in each iteration.

[0124] The cyclic iteration of the weight coefficient learning and the gradient moving average in the BP neural network is carried out according to the RMSProp algorithm. Adding the momentum method, the neural network learning process of optimizing the weights using the RMSProp algorithm is as follows:

[0125] Denote the gradient of the weight at time \(t\) as:

[0126] g o (t) = δ o (t) · f(input(t) · w_i(t - 1))

[0127] g i (t) = δ i (t) · input(t)

[0128] The average of the gradient movement is used as the coefficient of the adaptive learning rate and acts with the global learning rate. The calculation method of the average value of the gradient movement is:

[0129] E[g o (t)] = βE[g o (t - 1)] + (1 - β)g o 2 (t)

[0130] E[g i (t)] = βE[g i (t - 1)] + (1 - β)g i2 (t)

[0131] where β is the attenuation coefficient, which takes the value of 0.9 in this embodiment. This makes the moving average of the gradient consider both the weighted sum of historical gradient squares and the weight of the current gradient square; and the influence of the historical gradient on the learning rate will decay exponentially, thus preventing the learning rate from being too small. At the same time, for any weight w ij , the corresponding learning rate becomes:

[0132]

[0133] where η is the global learning rate, η ij (t) is the learning rate corresponding to the (i, j) - th element in the weight matrix at time t, ∈ is the initial minimum value to prevent division by zero in the denominator. Thus, the learning rate matrices of the input layer and the output layer are:

[0134]

[0135] On this basis, the momentum method is added to accelerate convergence, and the weight learning formula is as follows:

[0136] wo ij (t)=wo ij (t - 1)+αΔwo ij (t - 1)+η o (t)δ o (t)f(input(t)·wi(t - 1))

[0137] wi ij (t)=wi ij (t - 1)+αΔwi ij (t - 1)+η i (t)δ i (t)input(t)

[0138] where α is the inertia factor, Δwi ij (t - 1)=wi ij (t - 1)-wi ij (t - 2) is the momentum term. For any weight, the RMSProp algorithm will dynamically adjust the size of the learning rate according to the information of its historical and current gradients, preventing the learning rate from being too large to cause oscillations or too small to result in a too slow learning rate. Taking the step signal as the reference signal, setting the global learning rate to 0.3 and the inertia factor to 0.3, the learning effects of only using the BP neural network and using the PMSProp algorithm combined with the momentum method are compared as Figure 7 shown. From Figure 7As can be seen, while maintaining the adaptive learning rate, RMSProp introduces a decay coefficient by calculating the exponentially weighted average of the squared gradients for each parameter, avoiding the problem of the learning rate gradually decreasing during training. In the case of different feature scales, RMSprop can maintain stable parameter updates, avoiding the problem of some parameters being updated too fast or too slow, and there is no need for cumbersome learning rate debugging. The adaptive nature of the algorithm ensures fast convergence, and when training deep neural networks such as convolutional neural networks and recurrent neural networks, it can quickly find the optimal solution.

[0139] Figure 8 It is the parameter variation of the pitch PID controller controlled by the BP-RMSProp algorithm during the simulation process.

[0140] To verify the effect of the technical solution of this embodiment, statistical analysis is carried out on the performance of three wind turbine pitch control strategies, traditional PID, BP neural network, and BP-RMSProp, to examine the performance and stability of each algorithm under random wind speed conditions, providing a basis for selecting the optimal algorithm. The stability of the control system is evaluated by the mean value, variance, and overshoot. The stability of the control is evaluated by intercepting the time period from 200s to 300s when the wind speed is higher than the rated value and the control strategy is operating normally, and the comparison results are as Figure 9 、 Figure 10 shown.

[0141] As can be seen from the figure, the rated speed of the wind turbine is 0.83395 (rad / s), the average value of the wind turbine speed using the PID algorithm is 0.8375, BP is 0.8400, and BP-RMS is 0.8365. The overshoot is a key indicator, which reflects the oscillation degree of the system near the set value. If the control system has too much overshoot, it will lead to a decrease in reliability and lifespan. The output of BP-RMS is closer to the set target with a lower overshoot, more sensitive response, and better adaptability. In terms of statistical variance, BP-RMS shows the lowest variance (0.0000109), indicating that the output of this algorithm has the smallest fluctuation range and the highest stability compared to other algorithms. While BP has the highest variance (0.0000552), and this algorithm is affected by the fixed learning rate during the control process and will have a "sawtooth" oscillation when approaching the optimal value. The neural network using only the negative gradient algorithm has a slow convergence speed and is extremely prone to falling into local minima. The output has a large uncertainty and insufficient stability. It can be seen that BP-RMS has the lowest overshoot, significantly lower than other algorithms, and it can achieve a better output response during the control process. Further verifying the excellent control performance.

[0142] The technical solution provided in this embodiment combines the RMSProp algorithm and the BP neural network, which can adjust the learning rate of the neural network according to the moving average of the gradients of each weight and threshold, dynamically adjust the learning rates of different weights and thresholds, and realize the simulation online tuning of the pitch PID controller of the wind turbine generator set. Through the simulation of random wind speeds higher than the rated value, the controller adjusts the pitch angle to minimize the difference between the actual speed and the rated speed of the fan. The simulation results verify the expected performance of the proposed pitch controller. Its lowest mean value and overshoot indicate that the algorithm can better meet the control objectives and maintain a small oscillation during the response process, which is of great significance for the adaptability and robustness of the wind turbine pitch control under dynamic wind speed conditions; it can complete the online tuning of the pitch PID controller without performing the actual start-up operation of the wind turbine generator set.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered by the scope of the claims and the description of the present invention.

Claims

1. A simulation tuning method for a variable pitch PID controller based on an artificial neural network, characterized in that: include: Use the joint simulation pitch control model to obtain the real-time simulated power of the wind rotor based on the wind rotor speed and generator resistance torque; The rated power of the wind rotor, the real-time simulated power of the wind rotor, and the difference between the rated power of the wind rotor and the real-time simulated power of the wind rotor are used as the input of the artificial neural network, and the parameters of the variable pitch PID controller are calculated using the artificial neural network PID control algorithm; Generate a pitch control signal according to the pitch PID control parameter; The process of generating pitch control signals adopts weight coefficient learning and gradient moving average, and the RMSProp algorithm is iterated in the BP neural network to perform simulation online tuning.

2. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 1 is characterized in that: The joint simulation variable pitch control model includes a random wind speed generation module, an aerodynamic load simulation interface, a wind turbine whole machine dynamics model and a variable pitch control model; The random wind speed generation module is connected to the aerodynamic load simulation interface signal, and is used to generate random wind speed and input it into the aerodynamic load simulation interface, and then input it into the variable pitch control model through the aerodynamic load simulation interface and the wind turbine whole machine dynamics model; The aerodynamic load simulation interface is connected to the wind turbine whole machine dynamics model signal, and is used to calculate the aerodynamic load according to the random wind speed; The wind turbine whole machine dynamics model is respectively connected with the aerodynamic load simulation interface and the pitch control model signal, and is used to apply the aerodynamic load to the wind turbine whole machine dynamics model, obtain the wind rotor speed and the generator resistance torque and calculate the wind rotor power; is used to input the wind rotor power and the wind rotor speed into the pitch control model; and is also used to feed back the pitch angle to the aerodynamic load simulation interface; The pitch control model is used to calculate a pitch control signal according to the wind rotor power and the wind rotor speed. The pitch control signal includes an angular displacement, an angular velocity, and an angular acceleration signal of the pitch angle.

3. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 2 is characterized in that: The pitch control model determines whether to start the pitch control based on the rotor speed and random wind speed. An angular velocity limiting module and a filter are set at the output end of the pitch control signal, and the actuator is a first-order inertial system.

4. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 2 is characterized in that: The joint simulation pitch control model is collaboratively constructed using TurbSim, Aerodyn, Simpack and Simulink software.

5. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 2 is characterized in that: The working process of the joint simulation pitch control model includes the following steps: The random wind speed generation module generates random wind speeds, and the aerodynamic load simulation interface calculates aerodynamic loads based on the random wind speeds; Apply aerodynamic loads to the wind turbine dynamics model to obtain the rotor speed and generator resistance torque, and calculate the rotor power based on the rotor speed and generator resistance torque; The pitch control model calculates the pitch signal according to the wind rotor power and feeds the pitch signal back to the wind turbine overall dynamics model. At the same time, the wind turbine overall dynamics model feeds back the pitch angle and wind rotor speed to the aerodynamic load simulation interface for closed-loop control.

6. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 1, characterized in that: The artificial neural network includes a back propagation neural network, a radial basis function network, a recurrent neural network or a convolutional neural network.

7. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 6 is characterized in that: When the artificial neural network is a back propagation neural network, the artificial neural network PID control algorithm includes: Determine the overall structure of the BP neural network, including the specific number of nodes in the input layer and hidden layer, and set the initial weight coefficients for each layer; Select the learning rate and inertia coefficient; Performing data sampling to obtain a given reference signal and a system feedback signal, and calculating an error based on the reference signal and the feedback signal; Determine the input quantity and provide it to the BP neural network for processing; After completing the calculation of the control output, the BP neural network is trained and the weighted coefficients of the output layer and the hidden layer are adjusted to perform adaptive adjustment of the parameters of the variable pitch PID controller.

8. The artificial neural network-based variable pitch PID controller simulation tuning method according to claim 7 is characterized in that: The process of generating pitch control signals adopts weight coefficient learning and gradient moving average, and the adaptive learning rate is optimized according to the RMSProp algorithm during cyclic iteration in the BP neural network.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the artificial neural network-based variable pitch PID controller simulation tuning method as described in any one of claims 1-8.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the artificial neural network-based variable pitch PID controller simulation tuning method described in any one of claims 1 to 8 are implemented.

Images