Wind power plant active load shedding optimization method based on dual-input convex neural network

Abstract

The invention provides a wind power plant active load shedding optimization method based on a dual-input convex neural network, and belongs to the technical field of wind power plants, and the method comprises the steps: constructing a simulation environment, so as to obtain simulation data; fitting simulation data by adopting an input convex neural network model to obtain a nonlinear convex model of the wind turbine generator; fitting and generating time window load time sequence data by inputting a convex neural network model, and obtaining a nonlinear convex model of the fatigue load based on a mapping relation between the time window load time sequence data and the equivalent fatigue load; taking the main shaft torque output by the nonlinear convex model of the wind turbine generator as the input of the nonlinear convex model of the fatigue load, and obtaining a dual-input convex neural network pre-training model; and performing active power optimization control solution on the dual-input convex neural network pre-training model to obtain an active power distribution optimal solution meeting the wind power plant safety and stability constraint.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind farms, and in particular relates to a wind farm active load reduction optimization method based on a dual-input convex neural network. Background Art

[0002] In the traditional active power distribution method, after receiving the grid command, the wind farm active power controller directly sends the active power reference value to each wind turbine according to the principle of average distribution based on the available power of each unit, without considering the fatigue load of the wind turbine. The specific method is: obtain the available active power of each wind turbine and the grid dispatching command; the power of each wind turbine is: grid dispatching command × available power of the i-th wind turbine ÷ total available power of the wind farm, ignoring the wind speed conditions and control inertia of each wind turbine, which will further aggravate the fatigue damage of the wind turbine. On the other hand, in the evaluation index of fatigue load, equivalent fatigue load is often used as a quantitative index. The calculation of equivalent fatigue load requires the use of rain flow counting method to statistically calculate the cyclic mean and amplitude of load time series data. This calculation process is difficult to solve online and has strong nonlinearity and non-convexity.

[0003] In response to the above challenges, the present invention proposes an optimization control strategy for active load shedding in wind farms based on a dual-input convex neural network. Summary of the invention

[0004] The present invention provides a wind farm active load reduction optimization method based on a dual-input convex neural network, so as to at least solve the problem that the traditional active power allocation strategy in the prior art fails to fully consider the influence of fatigue load, thus restricting the economic benefits of the wind power industry.

[0005] The embodiment of the present application provides a method for optimizing active load shedding in a wind farm based on a dual-input convex neural network, the method comprising: Build a simulation environment to obtain simulation data; The simulation data is fitted by using the input convex neural network model to obtain the nonlinear convex model of the wind turbine, and the main shaft torque is output through the nonlinear convex model of the wind turbine; The time window load time series data is generated by fitting the input convex neural network model, and the nonlinear convex model of fatigue load is obtained based on the mapping relationship between the time window load time series data and the equivalent fatigue load; The main shaft torque output by the nonlinear convex model of the wind turbine is used as the input of the nonlinear convex model of fatigue load to obtain a dual-input convex neural network pre-training model; The active power optimization control solution is solved for the dual-input convex neural network pre-training model to obtain the optimal solution for active power distribution that meets the safety and stability constraints of the wind farm.

[0006] Further, a simulation environment is constructed to obtain simulation data, specifically including: Construct a simulation environment to obtain the mapping relationships between the main shaft torque and the inflow wind speed, generator speed, output power, pitch angle, generator torque, and active power reference value of the wind farm respectively, and obtain simulation data based on these mapping relationships; Mix the simulation data to obtain a data set and the test set ; Among them, represents the input data, represents the output data; the input data includes: inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference value, and the output data includes: main shaft torque; The simulation environment includes: an aerodynamic simulation environment, a drive train simulation environment, a generator simulation environment, a tower simulation environment, and a control system simulation environment.

[0007] Further, an input convex neural network model is used to fit the simulation data to obtain a non-linear convex model of the wind turbine. Among them, the expression of the input convex neural network model is:

[0008] In the formula, is a convex function with respect to , is the output vector of the input convex neural network model, is the set of weight and bias parameters of the input convex neural network model, is the output vector of the (i + 1)-th layer of the input convex neural network model, is the non-direct layer weight of the i-th layer of the input convex neural network model, is the output of the i-th layer of the input convex neural network model, is the direct layer weight of the i-th layer of the input convex neural network model, is the input vector of the input convex neural network model, is the bias of the i-th layer of the input convex neural network model, is the set of direct layer weights of the input convex neural network model, is the set of non-direct layer weights of the input convex neural network model, is the set of bias terms of the input convex neural network model, is the activation function; Use the data set to train the input convex neural network model; Improve the convergence and accuracy of the input convex neural network model through the loss function to obtain a non-linear convex model of the wind turbine. The expression of the loss function is:

[0009] In the formula, is the loss function, is to minimize the model error, is to minimize the model gradient error, is the output vector of the improved input convex neural network model, is the input vector of the improved input convex neural network model, is the gradient of the input convex neural network model, is the simulation environment gradient calculated by using an automatic gradient calculation tool.

[0010] Furthermore, the time window load time series data is generated by fitting with an input convex neural network model. Based on the mapping relationship between the time window load time series data and the equivalent fatigue load, a non-linear convex model of the fatigue load is obtained, specifically including: Slice the obtained spindle torque with a sliding window of a preset window length to divide it into a spindle torque test set and a spindle torque data set; Through the rain flow counting method, calculate the load cycle amplitude and mean value of the spindle torque of each sliding window to obtain the equivalent fatigue load, and its calculation formula is:

[0011] In the formula, is the load amplitude of the i-th level, is the number of cycles of the load at the level, m is the Wohler index of the S-N curve, is the equivalent fatigue load of the N-th segment of the load signal; Use the input convex neural network model to fit the spindle torque to generate time window load time series data;

[0012] Furthermore, perform active power optimization control solution on the dual input convex neural network pre-training model to obtain the optimal solution of active power distribution that meets the safety and stability constraints of the wind farm, specifically including: Construct the active power reduction optimization objective as:

[0013] Among them, is the equivalent fatigue load of the i-th wind turbine calculated according to the dual input convex neural network pre-training model, obtained from the lookup table constructed according to and is the active power reduction optimization objective function; The constraints for constructing the optimization problem are as follows:

[0014] Among them, is the reference value of active power, is the maximum available power of a single unit, is the minimum available power, is the equality constraint function, is the sum of active power of the whole field, is the power grid dispatching instruction, is the inequality constraint function, is the change in active power within 1 time step, is the maximum power ramp constraint of a single unit; The model predictive control algorithm is used to optimize and solve the pre-trained model of the dual-input convex neural network, and the optimal solution of active power distribution that meets the safety and stability constraints of the wind farm is obtained.

[0015] Furthermore, the model predictive control algorithm is the sequential least squares programming method; Using the model predictive control algorithm to optimize and solve the pre-trained model of the dual-input convex neural network, obtaining the optimal solution of active power distribution that meets the safety and stability constraints of the wind farm, including: According to the traditional proportional distribution algorithm, calculate the reference active power of each wind turbine in the wind farm as the initial point; Generate a fatigue load lookup table for wind turbines according to the pre-trained model of the dual-input convex neural network on a scale of 10W; Use the sequential least squares programming method to solve the optimization problem: At the k-th iteration, use the sequential least squares programming method to perform a first-order Taylor expansion on the active power reduction optimization objective and the constraints of the optimization problem: For the active power reduction optimization objective:

[0016] Among them, is the active power reduction optimization objective function of the reference value of active power, is the reference value of active power at the k-th iteration at for a first-order Taylor expansion, is the reference value of active power at the k-th iteration at the gradient vector at, is the search direction; For the constraints of the optimization problem:

[0017]

[0018] In the formula, For the first-order Taylor expansion of the equality constraint function at the point during the k-th iteration, is the gradient vector of the equality constraint function at the point during the k-th iteration; For the first-order Taylor expansion of the inequality constraint function at the point during the k-th iteration, is the gradient vector of the inequality constraint function at the point during the k-th iteration; The sequential least squares programming method transforms the original problem into a constrained quadratic programming problem by introducing Lagrange multipliers and penalty functions, and its expression is:

[0019] where, is the Hessian matrix of the objective function at the point; The search direction is obtained by solving the quadratic programming problem; The current point is updated using a fixed step size, and its expression is:

[0020] where, is the step size, which is taken as 1 here, is the update direction at the step, is the active power reference value updated after the step iteration; After completing the iteration, the optimal solution of the active power distribution that satisfies the safety and stability constraints of the wind farm is obtained; The iteration is terminated when any of the following conditions is met: The change in the objective function is less than the threshold; The gradient norm is 0; The maximum number of iterations is reached.

[0021] Furthermore, an aerodynamic simulation environment is constructed, including: Calculating the aerodynamic torque and thrust, and the calculation formulas are as follows:

[0022]

[0023] where, is the aerodynamic torque, is the thrust, is the power coefficient, is the thrust coefficient, is the blade length, is the air density, is the effective wind speed on the wind turbine blade, is the tip speed ratio, and its definition formula is ; Build a simulation environment for the drive train, including: Adopt a single mass block model, and combine the rotor mass and the generator mass into an equivalent mass. The formula is as follows:

[0024] According to the low-speed shaft motion equation, calculate the rotational speed relationship between the rotor and the generator. The formula is as follows:

[0025]

[0026] Among them, is the equivalent mass, is the rotor mass, is the generator mass, is the gearbox transmission ratio, is the generator torque, is the rotor, is the rotational speed of the generator.

[0027] Furthermore, build a simulation environment for the generator, including: In the torque control loop, adopt vector control with millisecond-level high-precision response and ignore the dynamic characteristics, that is:

[0028] Among them, is the generator reference torque.

[0029] Furthermore, build a simulation environment for the tower, including: The calculation formula for the front and rear bending moments of the tower is:

[0030] Among them, represents the front and rear bending moments of the tower, represents the height of the tower.

[0031] Furthermore, build a simulation environment for the control system, including: When the wind speed is lower than the rated wind speed , it is defined as the first mode; When the wind speed is greater than the rated wind speed , it is defined as the second mode; When operating in the first mode, the pitch control stops, ; When operating in the second mode, the measured generator speed is filtered through a low-pass filter, and the filtered speed is:

[0032] Where is the filtered speed, is the time constant of the filter, is the generator speed, is the Laplace operator; Based on the deviation between the filtered speed and the rated value of the speed , the reference value of the pitch angle is calculated, and its expression is:

[0033] Where, is the reference value of the pitch angle, is the proportional gain of the PI controller, is the integral gain of the PI controller, is the correction factor, , where, and are constants.

[0034] It can be seen from the above technical solutions that the present invention has the following advantages: In the active power shedding optimization method for a wind farm based on a dual-input convex neural network provided by this application, the fatigue load is suppressed by adjusting the reference value of the active power of each wind turbine, and the additional operation and maintenance costs caused by excessive fatigue of the units are reduced; The present invention adopts an input convex neural network model, models the equivalent fatigue load of the main shaft torque of the wind turbine as a convex function form based on a sliding time window, and transforms the non-convex optimization problem into a convex optimization problem for solution, reducing the solution difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of this application, the drawings required for description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 FIG. is a flowchart of the active power shedding optimization method for a wind farm based on a dual-input convex neural network.

[0037] Figure 2 FIG. is a structural diagram of the input convex neural network model.

[0038] Figure 3This is the prediction effect diagram of the main shaft torque of the nonlinear convex model of the wind turbine.

[0039] Figure 4 This is the equivalent fatigue load prediction effect diagram of the dual-input convex neural network pre-training model. DETAILED DESCRIPTION

[0040] In order to make the purpose, features and advantages of this application more obvious and easy to understand, the technical solution protected by this application will be clearly and completely described below using specific embodiments and drawings. Obviously, the embodiments described below are only part of the embodiments of this application, not all of them. Based on the embodiments in this patent, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this patent.

[0041] The existing active power allocation method ignores the wind speed conditions and control inertia of each wind turbine, which will further aggravate the fatigue damage of the wind turbine. The method proposed by the present invention allocates the total active power of the wind turbine required by the grid dispatching instruction according to the working state and wind speed of each wind turbine, taking into account the safety and stability constraints of the wind farm, thereby reducing the fatigue load of the entire farm.

[0042] In the calculation of equivalent fatigue load, in order to include it in the optimization index, it is required to have online computing capability. In order to reduce the difficulty of solving, the model should have convexity. To this end, the present invention adopts an input convex neural network to model the equivalent fatigue load of the wind turbine main shaft torque as a convex function based on a sliding time window, and transforms the non-convex optimization problem into a convex optimization problem for solution, thereby reducing the difficulty of solving.

[0043] The technical solution proposed in the embodiments of the present application is described in detail below with reference to the accompanying drawings.

[0044] Figure 1 The flowchart of a wind farm active load reduction optimization method based on a dual-input convex neural network provided in an embodiment of the present application. Figure 1 As shown, an optimization method for active load shedding of a wind farm based on a dual-input convex neural network is provided in an embodiment of the present application, and specifically comprises the following steps: Step S1: construct a simulation environment to obtain simulation data; Step S2: fitting the simulation data using the input convex neural network model to obtain a nonlinear convex model of the wind turbine, and outputting the main shaft torque through the nonlinear convex model of the wind turbine; Step S3: generating time window load time series data by fitting the input convex neural network model, and obtaining a nonlinear convex model of fatigue load based on the mapping relationship between the time window load time series data and the equivalent fatigue load; Step S4: Use the main shaft torque output from the non-linear convex model of the wind turbine as the input to the non-linear convex model of the fatigue load to obtain a pre-trained model of the dual-input convex neural network, and maintain the convexity of the pre-trained model of the dual-input convex neural network. The inputs of the pre-trained model of the dual-input convex neural network are: inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference, historical data of main shaft torque and tower thrust, and the output is the equivalent fatigue load. Figure 4 It is the prediction effect of the equivalent fatigue load (Damage Equivalent Load) of the pre-trained model of the dual-input convex neural network. Figure 4 In it, Damage Equivalent Load, that is, the equivalent fatigue load, is used to describe that in fatigue analysis, the complex load history is simplified into an equivalent simple load to facilitate the evaluation of the fatigue life of the entire main shaft of the wind farm under this load. The abscissa represents time, with the unit of second, indicating the change of data over time; the ordinate represents the equivalent fatigue load; Rain flowcounting method represents the rain flow counting method, which is used to analyze fatigue load data to determine the number and amplitude of load cycles, so as to evaluate the fatigue life of the entire main shaft of the wind farm; CNN method represents the Convolutional Neural Network method. The figure shows the variation of the equivalent fatigue load calculated by two methods (rain flow counting method and CNN method) over time. By comparing the results of these two methods, their performance and accuracy in predicting the equivalent fatigue load can be analyzed.

[0045] Step S5: Conduct active optimization control solution on the pre-trained model of the dual-input convex neural network to obtain the optimal solution of active power distribution that meets the safety and stability constraints of the wind farm.

[0046] Build a simulation environment to obtain simulation data, specifically including: Build a simulation environment to obtain the mapping relationships between the main shaft torque and the inflow wind speed, generator speed, output power, pitch angle, generator torque, and active power reference value of the wind farm respectively, and obtain simulation data based on these mapping relationships; Mix the simulation data to obtain a data set and the test set , where represents the input data, represents the output data; the input data includes: inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference value, and the output data includes: main shaft torque.

[0047] The simulation environment includes: an aerodynamic simulation environment, a drive train simulation environment, a generator simulation environment, a tower simulation environment, and a control system simulation environment.

[0048] The inflow wind speed includes high wind speed data and low wind speed data, and the high wind speed data and low wind speed data are used as the inflow wind speed. Simulation is carried out through the constructed simulation environment to obtain simulation data; after mixing the obtained simulation data, it is divided into a data set required for model training and a test set . is the input, including inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference value; is the output, including main shaft torque.

[0049] Construct an aerodynamic simulation environment, including: Calculate the aerodynamic torque and thrust, and the calculation formulas are as follows:

[0050]

[0051] Among them, is the aerodynamic torque, is the thrust, is the power coefficient, is the thrust coefficient, is the blade length, is the air density, is the effective wind speed on the wind turbine blade, is the tip speed ratio, and the definition formula is ; the calculation through the above formula provides basic data for subsequent obtaining of the mapping relationship.

[0052] Construct a drive train simulation environment, including: Assume that the drive train is rigidly connected, and adopt a single mass block model. Combine the wind turbine mass and the generator mass into an equivalent mass, and the formula is as follows:

[0053] According to the low-speed shaft motion equation, calculate the rotational speed relationship between the rotor and the generator, and the formula is as follows:

[0054]

[0055] Among them, is the equivalent mass, is the wind turbine mass, is the generator mass, is the gearbox transmission ratio, is the generator torque, is the rotor, is the rotational speed of the generator; The above calculation results reflect the mechanical relationships of various parts in the transmission system, are interrelated with other environmental parameters, and affect the final mapping relationship.

[0056] Build a generator simulation environment, including: In the torque control loop, vector control with millisecond-level high-precision response is adopted, ignoring the dynamic characteristics, that is:

[0057] where, is the generator reference torque; Here, the generator torque is approximately equal to the generator reference torque , indicating that under this control strategy, the actual generator torque can closely track the generator reference torque. Through precise control of the generator torque , it can ensure that the generator outputs stable electric energy to meet the requirements of the power grid. In the case where the power grid has strict requirements for power quality and power stability, precise regulation of the generator torque can enable the wind turbine to better connect to the power grid and improve the reliability and stability of power supply.

[0058] Build a tower simulation environment, including: The calculation formula for the bending moment of the tower in the front and rear directions is:

[0059] where, represents the bending moment of the tower in the front and rear directions, represents the height of the tower.

[0060] Build a control system simulation environment, including: The power coefficient of the wind turbine control system ; When the wind speed is lower than the rated wind speed , the pitch control stops, that is , and the generator torque is adjusted to track the optimal rotor speed. At this time, ; is the pitch angle (whether the physical meaning expressed by in other parts is the pitch angle, (it is necessary to ensure that one parameter has and only has one physical meaning)); When the wind speed is greater than the rated wind speed , the generator torque remains at the rated value, and the pitch control is started to prevent the generator speed Overspeed, at this time, the maximum power of the wind turbine is limited to the rated value, that is ; is the maximum power of the wind turbine, is the available power of the wind turbine, is the rated power of the wind turbine; When the wind speed is lower than the rated wind speed , it is defined as the first mode; When the wind speed is greater than the rated wind speed , it is defined as the second mode; When operating in the first mode, the pitch control stops, ; When operating in the second mode, the measured generator speed is filtered through a low-pass filter, and the filtered speed is:

[0061] where, is the filtered speed, is the time constant of the filter, is the generator speed, is the Laplace operator; Based on the deviation between the filtered speed and the rated value of the speed , the reference value of the pitch angle is calculated, and its expression is:

[0062] where, is the reference value of the pitch angle, is the proportional gain of the PI controller, is the integral gain of the PI controller, is the correction factor, , where, and are constants.

[0063] In order to model the mapping relationships between the main shaft torque and the inflow wind speed, generator speed, output power, pitch angle, generator torque, and active power reference value of the wind farm as a non-linear convex model, the present invention uses an input convex neural network model for fitting, and the structure of the input convex neural network model is as Figure 2 shown.

[0064] Using the input convex neural network model to fit the data set , a non-linear convex model of the wind turbine is obtained, where the expression of the input convex neural network model is:

[0065] In the formula, is a convex function with respect to ; is the output vector input to the convex neural network model; is the set of weight and bias parameters of the convex neural network model; is the output vector of the (i + 1)-th layer of the convex neural network model; is the non-direct layer weight of the i-th layer of the convex neural network model; is the output of the i-th layer of the convex neural network model; is the direct layer weight of the i-th layer of the convex neural network model; is the input vector of the convex neural network model; is the bias of the i-th layer of the convex neural network model; is the set of direct layer weights of the convex neural network model; is the set of non-direct layer weights of the convex neural network model; is the set of bias terms of the convex neural network model; is the activation function; In the convex neural network model constructed by the present invention, are all non-negative real numbers, and the convex non-decreasing function Smoothed relu is selected as the activation function . To ensure the accuracy and generalization of the convex neural network model, several direct layers are added to the convex neural network model , and a bias term is added to .

[0066] Before constructing the convex neural network model, a data set is obtained. The data set is used to train the convex neural network model. The data set includes the inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference value, and main shaft torque; the data set is used to train the convex neural network model so as to accurately predict the non-linear convex characteristics of the wind turbine. Specifically, the data set is obtained through a simulation environment and is divided into a data set for training the model and a test set for evaluating the model performance after being mixed.

[0067] These datasets are used to train an input convex neural network model, which aims to simulate the relationship between the equivalent fatigue load of the main shaft torque of a wind turbine and parameters such as wind speed, generator speed, and output power. Through training, the model can learn the non-linear relationship between these parameters and can predict the fatigue load of the wind turbine under different operating conditions, thereby being used to optimize the active load shedding control strategy of the wind farm.

[0068] The convergence and accuracy of the input convex neural network model are improved through the loss function to obtain a non-linear convex model of the wind turbine. The expression of the loss function is:

[0069] In the formula, is the loss function, is to minimize the model error, is to minimize the model gradient error, is the output vector of the improved input convex neural network model, is the input vector of the improved input convex neural network model, is the gradient of the input convex neural network model, is the simulation environment gradient calculated using an automatic gradient calculation tool, making the gradient of the input convex neural network model close to the simulation environment gradient.

[0070] It can be understood that is the simulation model gradient calculated using the automatic gradient function of Pytorch.

[0071] The calculated simulation model gradient is used as physical information to guide the training of the input convex neural network model, further improving the convergence and accuracy of the input convex neural network model. Thus, a non-linear convex model of the wind turbine can be obtained. Figure 3 is the prediction effect diagram of the main shaft torque (S spindle torque) of the non-linear convex model of the wind turbine. Figure 3In it, S spindle torque (kN) represents the spindle torque, with the unit of kilonewton (kN); the abscissa represents time, with the unit of second; the ordinate represents the spindle torque (Torque), with the unit of Newton - meters, which is used to describe the magnitude of the torsional force acting on the main shaft of the wind turbine. Torque is a measure of the rotational force and reflects the ability of the force to produce a rotational effect on an object. In a wind turbine, the spindle torque is one of the important operating parameters, which directly affects the operating efficiency and mechanical load of the wind generator. The vertical axis in the figure is marked as TS (N·m), showing the spindle torque values calculated by two methods, namely simulation (Simulation result) and (ICNN method), at different time points (the horizontal axis, with the unit of second). By comparing the results of these two methods, the accuracy and effectiveness of the non - linear convex model of the wind turbine in simulating or predicting the spindle torque of the wind turbine can be evaluated. Simulation result represents the prediction result of the non - linear convex model of the wind turbine, referring to the data obtained through the non - linear convex model of the wind turbine.

[0072] Figure 3 The curves of the spindle torque changing with time obtained by the two methods are shown. Among them, the red curve represents the prediction result of the non - linear convex model of the wind turbine, and the blue curve represents the result of the ICNN method. By comparing these two curves, the performance and accuracy of the two methods in simulating the spindle torque can be analyzed.

[0073] The input convex neural network model is used to fit the simulation data. In the constructed input convex neural network model, are all non - negative real numbers, and the convex non - decreasing function Smoothed relu is selected as the activation function , multiple straight - through layers are added and a bias term is added in . Through such a network structure, training is carried out based on the data set. The loss function includes minimizing the model error and minimizing the model gradient error. Finally, the non - linear convex model of the wind turbine is obtained, realizing the modeling of the mapping relationship between these parameters.

[0074] The time - window load time - series data is generated by fitting through the input convex neural network model. Based on the mapping relationship between the time - window load time - series data and the equivalent fatigue load, a non - linear convex model of the fatigue load is obtained, specifically including: Using a sliding window with a preset window length to slice the obtained spindle torque to divide it into a spindle torque test set and a spindle torque data set; in this embodiment, the window length is 10.

[0075] By using the rainflow counting method, calculate the load cycle amplitude and mean value of the main shaft torque for each sliding window to obtain the equivalent fatigue load. The calculation formula is as follows:

[0076] In the formula, is the load amplitude of the i-th level, is the number of cycles of the load at the level. m is the Wohler index of the S-N curve. N depends on the design life of the wind turbine. is the equivalent fatigue load of the N-th segment of the load signal; Use the input convex neural network model to fit the main shaft torque to generate the time window load time series data;

[0077] Based on the mapping relationship between the time window load time series data and the equivalent fatigue load, obtain the non-linear convex model of the fatigue load.

[0078] Use the input convex neural network model to fit the mapping relationship between the time window load time series data and the equivalent fatigue load. The input of this mapping relationship is the time window load time series data, and the output is the equivalent fatigue load. Construct the active power reduction optimization objective as:

[0079] Among them, is the equivalent fatigue load of the i-th wind turbine calculated according to the pre-trained model of the dual-input convex neural network, which is obtained from the lookup table constructed according to is the active power reduction optimization objective function, indicating minimizing the objective function, that is, minimizing the equivalent fatigue load of the whole field.

[0080] Construct the constraint conditions of the optimization problem as:

[0081] Among them, is the reference value of the active power, is the maximum available power of a single unit, is the minimum available power, is the equality constraint function, is the sum of the active power of the whole field, is the power grid dispatch instruction, is the inequality constraint function, is the change amount of the active power within 1 time step, is the maximum power ramp constraint of a single unit.​

[0082] The model predictive control algorithm is used to optimize and solve the pre-trained model of the dual-input convex neural network, and the optimal solution of the active power distribution that satisfies the safety and stability constraints of the wind farm is obtained.

[0083] In the present invention, through the pre-trained model of the dual-input convex neural network, the equivalent fatigue load of the wind turbine is modeled in the form of a convex function; and by generating a look-up table to accelerate the solution of the convex optimization problem, while satisfying the safety and stability constraints of the wind farm, the fatigue load of the entire wind farm is effectively suppressed.

[0084] The input convex neural network model can model the equivalent fatigue load of the main shaft torque of the wind turbine in the form of a convex function based on a sliding time window, transform the non-convex optimization problem into a convex optimization problem for solution, while maintaining the convexity of the model, and can also improve the convergence and accuracy of the model.

[0085] Taking the pre-trained model of the dual-input convex neural network as the prediction model, the model predictive control algorithm is used for optimization and solution, and the model predictive control algorithm is the sequential least squares quadratic programming method; Using the model predictive control algorithm to optimize and solve the pre-trained model of the dual-input convex neural network, the optimal solution of the active power distribution that satisfies the safety and stability constraints of the wind farm is obtained, including: According to the traditional proportional distribution algorithm, calculate the active power reference of each wind turbine in the wind farm as the initial point; Generate a look-up table of the wind turbine fatigue load according to the pre-trained model of the dual-input convex neural network at a scale of 10W; Use the sequential least squares quadratic programming method SLSQP to solve the optimization problem: At the k-th iteration, use the sequential least squares quadratic programming method to perform a first-order Taylor expansion (linear approximation) on the active power reduction optimization objective and the optimization problem constraint conditions: For the active power reduction optimization objective:

[0086] Among them, is the active power reduction optimization objective function of the active power reference value, is the first-order Taylor expansion of the active power reference value at the k-th iteration at , is the gradient vector of the active power reference value at the k-th iteration at , is the search direction.

[0087] For the optimization problem constraint conditions:

[0088]

[0089] In the formula, is the first-order Taylor expansion of the equality constraint function at during the k-th iteration, is the gradient vector of the equality constraint function at during the k-th iteration, is the first-order Taylor expansion of the inequality constraint function at during the k-th iteration, is the gradient vector of the inequality constraint function at during the k-th iteration.

[0090] The sequential least squares programming method transforms the original problem into a constrained quadratic programming problem by introducing Lagrange multipliers and penalty functions, and its expression is:

[0091] Among them, is the Hessian matrix of the objective function at estimated by the finite difference method; The search direction is obtained by solving the quadratic programming problem; The current point is updated using a fixed step size, and its expression is:

[0092] Among them, is the step size, which is taken as 1 here, is the update direction at the step, is the active power reference value updated after the step iteration; The multipliers λ and μ are updated according to the KKT conditions; After the iteration is completed, the optimal solution of the active power distribution that satisfies the safety and stability constraints of the wind farm is obtained and sent to each wind turbine for execution.

[0093] Compared with the traditional proportional distribution algorithm, the present invention can reduce the overall shaft fatigue load of a 10×5MW wind farm by 40% under the condition of satisfying the operation constraints of the wind farm. After accelerating the calculation using the look-up table, real-time solution can be achieved. Taking 1s as the control time step, the total solution time for 10 wind turbines is 0.83s.

[0094] The iteration is terminated when any of the following conditions is satisfied: The change in the objective function is less than the threshold, which is 1 here; The gradient norm is 0; The maximum number of iterations is reached, which is 100 here.

[0095] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0096] For those of ordinary skill in the art, according to the teachings of the present invention, designing different forms of control circuits does not require creative efforts. Without departing from the principles and spirit of the present invention, these changes, modifications, substitutions, and variations to the embodiments still fall within the protection scope of the present invention.

Claims

1. A wind farm active load reduction optimization method based on dual-input convex neural network, characterized in that: The method comprises: Build a simulation environment to obtain simulation data; The simulation data is fitted by using the input convex neural network model to obtain the nonlinear convex model of the wind turbine, and the main shaft torque is output through the nonlinear convex model of the wind turbine; The time window load time series data is generated by fitting the input convex neural network model, and the nonlinear convex model of fatigue load is obtained based on the mapping relationship between the time window load time series data and the equivalent fatigue load; The main shaft torque output by the nonlinear convex model of the wind turbine is used as the input of the nonlinear convex model of fatigue load to obtain a dual-input convex neural network pre-training model; The active power optimization control solution is solved for the dual-input convex neural network pre-training model to obtain the optimal solution for active power distribution that meets the safety and stability constraints of the wind farm.

2. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 1, characterized in that: Build a simulation environment to obtain simulation data, including: Construct a simulation environment to obtain the mapping relationship between the main shaft torque and the inflow wind speed, generator speed, output power, pitch angle, generator torque, and active power reference value of the wind farm, and obtain simulation data based on the mapping relationship; Mix the simulated data and get the data set With the test set ; in, Represents input data, Represents output data; input data includes: inflow wind speed, generator speed, output power, pitch angle, generator torque, active power reference value; output data includes: main shaft torque; The simulation environment includes: an aerodynamic simulation environment, a transmission system simulation environment, a generator simulation environment, a tower simulation environment, and a control system simulation environment.

3. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 2, characterized in that: The input convex neural network model is used to fit the simulation data to obtain the nonlinear convex model of the wind turbine. The expression of the input convex neural network model is: In the formula, About The convex function of is the output vector of the input convex neural network model, is the set of weights and bias parameters input to the convex neural network model, is the i+1th layer output vector of the input convex neural network model, is the weight of the i-th non-through layer of the input convex neural network model, is the i-th layer output of the input convex neural network model, is the weight of the i-th layer through the input convex neural network model, is the input vector of the convex neural network model, is the bias of the i-th layer of the input convex neural network model, is the set of pass-through layer weights input to the convex neural network model, is the set of non-through layer weights input to the convex neural network model, is the set of bias terms input to the convex neural network model, is the activation function; Adopting Dataset Train the input convex neural network model; The convergence and accuracy of the input convex neural network model are improved by the loss function, and the nonlinear convex model of the wind turbine is obtained. The expression of the loss function is: In the formula, is the loss function, is to minimize the model error, is to minimize the model gradient error, is the output vector of the improved input convex neural network model, is the input vector of the improved convex neural network model, is the gradient of the input convex neural network model, It is the simulation environment gradient calculated using the automatic gradient calculation tool.

4. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 3 is characterized in that: The time window load time series data is generated by fitting the input convex neural network model. Based on the mapping relationship between the time window load time series data and the equivalent fatigue load, the nonlinear convex model of the fatigue load is obtained, which specifically includes: Slice the acquired spindle torque using a sliding window of preset window length to divide it into a spindle torque test set and a spindle torque data set; The load cycle amplitude and mean value of the main shaft torque in each sliding window are calculated by the rain flow counting method to obtain the equivalent fatigue load. The calculation formula is: In the formula, is the load amplitude of the i-th level, For the The number of cycles of the level load, m is the Wohler index of the SN curve, is the equivalent fatigue load of the Nth segment load signal; The input convex neural network model is used to fit the main shaft torque and generate time window load time series data; Based on the mapping relationship between the time window load time series data and the equivalent fatigue load, a nonlinear convex model of fatigue load is obtained.

5. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 4, characterized in that: The active power optimization control solution of the dual-input convex neural network pre-training model is solved to obtain the optimal solution of active power distribution that meets the safety and stability constraints of the wind farm, including: The optimization target of active load shedding is: in, is the equivalent fatigue load of the i-th wind turbine calculated according to the dual-input convex neural network pre-training model, which is obtained by The constructed lookup table is obtained, is the active load shedding optimization objective function; The constraints for constructing the optimization problem are: in, is the active power reference value, is the maximum available power of a single machine, is the minimum available power, is the equality constraint function, is the sum of the active power of the entire field, is the grid dispatch instruction, is the inequality constraint function, is the change in active power within 1 time step, is the maximum power climbing constraint of a single machine; The model predictive control algorithm is used to optimize and solve the dual-input convex neural network pre-training model to obtain the optimal solution for active power distribution that meets the safety and stability constraints of the wind farm.

6. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 5, characterized in that: The model predictive control algorithm is the sequential least squares programming method; The model predictive control algorithm is used to optimize and solve the dual-input convex neural network pre-training model to obtain the optimal solution for active power distribution that meets the safety and stability constraints of the wind farm, including: According to the traditional proportional allocation algorithm, the active power reference of each wind turbine in the wind farm is calculated as the initial point; At the 10W scale, a wind turbine fatigue load lookup table is generated based on a dual-input convex neural network pre-trained model; Solve the optimization problem using sequential least squares programming: At the k-th iteration, the sequential least squares programming method is used to perform a first-order Taylor expansion of the active load reduction optimization objective and the optimization problem constraints: For active load shedding optimization objectives: in, is the active load shedding optimization objective function for the active power reference value, For the active power reference value, the kth iteration is A first-order Taylor expansion is made at is the active power reference value at the kth iteration The gradient vector at is the search direction; For the optimization problem constraints: In the formula, When the kth iteration of the equality constraint function is A first-order Taylor expansion is made at is the equality constraint function at the kth iteration The gradient vector at When the kth iteration of the inequality constraint function is A first-order Taylor expansion is made at is the inequality constraint function at the kth iteration The gradient vector at ; The sequential least squares programming method transforms the original problem into a constrained quadratic programming problem by introducing Lagrange multipliers and penalty functions. Its expression is: in, The objective function is The Hessian matrix at ; The search direction is obtained by solving the quadratic programming problem ; Use a fixed step size to update the current point, and its expression is: in, is the step length, here it is 1, It is The update direction of the step, It is The active power reference value updated after the step iteration; After completing the iteration, the optimal solution for active power allocation that meets the safety and stability constraints of the wind farm is obtained; The iteration is terminated when any of the following conditions are met: The objective function change is less than the threshold; The gradient norm is 0; The maximum number of iterations has been reached.

7. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 6, characterized in that: Build an aerodynamic simulation environment, including: Calculate the aerodynamic torque and thrust with the following formula: in, is the pneumatic torque, For thrust, is the power coefficient, is the thrust coefficient, is the blade length, is the air density, is the effective wind speed on the wind turbine blades, is the tip speed ratio, which is defined as ; Build a transmission system simulation environment, including: Using a single mass block model, the wind rotor mass and the generator mass are combined into an equivalent mass, and the formula is as follows: According to the low-speed shaft motion equation, the speed relationship between the rotor and the generator is calculated as follows: in, is the equivalent mass, is the mass of the wind wheel, is the generator mass, is the gearbox transmission ratio, is the generator torque, For the rotor, is the speed of the generator.

8. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 7, characterized in that: Build a generator simulation environment, including: In the torque control loop, vector control with millisecond-level high-precision response is adopted, ignoring the dynamic characteristics, that is: in, is the generator reference torque.

9. The method for optimizing active load shedding in a wind farm based on a dual-input convex neural network according to claim 8, characterized in that: Build a tower simulation environment, including: The calculation formula of the tower front and rear bending moment is: in, represents the fore-aft bending moment of the tower, Indicates the tower height.

10. The wind farm active load reduction optimization method based on dual-input convex neural network according to claim 9, characterized in that: Build a control system simulation environment, including: The wind speed Below rated wind speed When , it is defined as the first mode; The wind speed Greater than rated wind speed When , it is defined as the second mode; When operating in the first mode, pitch control is stopped. ; When running in the second mode, the measured generator speed is filtered through a low-pass filter, and the filtered speed is: in is the speed after filtering, is the time constant of the filter, is the generator speed, is the Laplace operator; Based on the filtered speed Rated value of speed The deviation of the blade pitch angle is calculated, and the expression is: in, is the reference value of the pitch angle, is the proportional gain of the PI controller, is the integral gain of the PI controller, is the correction factor, ,in, Hewei constant.

Images