Feed-forward control generating capacity increasing method for wind generating set

Abstract

The invention discloses a wind generating set feed-forward control generating capacity increasing method, which relates to the technical field of wind generating set control, and comprises the following steps: carrying out fusion processing on obtained target equipment observation data through an extended Kalman filtering algorithm to obtain a predicted wind speed, and correcting the predicted wind speed; according to the corrected predicted wind speed, generating a propeller pitch correction through a feed-forward mapping function; calculating a dynamic feed-forward control weight according to the predicted wind speed confidence coefficient and the real-time operation state of the target equipment, and combining the dynamic feed-forward control weight with the propeller pitch correction to generate a control instruction; obtaining operation data after the target device executes the control instruction, constructing a target reward function, and dynamically regulating and controlling parameters of the feedforward mapping function through a DDPG algorithm; according to the method, the contradictory problem of generating capacity improvement and load suppression is solved through propeller pitch safety optimization control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind turbine control, and more particularly, to a wind turbine feedforward control power generation improvement method. BACKGROUND

[0002] As an important way of renewable energy utilization, wind power generation has been rapidly developing worldwide in recent years. However, due to the randomness, intermittency and difficulty in accurate prediction of wind energy, wind turbine units often face problems such as significant fluctuation of power generation, low energy capture efficiency, and fatigue load on blades and transmission chain components during actual operation. Especially in complex wind field environment, the incoming flow wind speed and turbulence intensity show high dynamic changes. If the pitch and power angle of the unit control system cannot be adjusted in time and accurately, the unit output will be unstable, the power generation will decrease or the load will increase, thereby affecting the performance and service life of the unit. Therefore, how to improve the wind energy capture efficiency and the stability of the unit power output while meeting the structural safety requirements of the unit has become a key research problem in the field of wind power generation technology.

[0003] For example, the patent for invention with publication number CN113309661B discloses a method for improving the power generation of a wind turbine unit. The method obtains the wind direction deviation interval in which the optimal power curve of the unit is located through data analysis, selects the middle value of the interval as the re-correction angle of the nacelle wind direction marker based on the initial zero correction, and re-corrects the nacelle wind direction marker to optimize the power curve of the unit and thereby improve the power generation of a single unit or even the entire field of units, and increase the direct economic benefits of the wind farm.

[0004] For example, the patent for invention with publication number CN114517763A discloses a control method and system for improving the power generation of a large-capacity variable-speed wind turbine unit. The method includes: 1) reading sensor wind speed information; 2) calculating the minimum speed limit; 3) limiting the calculated minimum speed limit ωMin; and 4) modifying the actual minimum speed limit parameter. Without the need for additional hardware or sensor equipment, the method can improve power generation and achieve higher profits without additional costs.

[0005] The above-mentioned technical solutions at least have the following technical problems: In the traditional technology, the pitch adjustment amount is generated based on the predicted wind speed, without fully considering the blade load, unit vibration and historical operation rules, which can easily lead to over-regulation, imbalance between power generation and load, and the use of fixed proportion or static front weight in existing control strategies, which cannot be adjusted in real time according to the actual deviation degree of the unit and the prediction confidence, resulting in a decline in control performance under sudden changes in wind conditions or abnormal working conditions. To solve the above problems, the present application provides a solution. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiment of the present application provides a wind turbine front feed control power generation capacity improvement method, which realizes safe optimization control of the pitch by fine wind speed prediction and front feed adaptive regulation, and solves the contradiction problem between power generation capacity improvement and load suppression.

[0007] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions: A wind turbine front feed control power generation capacity improvement method, comprising: fusing the obtained target device observation data by an extended Kalman filtering algorithm to obtain a predicted wind speed, and correcting the predicted wind speed; generating a pitch correction amount by a front feed mapping function according to the corrected predicted wind speed; calculating a dynamic front feed control weight according to the predicted wind speed confidence and the real-time running state of the target device, and combining the pitch correction amount to generate a control instruction; obtaining the running data of the target device after executing the control instruction, constructing a target reward function, and dynamically regulating and controlling the parameters of the front feed mapping function by a DDPG algorithm.

[0008] In a preferred embodiment, the predicted wind speed is obtained by fusing the obtained target device observation data by an extended Kalman filtering algorithm, and the specific steps are as follows: obtaining the observation data of the target device, and unifying the time stamp and matching the sampling rate of the observation data to obtain a synchronous signal set; wavelet decomposing and low-pass filtering the synchronous observation data set to form filtered observation data; establishing a nonlinear state space model according to the filtered observation data set, wherein the nonlinear state space model includes a state equation and an observation equation; initializing the key parameters of the extended Kalman filter according to the state equation and the observation equation, wherein the key parameters include state estimation, state covariance, process noise covariance matrix and observation noise covariance matrix; predicting the state at the next moment by a state transition function according to the state estimation and the state covariance at the previous moment combined with the control input vector and the system parameters; outputting the state prediction covariance matrix according to the predicted state and the state covariance at the previous moment; calculating the predicted observation value by the observation equation according to the predicted state, and comparing it with the initial observation value to obtain the observation residual; optimizing the observation residual by a Gaussian process regression model to obtain the corrected predicted wind speed.

[0009] In a preferred embodiment, the observation residual is optimized by the Gaussian process regression model to obtain the corrected predicted wind speed, specifically as follows: the observation Jacobian matrix is obtained by taking partial derivative of the state vector in the observation equation according to the predicted state, and the Kalman gain is output according to the observation Jacobian matrix and the state prediction covariance matrix; the state estimation and the state prediction covariance matrix are updated according to the Kalman gain and the observation residual; the initial residual correction amount is obtained by the Gaussian process regression model taking the observation residual and the updated state prediction covariance matrix as input, and the final residual correction amount is obtained according to the weight distribution of the observation residual; the initial predicted wind speed is obtained by the wind speed solving function taking the updated state estimation as input, and the initial predicted wind speed is corrected according to the residual correction amount to obtain the corrected predicted wind speed.

[0010] In a preferred embodiment, the pitch correction amount is generated by the feedforward mapping function according to the corrected predicted wind speed, specifically as follows: the state prediction covariance matrix and the observation residual statistics are mapped to a fixed interval to obtain the predicted wind speed confidence, and the confidence adjustment coefficient is obtained by converting the predicted wind speed confidence by a linear function; a plurality of candidate pitch correction amounts are generated by mapping the corrected predicted wind speed to the power of the target device, and the uncertainty of each candidate pitch correction amount is estimated according to the state prediction covariance matrix; according to the current wind condition and the unit state, the historical similar operation segment is retrieved, the corresponding power change amplitude and blade load of each candidate pitch correction amount in the similar segment are statistically analyzed, and multi-objective Pareto screening is performed on the candidate pitch correction amount to obtain a non-dominated candidate set; the non-dominated candidate set is subjected to explainability analysis, and after the explainability analysis is completed, the non-dominated candidate set is taken as input, the candidate pitch correction amount is subjected to amplitude weighting according to the confidence adjustment coefficient and the uncertainty to obtain the confidence weighted pitch correction amount, and the confidence weighted pitch correction amount is subjected to amplitude limitation and rate limitation; the confidence weighted pitch correction amount is split into collective pitch and differential pitch components, and the influence intensity of the collective pitch and differential pitch components on the blade cycle load spectrum is evaluated; the influence intensity of the differential pitch component on the blade cycle load spectrum is adjusted according to the influence intensity; the candidate pitch correction amount after the distribution adjustment is input into the digital twin model for short-term virtual simulation, and according to the virtual simulation result, the safe candidate set is retained; according to the preset strategy, the final pitch correction amount is selected from the candidate pitch correction amount after the virtual verification and adjustment.

[0011] In a preferred embodiment, the step of calculating dynamic feedforward control weights based on the predicted wind speed confidence level and the real-time operating status of the target equipment, and combining them with the pitch correction to generate control commands, is as follows: The real-time operating status of the target equipment is obtained, and the unit deviation is output based on the Mahalanobis distance method; the pitch correction, predicted wind speed confidence level, and unit deviation are input into a nonlinear mapping function to generate dynamic feedforward control weights, and state-sensitive constraints are applied to the dynamic feedforward control weights; the dynamic feedforward control weights are combined with the feedforward pitch correction, and the contribution of the feedforward in the final command is controlled through a weighted method to generate the final control command.

[0012] In a preferred embodiment, the steps of acquiring the operational data of the target device after executing control commands, constructing a target reward function, and adjusting the parameters of the feedforward mapping function using the DDPG algorithm are as follows: receiving real-time operational data of the target device after executing control commands, setting the target reward function based on the real-time operational data; generating a set of candidate parameter actions based on the feedforward mapping function, and performing short-term virtual simulation on the set of candidate parameter actions to obtain a safe candidate set; inputting the safe candidate set into the DDPG algorithm for policy training, and extracting historical data related to the current state from the experience replay pool; combining the extracted historical data with the current safe candidate set and inputting it into the value network, whereby the value network learns from the historical data to predict the expected reward of each candidate parameter in the current state, thereby evaluating the candidate parameters; updating the policy network based on the evaluation results of the value network; applying the optimal parameters output by the policy network to the feedforward mapping function to achieve the generation of the next cycle's pitch, while storing the new state, action, and reward data into the experience replay pool.

[0013] In a preferred embodiment, the step of generating a set of candidate parameter actions based on the feedforward mapping function is as follows: obtaining the current system state vector, inputting the current state vector into the feedforward mapping function, calculating preliminary candidate control parameter actions, and generating a preliminary set of candidate parameter actions within a preset range to cover the adjustment space; filtering the set of candidate actions according to system constraints, eliminating actions that do not meet the constraints, and obtaining the final set of feasible candidate parameter actions.

[0014] The technical effects and advantages of the method for improving power generation of wind turbine generator sets using feedforward control according to the present invention are as follows: 1. This invention introduces a feedforward mapping function to generate pitch correction values, and emphasizes interpretability and safety during candidate scheme formation, screening, and confidence weighting. The system first generates multiple candidate pitch correction values ​​based on the corrected predicted wind speed, and calculates the prediction uncertainty using state covariance and observation residual information. Subsequently, through historical similar operation segment retrieval and multi-objective Pareto screening, it ensures that candidate schemes balance power generation improvement and blade load constraints. Furthermore, this invention enables the control strategy to directly optimize the load spectrum distribution by decomposing the aggregate pitch and differential pitch components and assessing the influence intensity of cyclic loads. Finally, digital twin simulation is introduced to verify the short-term execution safety of candidate schemes, avoiding potential over-adjustment, vibration, or fatigue problems in traditional feedforward control from the source.

[0015] 2. This invention constructs a dynamic feedforward control weighting mechanism, enabling control commands to adaptively adjust according to the operating state. By introducing the Mahalanobis distance to the generator set deviation and combining it with the predicted wind speed confidence level, these factors, along with the pitch correction, are input into the nonlinear mapping solution weights, achieving real-time sensitive response of feedforward control to changes in generator operating conditions. This mechanism avoids output instability caused by the constant superposition of feedforward quantities under unsteady wind conditions, giving the control commands smoothness, controllability, and time-varying adaptability. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the process for a method to improve the power generation of a wind turbine generator set using feedforward control, according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1, Figure 1 This invention provides a method for improving the power generation of a wind turbine generator set through feedforward control, comprising: S1, the acquired target equipment observation data is fused and processed by the extended Kalman filter algorithm to obtain the predicted wind speed, and the predicted wind speed is corrected. In this embodiment, the acquired target device observation data is fused using the extended Kalman filter algorithm to obtain the predicted wind speed, and the predicted wind speed is then corrected, as follows: The observation data of the target equipment is acquired, the timestamp of the observation data is unified and the sampling rate is matched to obtain a synchronization signal set, and the time delay between each observation data is calculated by phase estimation. The signal is time-compensated to eliminate the deviation caused by the sampling time difference of different observation data, forming a synchronous observation dataset. The observation data includes nacelle LIDAR wind speed, nacelle anemometer, blade root strain, blade pitch angle feedback and generator speed. Wavelet decomposition and low-pass filtering are performed on the synchronous observation dataset to extract low-frequency trend components and filter out high-frequency noise, resulting in filtered observation data. A nonlinear state-space model is established based on the filtered observation dataset. The nonlinear state-space model includes state equations. and observation equations ,in, This represents the system state vector at time step k, composed of observed data. It is the control input vector, the control quantity at time step k. It is a preset process noise. These are system parameters, including but not limited to blade lift / drag coefficients. It is the observation vector at time step k. It is a nonlinear observation equation that maps the system state to observable outputs. It is a preset observation noise, describing the errors and noise during the measurement process. It is a nonlinear state transition function; Based on the state equation and the observation equation, the key parameters of the extended Kalman filter (EKF) are initialized. The key parameters include the state estimate, state covariance, process noise covariance matrix, and observation noise covariance matrix. Based on the state estimate and state covariance from the previous moment, combined with the control input vector and system parameters Predict the state at the next time step using the state transition function. ; Based on the predicted state and the state covariance of the previous time step, output the state prediction covariance matrix. ,in, Is the state equation in The transpose of the Jacobian matrix at that point. It is the process noise covariance; The predicted observations are calculated using the observation equation based on the predicted state. and compared with the initial observations Compare and calculate the observed residuals ; The observation Jacobian matrix is ​​obtained by taking the partial derivative of the predicted state with respect to the state vector in the observation equation. The Kalman gain is then calculated based on the observation Jacobian matrix and the state prediction covariance matrix. Update the state estimate based on Kalman gain and observation residuals. and state prediction covariance matrix ,in, It is Kalman gain. It is the observed Jacobian matrix. It is an identity matrix with the same dimensions as the state prediction covariance matrix; The observation residuals and the updated state prediction covariance matrix are used as inputs. The initial residual correction is obtained by using a Gaussian process regression model. The final residual correction is obtained by weighting the observation residuals. The updated state estimate is used to obtain the initial predicted wind speed through the wind speed solution function, and the initial predicted wind speed is corrected according to the residual correction amount to obtain the corrected predicted wind speed. The wind speed solution function is a wind speed inversion function derived from the aerodynamic model of the wind turbine generator. This function maps the updated system state estimate to the corresponding effective wind speed and is a commonly used wind speed inversion model in the existing wind turbine aerodynamic modeling.

[0019] The Kalman gain calculation formula is as follows: In the formula: It is Kalman gain. It is the state prediction covariance matrix. It is the transpose of the observed Jacobian matrix. It is the observation noise covariance matrix.

[0020] S2, Based on the corrected predicted wind speed, the pitch correction amount is generated through the feedforward mapping function; In this embodiment, the pitch correction is generated based on the corrected predicted wind speed using a feedforward mapping function, as follows: By using causal analysis, the time lag of the effect of predicted wind speed on changes in unit speed, output power and load is identified, which is used to determine the time alignment of feedforward control. The state prediction covariance matrix and observation residual statistics are mapped to the [0,1] interval to obtain the prediction wind speed confidence. The prediction wind speed confidence is then transformed by a linear function to obtain the confidence adjustment coefficient, which is used for weighting the amplitude of subsequent candidate pitch corrections. The corrected predicted wind speed is mapped to the power of the target equipment to generate several candidate pitch corrections, and the uncertainty of each candidate pitch correction is estimated based on the state prediction covariance matrix. Based on the current wind conditions and unit status, historical similar operating segments are retrieved. Statistical analysis is performed on the power change amplitude and blade load corresponding to each candidate pitch correction in the similar segments. Multi-objective Pareto screening (considering power fluctuations, load change amplitude and uncertainty) is then performed on the candidate pitch corrections to obtain a non-dominated candidate set. Interpretability analysis is performed on the non-dominated candidate set. By performing feature contribution analysis on the non-dominated candidate set, the contribution of each candidate pitch correction to the uncertainty of output power fluctuation, predicted wind speed confidence and blade load is evaluated, so as to ensure that the candidate selection is traceable and engineering interpretable. After completing the interpretability analysis, the non-dominated candidate set is used as input. The candidate pitch corrections are weighted by magnitude according to the confidence adjustment coefficient and uncertainty to obtain the confidence-weighted pitch correction. The confidence-weighted pitch correction is then subject to amplitude and rate constraints. The confidence-weighted pitch correction is decomposed into aggregate pitch and differential pitch components, and the influence of aggregate pitch and differential pitch components on the blade periodic load spectrum is evaluated. If the differential pitch component has a significant impact on the cyclic load, then the distribution is adjusted to transfer some components to the combined pitch. The candidate pitch correction values ​​after allocation and adjustment are input into the digital twin model for short-time virtual simulation to simulate its instantaneous response to unit speed, output power and blade load. Based on the virtual simulation results, candidate pitch correction values ​​that may cause blade load or unit speed to exceed the limit are eliminated, and a safe and feasible candidate set is retained. Among the candidate pitch correction values ​​that have been verified and adjusted through virtual testing, the final pitch correction value is selected according to a preset strategy (such as prioritizing blade load safety or power stability).

[0021] In this embodiment, the confidence-weighted pitch correction is decomposed into aggregate pitch and differential pitch components, as follows: Obtain the confidence-weighted pitch correction and the actual pitch angle of each blade, and average the confidence-weighted pitch correction for each blade to obtain the aggregate pitch components. Subtract the confidence-weighted correction for each blade from the aggregate pitch component to obtain the differential pitch component. Based on the observed noise covariance of the blades, the confidence weight of each blade is calculated, and the differential pitch component is multiplied by the confidence weight to obtain the weighted differential pitch. The combined pitch components are added to the original pitch angle of each blade to obtain the initial control pitch. The weighted differential pitch is superimposed on the initial control pitch to complete the balance correction between blades and obtain the final control pitch.

[0022] S3 calculates the dynamic feedforward control weights based on the predicted wind speed confidence level and the real-time operating status of the target equipment, and combines them with the pitch correction amount to generate control commands. In this embodiment, dynamic feedforward control weights are calculated based on the predicted wind speed confidence level and the real-time operating status of the target equipment. These weights are then combined with the pitch correction amount to generate control commands, as follows: The real-time operating status of the target equipment is obtained, and the deviation of the unit is output based on the Mahalanobis distance method. The real-time operating status includes the rotational speed, output power, blade load, and historical rate of change of pitch. The pitch correction, predicted wind speed confidence, and unit deviation are input into a nonlinear mapping function to generate dynamic feedforward control weights, which are used to adjust the proportion of feedforward contribution in the final control command. State-sensitive constraints are applied to the dynamic feedforward control weights to ensure that the feedforward control is within a safe range, including: limiting the feedforward contribution when the blade load change rate is too high, limiting the feedforward contribution when the pitch change rate is close to the limit, and automatically attenuating the feedforward contribution when the output power or speed is close to the safe limit. Based on historical control data and the predicted residuals and execution errors obtained from rolling window statistics, the parameters of the nonlinear mapping function are dynamically updated to achieve dynamic adaptive adjustment of feedforward weights. By combining the dynamic feedforward control weights with the feedforward pitch correction, the contribution of the feedforward in the final command is controlled in a weighted manner to generate the final control command. The final pitch control command is sent to the wind turbine actuator, and the control input and equipment output status, including output power, speed, blade load and pitch response, are recorded for the next cycle dynamic feedforward weight calculation and adaptive optimization of the mapping function.

[0023] The formula for calculating the deviation of the unit is as follows: In the formula: It is the deviation of the generator set. It is the transpose of the state residual vector, defined as the difference between the real-time state and the desired state of the wind turbine. It is the inverse of the state vector covariance matrix, used to describe the variance and correlation between the state variables.

[0024] S4. Obtain the running data of the target device after executing the control command, construct the target reward function, and dynamically adjust the parameters of the feedforward mapping function through the DDPG algorithm; In this embodiment, the operating data of the target device after executing control commands is obtained, a target reward function is constructed, and the parameters of the feedforward mapping function are adjusted using the DDPG algorithm, as follows: Receive real-time operating data after the target equipment executes control commands, including output power, rotational speed, blade load, pitch response, and ambient wind speed; The target reward function is set based on real-time operating data, including: the square of the deviation of the rotational speed from the target value, the square of the deviation of the blade load from the target value, the square of the pitch change rate, and the short-term power fluctuation value, which are used to evaluate the control effect of each candidate parameter. A set of candidate parameter actions is generated based on the feedforward mapping function, and a short-time virtual simulation is performed on the set of candidate parameter actions to evaluate the impact of each candidate parameter on output power, speed and blade load. Candidate actions that cause exceedances are eliminated to obtain a safe candidate set. The set of safety candidates is input into the DDPG algorithm for policy training. Historical data with the current state is extracted from the experience replay pool, including the historical state, the corresponding candidate parameter action, the reward after execution, and the next state, which are used to evaluate the control effect of the candidate parameters in similar states. The extracted historical data is combined with the current set of safety candidates and input into the value network. The value network learns from the historical data to predict the control effect or expected reward that each candidate parameter may obtain in the current state, thereby evaluating the performance of the candidate parameters. Based on the evaluation results of the value network, the policy network is updated so that it can select the candidate parameters that perform best in the current state, that is, select the parameters that can achieve the maximum expected control effect in the current device state. The optimal parameters output by the policy network are applied to the feedforward mapping function to generate the paddle pitch for the next cycle. At the same time, the new state, action and reward data are stored in the experience replay pool to provide input for the next round of training, thus realizing the closed-loop adaptive optimization of the policy.

[0025] The process of generating the candidate parameter action set based on the feedforward mapping function is as follows: Obtain the current system state vector, including filtered observation data, pitch angle, blade strain, etc. The current state vector is input into the feedforward mapping function to calculate the preliminary candidate control parameter action, and a preliminary candidate parameter action set is generated within the allowable range to cover the adjustment space. The candidate action set is filtered based on system constraints (such as maximum and minimum pitch angle, blade load limits, etc.), and actions that do not meet the constraints are eliminated to obtain the final feasible candidate parameter action set.

[0026] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0027] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0028] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0029] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0031] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for increasing the power generation of a wind turbine generator set through feedforward control, characterized in that, include: The acquired observation data of the target equipment is fused using the extended Kalman filter algorithm to obtain the predicted wind speed, and the predicted wind speed is then corrected. Based on the corrected predicted wind speed, the pitch correction is generated through a feedforward mapping function. Based on the predicted wind speed confidence level and the real-time operating status of the target equipment, the dynamic feedforward control weight is calculated and combined with the pitch correction to generate control commands. The system acquires the operational data of the target device after executing control commands, constructs the target reward function, and dynamically adjusts the parameters of the feedforward mapping function using the DDPG algorithm.

2. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 1, characterized in that, The method involves fusing the acquired target device observation data using the extended Kalman filter algorithm to obtain the predicted wind speed, and then correcting the predicted wind speed, as detailed below: Acquire observation data from the target device, and unify the timestamps and sampling rates of the observation data to obtain a synchronization signal set; Wavelet decomposition and low-pass filtering are performed on the synchronous observation dataset to form filtered observation data; A nonlinear state-space model is established based on the filtered observation dataset. The nonlinear state-space model includes state equations and observation equations. Based on the state equation and the observation equation, the key parameters of the extended Kalman filter are initialized, including the state estimate, state covariance, process noise covariance matrix, and observation noise covariance matrix. Based on the state estimate and state covariance of the previous moment, combined with the control input vector and system parameters, the state of the next moment is predicted through the state transition function. Based on the predicted state and the state covariance of the previous time step, output the state prediction covariance matrix; The predicted observations are calculated using the observation equation based on the predicted state, and then compared with the initial observations to obtain the observation residuals. The predicted wind speed is obtained by optimizing the observation residuals using a Gaussian process regression model.

3. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 2, characterized in that, The corrected predicted wind speed is obtained by optimizing the observed residuals using a Gaussian process regression model, as detailed below: The observation Jacobian matrix is ​​obtained by taking the partial derivative of the predicted state with respect to the state vector in the observation equation. The Kalman gain is then output based on the observation Jacobian matrix and the state prediction covariance matrix. Update the state estimate and state prediction covariance matrices based on the Kalman gain and observation residuals; The observed residuals and the updated state prediction covariance matrix are used as inputs. The initial residual correction is obtained through a Gaussian process regression model. The final residual correction is obtained by weighting the observed residuals. The updated state estimate is used to obtain the initial predicted wind speed through the wind speed calculation function, and the initial predicted wind speed is corrected according to the residual correction amount to obtain the corrected predicted wind speed.

4. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 3, characterized in that, The process of generating a pitch correction amount based on the corrected predicted wind speed using a feedforward mapping function is as follows: The state prediction covariance matrix and observation residual statistics are mapped to a fixed interval to obtain the prediction wind speed confidence level. The prediction wind speed confidence level is then transformed by a linear function to obtain the confidence level adjustment coefficient. The corrected predicted wind speed is mapped to the power of the target equipment to generate several candidate pitch corrections, and the uncertainty of each candidate pitch correction is estimated based on the state prediction covariance matrix. Based on the current wind conditions and unit status, historical similar operating segments are retrieved, and statistical analysis is performed on the power change amplitude and blade load corresponding to each candidate pitch correction in the similar segments. Multi-objective Pareto screening is then performed on the candidate pitch corrections to obtain a non-dominated candidate set. An interpretability analysis is performed on the non-dominated candidate set. After the interpretability analysis is completed, the non-dominated candidate set is used as input. The candidate pitch correction is weighted by amplitude according to the confidence adjustment coefficient and uncertainty to obtain the confidence-weighted pitch correction. The confidence-weighted pitch correction is then subject to amplitude and rate constraints. The confidence-weighted pitch correction is decomposed into aggregate pitch and differential pitch components, and the influence of aggregate pitch and differential pitch components on the blade periodic load spectrum is evaluated. The distribution and adjustment are based on the intensity of the influence of the differential pitch component on the blade periodic load spectrum; The adjusted candidate pitch correction values ​​are input into the digital twin model for short-term virtual simulation. Based on the virtual simulation results, a safe candidate set is retained. From the candidate pitch correction values ​​that have been virtually verified and adjusted, the final pitch correction value is selected according to a preset strategy.

5. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 4, characterized in that, The dynamic feedforward control weights are calculated based on the predicted wind speed confidence level and the real-time operating status of the target equipment, and then combined with the pitch correction to generate control commands, as detailed below: Obtain the real-time operating status of the target equipment and output the unit deviation based on the Mahalanobis distance method; The pitch correction, predicted wind speed confidence, and unit deviation are input into a nonlinear mapping function to generate dynamic feedforward control weights, and state-sensitive constraints are applied to the dynamic feedforward control weights. By combining the dynamic feedforward control weights with the feedforward pitch correction, the contribution of the feedforward quantity to the final command is controlled in a weighted manner, thereby generating the final control command.

6. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 5, characterized in that, The process involves acquiring the operational data of the target device after executing control commands, constructing a target reward function, and adjusting the parameters of the feedforward mapping function using the DDPG algorithm, as detailed below: Receive real-time operational data after the target device executes control commands, and set the target reward function based on the real-time operational data; A set of candidate parameter actions is generated based on the feedforward mapping function, and a short-time virtual simulation is performed on the set of candidate parameter actions to obtain a safe candidate set. The set of safe candidates is input into the DDPG algorithm for policy training, and historical data related to the current state is extracted from the experience replay pool. The extracted historical data is combined with the current set of safe candidates and input into the value network. The value network learns from the historical data to predict the expected reward of each candidate parameter in the current state, thereby evaluating the candidate parameters. Update the policy network based on the evaluation results of the value network; The optimal parameters output by the policy network are applied to the feedforward mapping function to generate the pitch for the next cycle, while the new state, action and reward data are stored in the experience replay pool.

7. The method for increasing power generation of a wind turbine generator set via feedforward control according to claim 6, characterized in that, The process of generating the candidate parameter action set based on the feedforward mapping function is as follows: Obtain the current system state vector, input the current state vector into the feedforward mapping function, calculate the preliminary candidate control parameter actions, and generate a preliminary candidate parameter action set within a preset range to cover the adjustment space; The candidate action set is filtered according to the system constraints, and actions that do not meet the constraints are eliminated to obtain the final feasible candidate parameter action set.

Images