Disturbance adaptive compensation-based rapid frequency modulation method for wind turbine generator

Abstract

The invention relates to the technical field of devices for adjusting, controlling or stabilizing power or frequency in a power grid, in particular to a rapid frequency modulation method for a wind turbine generator based on disturbance adaptive compensation, which comprises the following steps of: constructing a dynamic disturbance sensing data set by collecting frequency deviation, a frequency deviation change rate and fan state parameters; a window length and a weight coefficient of a short-time window sliding mean algorithm are dynamically optimized by adopting a genetic algorithm, background noise interference of a power grid is suppressed, and high-precision disturbance characteristics are extracted. And reducing frequency interference of inertia identification by using a reverse test signal. And updating boundary layer parameters of the sliding mode control model through an online incremental learning algorithm, dynamically adjusting the output priority based on the disturbance energy entropy, and eliminating power conflicts. A multi-island genetic algorithm is introduced to optimize a mode switching threshold value, and frequency modulation safe exit is realized in combination with adaptive power ramp rate limitation. According to the method, the frequency response speed and the multi-source cooperation efficiency in the dynamic multi-disturbance and inertia time-varying scene are remarkably improved, and the frequency secondary drop risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of devices for adjusting, controlling or stabilizing power or frequency in a power grid, and particularly relates to a fast frequency modulation method for a wind turbine based on disturbance adaptive compensation. Background Art

[0002] With the access of a high proportion of new energy and distributed power sources to the power grid, the power system faces complex operation scenarios of multi-disturbance superposition and inertia dynamic change. Most of the existing frequency modulation methods for wind turbines are based on fixed control logics under single disturbances, and it is difficult to adjust the frequency modulation strategy in real time in a dynamic disturbance environment. Especially under continuous power disturbances or inertia time-varying conditions, traditional methods cannot quickly sense the change of system inertia and the disturbance superposition effect, resulting in a lag in frequency modulation response and exacerbating frequency fluctuations. At the same time, there is a lack of a dynamic coordination mechanism among multi-source frequency modulation devices, which easily leads to redundancy or insufficiency of frequency modulation power, causing an imbalance in resource allocation and further affecting the frequency stability of the system. There is an urgent need for a control method for wind turbines that can adapt to dynamic multi-disturbance scenarios and coordinate frequency modulation resources in real time to improve the frequency response accuracy and robustness in a complex power grid environment. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention provides a fast frequency modulation method for a wind turbine based on disturbance adaptive compensation, which solves the problem of the decline in the frequency stability of the power system caused by the lag in the frequency modulation response of the wind turbine and the insufficient coordination of multi-source frequency modulation resources under dynamic multi-disturbance and inertia time-varying conditions.

[0004] To solve the above technical problems, the specific technical solutions of the present invention are as follows: A fast frequency modulation method for a wind turbine based on disturbance adaptive compensation provided by the present invention includes: Step S1, collecting the frequency deviation, the rate of change of frequency deviation and the operating state parameters of the fan of the power system, and constructing a dynamic disturbance perception data set, where the operating state parameters of the fan include real-time monitoring data of the rotor kinetic energy reserve of the wind turbine; Step S2, based on the dynamic disturbance perception data set, dynamically optimizing the window length and weight coefficient of the short-time window sliding mean algorithm through a genetic algorithm to generate disturbance feature data after noise suppression; Step S3, generating a reverse test signal according to the disturbance feature data, and combining the recursive least squares method to estimate the equivalent inertia and damping parameters of the power grid in real time; Step S4, using an online incremental learning algorithm to perform time series modeling on continuous disturbance events in the dynamic disturbance perception data set, and updating the boundary layer thickness and switching gain parameters of a preset adaptive sliding mode control model; Step S5: Based on the equivalent inertia and damping parameters of the power grid, generate a compensation power command through a preset fuzzy adaptive ratio and a preset integral controller. The compensation power command includes the triggering conditions for the fast power support mode and the droop control mode; Step S6: According to the compensation power command, coordinate the allocation of the rotor kinetic energy reserve of the wind turbine generator and the power response of the multi-source frequency modulation equipment of the power grid through a preset microservice architecture, where the multi-source frequency modulation equipment of the power grid includes a virtual synchronous machine and an energy storage device; Step S7: Introduce the genetic algorithm into the adaptive sliding mode control model predictive control to optimize the triggering threshold of dynamic switching, and switch between the fast power support mode and the droop control mode according to the frequency deviation change rate; Step S8: When the frequency deviation returns to the preset dead zone range, exit the frequency modulation through the adaptive power ramp rate limit microservice and switch to the maximum power point tracking operation.

[0005] Furthermore, for the fast frequency modulation method of the wind turbine generator based on disturbance adaptive compensation of the present invention, Step S2 includes: Based on the dynamic disturbance perception data set, dynamically adjust the crossover probability and mutation probability according to the population fitness distribution, and increase the mutation probability to the range of 0.2 - 0.4 when the population diversity is lower than the set threshold; Adopt the elite retention strategy to directly transfer the optimal individual of each generation to the next generation, and accelerate the parameter convergence in combination with the distribution characteristics of the disturbance feature data; Combine the genetic algorithm with GPU parallel computing to perform parallel evaluation on the window length and weight coefficient of the short-time window sliding mean algorithm; Initialize the test signal amplitude parameter based on the historical disturbance data in the dynamic disturbance perception data set, and reduce the frequency interference of inertia estimation through reverse signal excitation.

[0006] Furthermore, the fast frequency modulation method of the wind turbine generator based on disturbance adaptive compensation of the present invention further includes: According to the execution feedback data of the compensation power command, adopt a reinforcement learning mechanism to adjust the activation weight of the fuzzy rules; Dynamically divide the fuzzy interval of the equivalent inertia parameter of the power grid through a clustering algorithm, and optimize the center point and width parameters of the membership function; Establish a parameter correlation matrix between the adaptive sliding mode control model and the fuzzy controller, and synchronously update the boundary layer thickness and input range; When a high-frequency disturbance mode is detected, preferentially optimize the high-frequency segment membership function parameters of the equivalent inertia parameter of the power grid and enhance the weight distribution of the corresponding rules.

[0007] Further, in the method for fast frequency modulation of a wind turbine based on disturbance adaptive compensation according to the present invention, step S6 includes: In the containerized deployment framework of the model predictive control, optimize the power distribution ratio of the rotor kinetic energy reserve and the converter capacity through the dynamic programming microservice; Integrate the communication protocols of the multi-source frequency modulation devices of the power grid through the API gateway to generate a cross-device frequency modulation instruction synchronization protocol; Calculate the disturbance energy entropy based on the dynamic disturbance perception dataset, and design an event-driven priority weight microservice to dynamically adjust the output ratio; When a power conflict is detected, trigger a redundancy elimination algorithm to reallocate the frequency modulation tasks of the virtual synchronous machine and the energy storage device.

[0008] Further, in the method for fast frequency modulation of a wind turbine based on disturbance adaptive compensation according to the present invention, step S7 includes: Input the historical disturbance data and real-time inertia parameters in the dynamic disturbance perception dataset into a rolling horizon optimization model to predict the frequency trend in the next 3-5 sampling periods; Use the multi-island genetic algorithm to divide the initial population of the genetic algorithm, and regularly migrate elite individuals to generate a differential trigger threshold; When the rate of change of the frequency deviation exceeds the start threshold, switch to the fast power support mode to release the rotor kinetic energy reserve; In the droop control stage, generate a power slope curve based on the compensation power instruction, and combine the fan speed recovery rate calculated from the speed change gradient after the release of the rotor kinetic energy reserve of the wind turbine to dynamically update the ramp rate limit.

[0009] Further, the method for fast frequency modulation of a wind turbine based on disturbance adaptive compensation according to the present invention further includes: Use the fan speed recovery rate and the frequency deviation as hard constraint conditions to construct a gradient search space; Solve the optimal power descent gradient parameter in the constraint space through the genetic algorithm; When a rotational speed mutation risk is detected, trigger a dynamic relaxation mechanism to expand the constraint boundary of the sliding mode control model; Initialize the search space with the optimal gradient parameter in the historical frequency modulation exit data to shorten the generation time of the safety recovery strategy.

[0010] Further, in the method for fast frequency modulation of a wind turbine based on disturbance adaptive compensation according to the present invention, step S8 includes: Collect the real-time data stream of the continuous disturbance event through a sliding time window, and perform principal component analysis for dimensionality reduction processing; The boundary layer parameters of the adaptive sliding mode control model are updated only based on incremental data by using the recursive least squares method; When data anomalies are detected, a local model rollback mechanism is triggered to restore to the effective parameters of the previous time period; The learning rate is dynamically adjusted according to the change rate of the model prediction error, and the parameter update speed is increased when the error gradient increases.

[0011] Advantages of the present invention; Through the construction of a dynamic disturbance perception data set and the optimization of the genetic algorithm, the present invention significantly improves the noise suppression ability of the short-time window moving average algorithm and enhances the accuracy of disturbance feature extraction; combined with the recursive least squares method to estimate the equivalent inertia parameters of the power grid in real time, it effectively adapts to the parameter dynamic tracking requirements in the inertia time-varying scenario and reduces the identification error. The online incremental learning algorithm and the fuzzy adaptive proportional-integral controller cooperate to optimize and compensate the power command generation logic, strengthen the dynamic response ability under high-frequency disturbances, and realize the efficient cooperation of multi-source frequency modulation devices through the microservice architecture, eliminating power redundancy and conflicts. The genetic algorithm is introduced to optimize the trigger threshold and the safe exit mechanism, balance the switching timing between the fast power support mode and the droop control, and realize the smooth transition of the frequency modulation exit process in combination with the adaptive power ramp rate limit. The above-mentioned technologies work together to solve the problems of response lag and insufficient resource coordination of traditional methods in the dynamic multi-disturbance superposition and inertia time-varying scenarios, improve the real-time performance, stability and multi-source device cooperation efficiency of power system frequency regulation, and reduce the risk of secondary frequency drop. Brief description of the drawings

[0012] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, other drawings can also be obtained according to the drawings without creative efforts.

[0013] Figure 1 It is a flowchart of a fast frequency modulation method for a wind turbine based on disturbance adaptive compensation provided by an embodiment of the present invention. Detailed implementation manners

[0014] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings. To better understand the objectives of the present invention, the present invention is further described in detail below.

[0015] Please refer to Figure 1 , the fast frequency modulation method for wind turbines based on disturbance adaptive compensation provided by the present invention includes: Step S1, collect the frequency deviation, frequency deviation change rate of the power system and the operating state parameters of the fan, and construct a dynamic disturbance perception data set. The fan operating state parameters include the real-time monitoring data of the rotor kinetic energy reserve of the wind turbine; When collecting the frequency deviation of the power system, the frequency measurement unit deployed at the grid node obtains the AC voltage frequency signal of the grid connection point in real time at a preset sampling frequency. The frequency measurement unit includes a digital signal processing module, which converts the original voltage signal into frequency deviation time series data through Fourier transform. When synchronously collecting the frequency deviation change rate, a differential tracking filter is used to perform a first-order difference operation on the frequency deviation signal, and the mean square value of the frequency deviation is statistically calculated in combination with a sliding time window to generate a frequency deviation change rate feature vector. The collection of the fan operating state parameters is realized by vibration sensors and speed encoders deployed on the drive chain of the wind turbine, and the angular acceleration and kinetic energy reserve of the generator rotor are monitored in real time, where the rotor kinetic energy reserve is calculated based on the product of the rotor mass inertia parameter and the square of the measured speed.

[0016] When constructing the dynamic disturbance perception data set, align the frequency deviation time series data, frequency deviation change rate feature vector and rotor kinetic energy reserve monitoring data according to the unified time stamp, and use the sliding window mechanism to synchronously fuse the multi-source data. The length of the sliding window is set to 5-10 power frequency cycles according to the typical disturbance duration, and the abnormal sampling points are removed from the data in the window through the outlier detection algorithm. The synchronously fused data set is stored in a time series database, and the data record includes fields such as time stamp, frequency deviation, frequency deviation change rate, rotor speed and kinetic energy reserve, and a grid node topology identifier is added to distinguish the data sources of different regions.

[0017] The real-time update mechanism of the dynamic disturbance perception data set adopts an event-driven mode. When it is detected that the frequency deviation exceeds the preset threshold or the fan speed change gradient is abnormal, the data acquisition module is triggered to increase the sampling frequency to 1.5-2 times the preset value. The newly added data is appended to the end of the time series database after verification, and the historical data exceeding the storage capacity is overwritten through the circular buffer management mechanism to maintain the time series continuity of the data set. In the data preprocessing stage, the principal component analysis method is introduced to perform dimensionality reduction processing on the multi-dimensional features, and the key principal component vectors representing the dynamic characteristics of the system are extracted to reduce the computational complexity of the subsequent feature extraction algorithm.

[0018] The real-time monitoring data of the rotor kinetic energy reserve of the wind turbine is obtained through the energy management module of the converter control system, and the module integrates a speed-power mapping table and a rotor kinetic energy prediction model. During the real-time monitoring process, the speed signal collected by the encoder is input into the prediction model after Kalman filtering, and the percentage of available kinetic energy reserve is calculated in combination with the current wind speed and pitch angle parameters. The prediction model is trained based on historical operation data, and a long short-term memory network is used to establish a non-linear mapping relationship between speed fluctuations and kinetic energy release capabilities, and the output result is used as the core state parameter of the dynamic disturbance perception data set.

[0019] The construction logic of the dynamic disturbance perception data set includes a multi-level data verification mechanism. During the data fusion stage, redundant check codes are used to verify the integrity of the data of each sensor, and the sampling points with failed verification are discarded. The timestamp alignment mechanism is implemented based on the IEEE 1588 precise time protocol to synchronize the clock of the acquisition devices at different nodes to the microsecond level of accuracy. The data storage structure uses a columnar database to optimize the real-time reading and writing performance, and multiple data copies are deployed through a distributed architecture to improve the reliability of data access and the concurrent processing ability.

[0020] The dynamic disturbance perception data set provides basic data support for closed-loop control, and its construction process and the subsequent feature extraction module form a feed-forward link. The frequency deviation and change rate data characterize the dynamic characteristics of the power grid, and the rotor kinetic energy reserve parameter reflects the frequency modulation ability reserve of the wind turbine. The spatio-temporal correlation characteristics of multi-dimensional data capture the disturbance propagation law through a sliding window mechanism. The dynamic update characteristic of the data set adapts to the time-varying scenario of the power system inertia, provides the original input for the noise suppression algorithm, and supports the real-time optimization of the subsequent control strategy.

[0021] Step S2, based on the dynamic disturbance perception data set, dynamically optimize the window length and weight coefficient of the short-time window sliding mean algorithm through a genetic algorithm to generate disturbance feature data after noise suppression; When constructing the initial population based on the dynamic disturbance perception data set, the individual coding of the genetic algorithm adopts a mixed coding method of binary and real numbers, maps the window length parameter of the short-time window sliding mean algorithm to an integer gene segment, and encodes the weight coefficient as a floating-point gene segment. The population size is set to 50-100 groups of candidate solutions according to the possible value range of the sliding window parameters. The initial individuals are generated by the Monte Carlo random sampling method, covering the typical value range of the sliding window length parameter, and the weight coefficient is evenly distributed in the normalized interval from 0 to 1. The calculation of the individual fitness function is based on the noise suppression effect evaluation index, including the weighted combination of the signal-to-noise ratio improvement rate of the disturbance feature signal and the trend fitting error.

[0022] During the population evolution process, the roulette wheel selection mechanism screens out individuals with high fitness values according to the fitness values and enters them into the mating pool. At the same time, the elite retention strategy is adopted to directly retain the optimal individual of each generation to the next generation population. The crossover operation adopts the multi-point crossover method, sets fixed crossover points in the sliding window length gene segment, and uses arithmetic crossover in the weight coefficient gene segment to generate new individuals. The mutation probability is dynamically adjusted according to the monitoring results of population diversity. When the genotype similarity of individuals exceeds the preset threshold, the adaptive mutation mechanism is triggered, and a random perturbation of ±10% is applied to the sliding window length parameter, and the weight coefficient is re-initialized within the range of 0.1 to 0.9.

[0023] The parameter optimization process of the short-time window sliding mean algorithm is accelerated by combining with the GPU parallel computing framework. The population individuals are assigned to different computing units for parallel evaluation. Each thread block independently calculates the noise suppression effect of the corresponding window length and weight coefficient combination, and exchanges the intermediate calculation results through shared memory. The parallel computing architecture is implemented based on the CUDA platform, and the dynamically perturbed perception dataset is divided into multiple data blocks and loaded into the video memory to reduce the data transmission delay. After the fitness evaluation results are summarized, the tournament selection mechanism is used to screen out the optimal parameter combination as the real-time configuration parameters of the sliding mean algorithm.

[0024] During the generation process of the perturbed feature data after noise suppression, the optimized sliding window length parameter is applied to the real-time data stream processing. The weight coefficients of the data points within the window are dynamically weighted according to the Gaussian distribution function, with the highest weight value assigned to the center point and the weights of the edge data points gradually decreasing. The calculation result of the weighted mean is smoothed twice by a moving average filter to eliminate the remaining high-frequency noise components. The processed perturbed feature data is stored in a structured data format containing timestamps, frequency deviation correction values, and feature intensities for subsequent calls by the inertia parameter identification module.

[0025] Based on the historical perturbation data in the dynamically perturbed perception dataset, the amplitude parameter of the reverse test signal is initialized by the energy normalization method. The energy distribution characteristics of historical perturbation events are statistically analyzed through a sliding time window, and the 80th percentile of the maximum energy value is extracted as the reference amplitude of the test signal. The signal timing characteristics are synchronously matched with the current perturbation characteristics, and the phase compensation algorithm is used to adjust the time alignment accuracy between the signal waveform and the real-time perturbation. When the reverse test signal is injected into the grid equivalent inertia estimation link, the signal frequency band range is restricted by a band-pass filter to avoid introducing additional frequency interference.

[0026] The recursive least squares method introduces a forgetting factor in the process of inertia parameter estimation, dynamically adjusting the weight distribution ratio of historical data. The value of the forgetting factor is dynamically adjusted according to the signal-to-noise ratio of the disturbance characteristic data. When the noise suppression effect is improved, the forgetting factor is increased to more than 0.95 to strengthen the parameter correction effect of real-time data. The parameter estimation result is iteratively optimized through the covariance matrix update mechanism. Only the parameter correction amount of the incremental data is calculated when each new data window arrives, reducing the computational complexity. The identified equivalent inertia parameter of the power grid is used as the input variable of the fuzzy controller to participate in the generation logic of the subsequent compensation power command.

[0027] Step S3: Generate a reverse test signal according to the disturbance characteristic data, and estimate the equivalent inertia and damping parameters of the power grid in real time by combining the recursive least squares method; When generating the reverse test signal, extract the spectral principal components based on the disturbance characteristic data after noise suppression, and generate a reverse waveform with the opposite phase to the original disturbance through the band-limited signal reconstruction algorithm. The signal amplitude is initialized according to the energy distribution characteristics of historical disturbance events. The 80th percentile value of the maximum disturbance energy is used as the reference amplitude by using a sliding time window, and the signal intensity is dynamically adjusted in combination with the current power grid frequency fluctuation amplitude. The signal timing characteristics are aligned with the real-time disturbance characteristics through the phase compensation algorithm to eliminate the signal superposition error caused by time delay. When injecting into the power grid, the anti-aliasing filter is used to limit the signal frequency band range to the interval of 0.5 - 5 Hz to avoid introducing high-frequency interference components.

[0028] In the process of estimating the equivalent inertia parameter of the power grid by the recursive least squares method, an input-output matrix is constructed with the reverse test signal and the power grid frequency response data. The input matrix contains the timing amplitude and injection time information of the reverse test signal, and the output matrix is the observed value of the frequency deviation change rate. The forgetting factor is dynamically adjusted according to the signal-to-noise ratio of the disturbance characteristic data. When the noise suppression effect is improved, it is gradually increased to more than 0.95 to strengthen the correction effect of real-time data on parameter estimation. The covariance matrix is iteratively optimized by using the rank-one update mechanism. Only the correction amount of the incremental data to the covariance matrix is calculated when each new data window arrives, reducing the computational complexity.

[0029] When estimating the equivalent inertia parameter of the power grid in real time, the frequency response data after injecting the reverse test signal is intercepted through a sliding time window to construct a dynamic regression model. The model takes the equivalent inertia and damping parameters of the power grid as variables to be identified, and iteratively optimizes the parameter estimation value based on the minimum prediction error criterion. A regularization term is introduced during the parameter update process to constrain the parameter change amplitude and prevent the estimation value from diverging due to data mutations. The identification result is smoothed by a moving average filter to eliminate instantaneous noise interference and output a stable equivalent inertia parameter sequence.

[0030] The estimation of the damping parameter is combined with the analysis of the grid frequency oscillation attenuation characteristics, and the frequency free response curve is intercepted after the reverse test signal is removed. The attenuation rate of the oscillation amplitude is calculated by the exponential fitting algorithm, and the estimated value of the equivalent damping coefficient is deduced. The fitting process uses the nonlinear least squares optimization algorithm, the initial value is set based on historical operation data, and the iteration step size is dynamically adjusted according to the gradient descent rate to improve the parameter convergence speed. The fitting result is weighted and fused with the damping parameter output by the recursive least squares method to form the final damping parameter estimation value.

[0031] The cooperative working mechanism of the reverse test signal and the parameter estimation module is realized through event-triggered logic. When it is detected that the frequency deviation exceeds the preset threshold or the disturbance energy entropy changes significantly, the signal injection process is started. The signal duration is determined according to the autocorrelation function analysis of the disturbance characteristics, covering more than 90% of the time period of the main disturbance energy. The parameter estimation result is transmitted to the fuzzy adaptive controller in real time and used as the core input variable for generating the compensation power command to support the optimization of the dynamic frequency modulation strategy under different disturbance scenarios.

[0032] The online identification accuracy of the grid equivalent inertia parameter is improved through a double verification mechanism, including historical data backtracking verification and real-time prediction error monitoring. The historical data backtracking verification uses a sliding window mechanism to compare the parameter estimation value with the off-line simulation result, and triggers the parameter recalibration process when the deviation exceeds 5%. The real-time prediction error monitoring compares the model output frequency with the actual observed value, dynamically adjusts the forgetting factor and the regularization coefficient, and maintains the stability and tracking ability of the parameter estimation. The verified equivalent inertia parameter is used as the basis for the analysis of the dynamic characteristics of the power system and participates in the cooperative control logic of multi-source frequency modulation equipment.

[0033] Step S4, an online incremental learning algorithm is used to perform time series modeling on the continuous disturbance events in the dynamic disturbance perception dataset, and update the boundary layer thickness and switching gain parameters of the preset adaptive sliding mode control model; When using the online incremental learning algorithm to perform time series modeling on continuous disturbance events, the real-time data stream in the dynamic disturbance perception dataset is intercepted through a sliding time window, and the window length is dynamically adjusted to 3-5 power frequency cycles according to the periodic characteristics of the disturbance events. The data in the window is processed by principal component analysis for dimensionality reduction, and the principal component vector with a contribution rate exceeding 85% is extracted to eliminate redundant noise interference and generate low-dimensional input features suitable for the sliding mode control model. The dimensionality-reduced feature vector is standardized to eliminate the dimension difference and form the input matrix of the time series model, improving the calculation efficiency of subsequent parameter updates.

[0034] When updating the boundary layer thickness and switching gain parameters of the adaptive sliding mode control model, the recursive least squares method is used to iteratively optimize the model parameters based only on incremental data. The recursive algorithm retains the covariance matrix information of historical parameters. Each time a new data window is received, the covariance matrix and parameter vector are updated through the matrix inversion lemma, avoiding recomputation of all data. The update amount of the boundary layer thickness is dynamically adjusted according to the steady-state error gradient of the frequency deviation, and the switching gain parameter is weighted and corrected by the change rate of the disturbance characteristics to balance the rapidity and smoothness of the control response.

[0035] When data anomalies are detected, a local model rollback mechanism is triggered to restore to the effective parameters of the previous time period. The anomaly detection module continuously monitors the statistical characteristics of the input data, including mean drift, variance mutation, and data distribution deviation metrics. If the data deviation of three consecutive sampling windows exceeds the threshold, the rollback mechanism calls the historical parameter database to load the verified effective sliding mode control parameters of the previous time period, overwriting the update results driven by the current abnormal data. The model after rollback maintains a stable output until subsequent data passes the verification, ensuring the continuity of the control instruction.

[0036] During the process of dynamically adjusting the learning rate, the parameter update step size is adaptively corrected according to the change rate of the model prediction error. The gradient of the prediction error is calculated through the error difference within the sliding window, and the exponential weighted average is combined to eliminate instantaneous fluctuation interference. When the absolute value of the error gradient exceeds the preset threshold, the learning rate is increased to 1.5 - 2 times the reference value according to the logarithmic function law to accelerate parameter convergence; when the error gradient approaches zero, the learning rate is gradually attenuated to the initial value to refine the parameter search accuracy. The dynamic adjustment mechanism matches the model convergence requirements under different disturbance scenarios, preventing overfitting or underfitting problems.

[0037] The collaborative working mechanism of the time series modeling module and the sliding mode controller is realized through event-triggered logic. When it is detected that the change rate of the frequency deviation exceeds the set threshold or the disturbance energy entropy increases significantly, an incremental learning process is triggered. The model update results are immediately applied to the switching logic of the sliding mode controller, dynamically correcting the boundary layer thickness to adapt to the control requirements of different disturbance intensities. The updated switching gain parameter is synchronized to the input universe of discourse of the fuzzy controller through the parameter correlation matrix, eliminating the response delay between multi-level controllers and improving the collaborative control efficiency.

[0038] The closed-loop parameter update link constructed by the online incremental learning algorithm captures the dynamic characteristics of disturbances through the sliding window mechanism, recursively optimizes to reduce the computational load, and abnormal rollback ensures the system robustness. The dynamic learning rate matches the convergence requirements. The data interaction of each technical link strictly follows the timing logic, forming a complete closed loop from disturbance perception to control parameter optimization, enhancing the adaptive ability of the sliding mode control in continuous disturbance scenarios and improving the dynamic response accuracy of the wind turbine frequency regulation.

[0039] Step S5: Based on the equivalent inertia and damping parameters of the power grid, generate a compensation power command through a preset fuzzy adaptive proportional and preset integral controller. The compensation power command includes the triggering conditions for the fast power support mode and the droop control mode; When generating the compensation power command, the input variables of the fuzzy adaptive proportional-integral controller include the equivalent inertia parameter of the power grid, the real-time frequency deviation, and the change rate. The universe of discourse division of the input variables is dynamically adjusted by a clustering algorithm. The inertia parameter is divided into 3 - 5 fuzzy intervals based on historical operation data, and the frequency deviation and change rate are divided into overlapping membership functions according to the normal distribution characteristics. The center point and width parameters of the membership function are updated through an online learning mechanism. When high-frequency disturbance characteristics are detected, the width of the high-frequency segment function is reduced to 60% of the reference value to improve the rule matching resolution.

[0040] The construction of the fuzzy rule base adopts a combination of expert experience and data-driven methods. The initial rule set is defined based on the frequency response characteristics under typical disturbance scenarios. The reinforcement learning mechanism dynamically adjusts the rule activation weights according to the execution feedback of the compensation power command. The feedback data includes the frequency fluctuation suppression rate, the frequency modulation energy consumption, and the power overshoot. The Q-learning algorithm iteratively updates the rule weight matrix, preferentially strengthening the optimal control path under high-frequency disturbance scenarios. The matching results of historical similar disturbance patterns are weighted and fused into the reward function to improve the dynamic adaptability of the rule base.

[0041] The triggering condition of the fast power support mode is realized by a dynamic threshold comparator. When the change rate of the frequency deviation exceeds 120% of the preset threshold or the equivalent inertia parameter is lower than the critical value, a fast power support command is triggered. The threshold parameter is generated by rolling optimization of the multi-island genetic algorithm and dynamically adjusts the threshold boundary in combination with the real-time inertia parameter prediction model. After the trigger signal is generated, the control command preferentially releases 80% - 90% of the pre-stored rotor kinetic energy of the wind turbine and quickly injects the compensation power through the converter to suppress the further drop of the frequency.

[0042] The activation of the droop control mode is based on the steady-state error range of the frequency deviation. When the frequency deviation continuously stays outside the preset dead zone range and the change rate is lower than the fast mode threshold, it switches to the droop control stage. In this mode, a power slope curve command is generated. The initial value of the slope is set according to the equivalent damping parameter and the remaining capacity of the rotor kinetic energy, and the power drop gradient is dynamically adjusted in combination with the real-time monitoring value of the speed recovery rate. The dead zone control logic is introduced in the slope adjustment process to shield the influence of small speed fluctuations on the command generation and maintain the continuity of the power output.

[0043] The collaborative working mechanism of the fuzzy controller and the sliding mode controller is realized through the parameter correlation matrix. The thickness parameter of the sliding mode boundary layer is dynamically correlated with the range of the fuzzy input universe of discourse. When the sliding mode control model updates the boundary layer parameters, the correlation matrix drives the synchronous adjustment of the input range of the fuzzy controller, eliminating the response delay between the two-level controllers. The output interface of the control instruction is encapsulated by a standardized protocol, including the power support mode identifier, instruction amplitude, and slope parameters, which is compatible with the communication protocol of the microservice architecture to ensure the real-time transmission of cross-module instructions.

[0044] The generation logic of the compensation power instruction is closed-loop connected to the collaborative control link of the multi-source frequency modulation equipment. The instruction parameters are allocated to the virtual synchronous machine and the energy storage device after being parsed by the dynamic programming microservice. The event-driven priority weight microservice dynamically adjusts the output ratio of each device based on the disturbance energy entropy, and triggers the redundancy elimination algorithm to reallocate tasks when instruction conflicts are detected. The generated control instruction is simultaneously fed back to the online learning module to form a closed-loop optimization link from parameter estimation, instruction generation to execution feedback, improving the frequency modulation control accuracy in complex disturbance scenarios.

[0045] Step S6, according to the compensation power instruction, coordinate the allocation of the rotor kinetic energy reserve of the wind turbine generator set and the power response of the multi-source frequency modulation equipment of the power grid through a preset microservice architecture, where the multi-source frequency modulation equipment of the power grid includes a virtual synchronous machine and an energy storage device; When coordinating the allocation of the rotor kinetic energy reserve of the wind turbine generator set, the dynamic programming microservice parses the power demand parameters and time constraint conditions in the compensation power instruction to construct a multi-objective optimization model. The model takes the maximization of the frequency stability recovery speed and the minimization of the rotor mechanical stress as the optimization objectives, and takes the rotor kinetic energy reserve capacity of the wind turbine generator set, the upper limit of the converter output, and the response rate of the energy storage device as the constraint conditions. In the optimization process, a linear programming algorithm is used to solve the optimal combination of the rotor kinetic energy release ratio and the converter power output, generate a power allocation scheme adapted to the dynamic disturbance scenario, and distribute the allocation instruction to the frequency modulation equipment node through a lightweight message queue.

[0046] In the microservice architecture, the API gateway integrates the heterogeneous communication protocols of the virtual synchronous machine and the energy storage device, and encapsulates the Modbus and IEC 61850 protocols into a standardized JSON instruction format through RESTful interfaces. The gateway real-time monitors the communication delay and bandwidth occupancy rate of each frequency modulation equipment, and dynamically adjusts the data packet transmission priority to ensure the timing consistency of cross-device frequency modulation instructions. The instruction synchronization protocol includes the device status query, power instruction issuance, and execution feedback confirmation processes, and adopts a timeout retransmission mechanism to improve the reliability of instruction transmission, while being compatible with the response characteristic differences of different devices.

[0047] The event-driven priority-weighted microservice calculates the disturbance energy entropy based on the dynamically perturbed perception dataset to quantify the spatio-temporal energy concentration degree of disturbance events. When the entropy value exceeds the preset threshold, the output weights of each device are dynamically allocated according to the real-time available capacity, response rate, and frequency regulation cost coefficient of the virtual synchronous machine and the energy storage device. The weight allocation results are shared in real time through the distributed cache cluster, supporting each frequency regulation node to adjust the output target proportionally, and synchronously updated to all associated devices through the publish-subscribe mechanism.

[0048] When a power conflict is detected, the redundancy elimination algorithm constructs an operating state parameter matrix of the conflicting devices to identify device combinations with overshoot or contradictory directions of output. The algorithm quickly solves the conflict resolution scheme in the hard constraint space based on the genetic algorithm. The constraint conditions include the upper limit of device output, frequency restoration requirements, and mechanical stress limitations. The initial population accelerates convergence using the historical feasible solution set, and the crossover and mutation operations generate new candidate solutions. After being evaluated by the fitness function, the optimal task allocation strategy is output, and the corrected frequency regulation instructions are reissued through containerized microservices.

[0049] During the execution of the frequency regulation task, the release ratio of the rotor kinetic energy reserve of the wind turbine is dynamically adjusted through the converter control logic. In the initial stage of release, 80%-90% of the pre-stored kinetic energy is preferentially called to quickly inject into the power grid. In the subsequent stage, the release rate is gradually reduced in combination with the rotational speed recovery gradient and the frequency deviation change rate. The power response of the energy storage device adopts a segmented ramp control strategy. In the initial stage, it quickly outputs at 120% of the rated power, and switches to the power tracking mode coordinated with the virtual synchronous machine after entering the steady state to maintain the smoothness of the frequency recovery curve.

[0050] The collaborative control link of the microservice architecture forms a closed-loop optimization mechanism. The initial allocation strategy is generated by dynamic programming. The API gateway ensures the cross-device synchronous transmission of instructions. The priority weights are dynamically adjusted to optimize the resource utilization efficiency, and the redundancy elimination algorithm solves the execution conflicts. Each module is tightly coupled through the standardized data interface and event trigger logic to respond in real time to the changes in the power grid disturbance characteristics, improving the collaborative efficiency of multi-source frequency regulation devices and the system frequency stability.

[0051] Step S7, introducing the genetic algorithm in the adaptive sliding mode control model predictive control to optimize the trigger threshold of dynamic switching, and switching the fast power support mode and the droop control mode according to the frequency deviation change rate; When constructing a trigger threshold optimization module in adaptive sliding mode control model predictive control, the multi-island genetic algorithm divides the search space into 4-6 independently evolving subpopulations. Each subpopulation takes the historical disturbance data intercepted by a sliding time window and real-time inertia parameters as inputs to generate differential trigger threshold candidate solutions. The initial population is initialized based on the feasible threshold combinations in historical frequency regulation scenarios. The fitness function takes the frequency stability index as the core evaluation parameter and comprehensively evaluates the weighted scores of the threshold switching times and the frequency recovery rate. After every 5-10 generations of iteration, the top 10% of the elite individuals in each subpopulation are migrated to adjacent subpopulations, and new threshold combinations are generated through arithmetic crossover operations to expand the global search range of the trigger threshold.

[0052] The triggering conditions of the dynamic switching logic are real-time tuned through a rolling horizon optimization model, and the predicted future 3-5 sampling period frequency trend data is input into a threshold comparator. When the rate of change of frequency deviation exceeds 120% of the current threshold or the equivalent inertia parameter is lower than the critical value, a fast power support mode switching instruction is triggered. The critical value is set based on the wind turbine rotor stress analysis model, and the threshold boundary is dynamically adjusted in combination with the real-time rotational speed recovery gradient to prevent false triggering caused by parameter mutations. After the trigger signal is generated, the control instruction preferentially releases 85%-95% of the pre-stored rotor kinetic energy, and smooths the power output curve through the inertia compensation algorithm of the converter control loop.

[0053] A dynamic association matrix is established between the boundary layer parameters of the sliding mode control model and the trigger threshold. When the genetic algorithm optimizes and generates a new threshold combination, the boundary layer thickness is synchronously adjusted through the matrix mapping relationship. The update amount of the boundary layer parameters is calculated according to the steady-state error gradient of the frequency deviation, and the thickness coefficient is iteratively corrected using the recursive least squares method to balance the rapidity and smoothness of the control response. The switching gain parameter is weighted and adjusted according to the energy distribution of the disturbance characteristics, and the gain value is increased to 1.2-1.5 times the reference level in high-frequency disturbance scenarios to enhance the ability to suppress mutation disturbances.

[0054] A double verification mechanism is introduced in the mode switching process. Before the control instruction is issued, the feasibility of the switching strategy is verified through a virtual simulation module. The simulation module operates based on the real-time power grid equivalent model, simulates the frequency response curves under different threshold combinations, and screens the candidate solutions that meet the frequency fluctuation suppression rate and mechanical stress constraints. The verified switching instructions are encapsulated through a standardized protocol and distributed to the actuators, and at the same time, they are fed back to the genetic algorithm population database to update the historical feasible solution set to accelerate the subsequent optimization process.

[0055] The anomaly handling mechanism is activated in real time by monitoring the deviation degree of the sliding mode surface trajectory. When it is detected that the control quantity exceeds the preset safety boundary or there is a risk of secondary drop in frequency deviation, the local model rollback is triggered. The rollback mechanism calls the last three valid parameter combinations in the historical parameter database and uses the majority voting mechanism to select the optimal parameters to cover the current abnormal state. The control model after rollback maintains stable operation until the incremental learning module completes a new round of parameter optimization, forming a closed-loop control link for fault self-healing and parameter update.

[0056] Step S8, when the frequency deviation returns to the preset dead zone range, the micro-service for adaptive power ramp rate limitation exits the frequency regulation and switches to the maximum power point tracking operation through the adaptive power ramp rate limitation micro-service.

[0057] When the frequency deviation enters the preset dead zone range, the micro-service for adaptive power ramp rate limitation starts the frequency regulation exit process and monitors the steady-state holding time of the frequency deviation through a sliding time window. The dead zone range is dynamically set based on the grid frequency regulation regulations. When the absolute values of the frequency deviations of 5 consecutive sampling points within the window are all lower than the threshold, it is determined that the steady-state recovery condition is established. The micro-service calls the genetic algorithm to solve the optimal power descent gradient parameters. The constraint conditions include the upper limit of the fan speed recovery rate and the frequency secondary drop risk coefficient. The initial population of the algorithm uses the feasible solution set of the historical frequency regulation exit scenarios to accelerate convergence and outputs a ramp rate parameter combination that takes into account both safety and economy.

[0058] During the power ramp rate adjustment stage, the converter control module gradually reduces the compensation power output according to the slope curve generated by the genetic algorithm. The slope parameter is dynamically corrected according to the real-time speed recovery gradient. When it is detected that the speed change rate is lower than the preset safety threshold, the gradient relaxation mechanism is triggered to temporarily increase the allowed drop amplitude to prevent the mechanical transmission chain from being overloaded. The corrected slope instruction is encapsulated by the standardized protocol and distributed to the virtual synchronous machine and the energy storage device through the API gateway to coordinate the multi-source devices to synchronously reduce the output and maintain the power balance of the power grid.

[0059] When switching to the maximum power point tracking operation, the fan control system adopts a progressive mode transition strategy. In the initial stage, 10%-15% of the rotor kinetic energy is reserved as inertial buffer. The pitch angle and the converter reference power are dynamically adjusted through the speed-power mapping table to make the output power smoothly transition along the maximum power point curve. During the transition period, the change trend of the frequency deviation is continuously monitored. If it is detected that the deviation exceeds the dead zone range again, the exit process is immediately interrupted and the frequency regulation mode is reactivated to form a closed-loop protection mechanism for mode switching.

[0060] The anomaly handling module identifies potential rotational speed mutation risks by monitoring the frequency volatility and equipment status parameters during the power decline process. When the second derivative of the rotational speed recovery rate exceeds the mechanical strength limit, a dynamic relaxation mechanism is triggered to temporarily expand the constraint boundary of the genetic algorithm and re-solve the safe exit strategy. Meanwhile, the local model rollback function is called to load the ramp rate parameters that have been verified effective in the previous time period, overwrite the optimization results driven by the current abnormal data, and maintain the stability of the control process.

[0061] The adaptive power ramp rate limit microservice constructs a complete closed-loop for frequency modulation exit. Dead zone monitoring triggers the optimization process, the genetic algorithm generates safety parameters, multi-source devices cooperate to execute power ramp control, and anomaly handling ensures transition reliability. Each technical link realizes state synchronization through the event bus and data pipeline, dynamically adjusts the instruction parameters and device response logic, realizes seamless switching from the frequency modulation mode to the maximum power point tracking mode, and improves the long-term stability of the grid frequency.

[0062] The fast frequency modulation method for wind turbines based on disturbance adaptive compensation provided by the present invention includes the following steps: Real-time collect the frequency deviation, frequency deviation change rate of the power system and the operating state parameters of the wind turbine, and construct a dynamic disturbance perception data set. The data collection is completed through a sensor network deployed at key nodes of the power grid. The sensor types include frequency measurement units, phasor measurement units and wind turbine state monitoring modules, and the sampling frequency is dynamically adjusted according to the disturbance characteristics of the power grid. After the collected data is preprocessed by timestamp alignment and outlier removal, it is stored as a data structure containing the time series of frequency deviation, statistical characteristics of frequency deviation change rate and the rotational speed-power mapping relationship of the wind turbine, forming a dynamic disturbance perception data set, which provides input for subsequent feature extraction and control model update.

[0063] Based on the dynamic disturbance perception data set, the window length and weight coefficient of the short-time window sliding mean algorithm are dynamically optimized through the genetic algorithm to generate disturbance feature data after noise suppression. The initial population of the genetic algorithm is composed of sliding window parameter combinations, and the fitness function takes the signal-to-noise ratio and trend retention ability of the disturbance features as evaluation indicators. During the population evolution process, the roulette wheel selection mechanism is used to screen individuals, and the crossover and mutation probabilities are dynamically adjusted based on the population diversity index. The fitness evaluation is accelerated by GPU parallel computing. Each thread block independently calculates the noise suppression effect of different window parameter combinations and outputs the optimal combination of window length and weight coefficient to suppress the interference of grid background noise on the disturbance features.

[0064] Generate a reverse test signal based on the disturbance characteristic data after noise suppression, and combine the recursive least squares method to estimate the equivalent inertia and damping parameters of the power grid in real time. The amplitude of the reverse test signal is initialized based on the historical disturbance energy distribution, and the timing parameters are synchronously adjusted with the current disturbance characteristics to reduce additional frequency interference. The equivalent inertia and damping parameters of the power grid are identified online through the recursive least squares method. A forgetting factor is introduced during the identification process to optimize the weight of historical disturbance data and improve the dynamic tracking ability of parameter estimation. The identification results are used as the input parameters of the fuzzy controller to support the generation of subsequent compensation power commands.

[0065] Use an online incremental learning algorithm to perform time series modeling on continuous disturbance events in the dynamic disturbance perception dataset, and update the boundary layer thickness and switching gain parameters of the adaptive sliding mode control model. The incremental learning algorithm intercepts the real-time data stream of continuous disturbance events through a sliding time window, and inputs it into the sliding mode control model after dimensionality reduction by principal component analysis. The recursive least squares method updates the boundary layer parameters only based on incremental data, reducing the computational complexity. When data anomalies are detected, a local model rollback mechanism is triggered to load the effective parameters of the previous time period to maintain control stability. The learning rate is dynamically adjusted according to the gradient of the model prediction error, accelerating parameter updates when the error increases and refining the search accuracy when the error converges.

[0066] Based on the real-time estimated equivalent inertia and damping parameters of the power grid, generate a compensation power command through a fuzzy adaptive proportional-integral controller. The membership function parameters of the fuzzy controller are dynamically divided by a clustering algorithm, and the activation weights of the fuzzy rule base are iteratively optimized by a reinforcement learning mechanism according to the historical frequency modulation effect. The input variables of the controller include equivalent inertia, frequency deviation and change rate, and the output is a multi-level compensation power command. In high-frequency disturbance scenarios, the resolution of the membership function in the high-frequency band is preferentially adjusted to enhance the weight allocation of corresponding rules and improve the dynamic adaptability of the compensation command.

[0067] According to the compensation power command, coordinate the allocation of the rotor kinetic energy reserve of the wind turbine and the power response of the multi-source frequency modulation equipment of the power grid through a microservice architecture. The microservice architecture adopts a containerized deployment mode, dynamically plans microservices to parse compensation commands and optimize the rotor kinetic energy release ratio and converter power output. The API gateway integrates the heterogeneous communication protocols of virtual synchronous machines, energy storage devices and traditional units to generate a standardized frequency modulation command format. The event-driven priority weight microservice dynamically adjusts the device output weight based on the disturbance energy entropy. When power conflicts are detected, a redundancy elimination algorithm reallocates the frequency modulation tasks to optimize the collaborative efficiency of multi-source devices.

[0068] Introduce a genetic algorithm in model predictive control to optimize the triggering threshold of dynamic switching, and switch between the fast power support mode and the droop control mode according to the rate of change of frequency deviation. The input of the rolling horizon optimization model is historical disturbance data and real-time inertia parameters to predict the frequency trend in the future sampling period. The multi-island genetic algorithm divides the initial population and migrates elite individuals to generate candidate solutions for different triggering thresholds. When the rate of change of frequency deviation exceeds the preset threshold, switch to the fast power support mode to release the pre-stored rotor kinetic energy; in the droop control stage, dynamically adjust the power slope curve in combination with the fan speed recovery rate to balance the frequency modulation exit speed and mechanical stress limit.

[0069] When the frequency deviation returns to the preset dead zone range, the micro-service exits the frequency modulation through the adaptive power ramp rate limit and switches to the maximum power point tracking operation. The power ramp rate parameter solves the optimal solution in the hard constraint space based on the genetic algorithm, and the constraint conditions include the fan speed recovery rate and the frequency deviation range. The feasible solution set in the historical frequency modulation exit data is used to initialize the search space to shorten the optimization time. The dynamic relaxation mechanism temporarily expands the constraint boundary when detecting the risk of speed mutation, re-solves the safe exit strategy, and realizes the seamless switching from the frequency modulation mode to the maximum power tracking mode.

[0070] The above steps improve the frequency modulation response speed and stability of wind turbines in complex power grid environments through a logical closed-loop of dynamic disturbance perception, parameter optimization, control instruction generation, and multi-source collaborative execution. The data flow and control instructions between the steps are tightly coupled. The genetic algorithm and incremental learning jointly optimize the model parameters, and the micro-service architecture ensures the real-time synchronization of cross-device instructions, ultimately achieving high-precision control of the power system frequency.

[0071] Specifically, for the fast frequency modulation method of wind turbines based on disturbance adaptive compensation described in the present invention, step S2 includes: Based on the dynamic disturbance perception data set, dynamically adjust the crossover probability and mutation probability according to the population fitness distribution, and increase the mutation probability to the range of 0.2 - 0.4 when the population diversity is lower than the set threshold; Adopt the elite retention strategy to directly transfer the optimal individual of each generation to the next generation, and accelerate the parameter convergence in combination with the distribution characteristics of the disturbance feature data; Combine the genetic algorithm with GPU parallel computing to perform parallel evaluation on the window length and weight coefficient of the short-time window sliding mean algorithm; Based on the historical disturbance data in the dynamic disturbance perception data set, initialize the test signal amplitude parameter, and reduce the frequency interference of inertia estimation through reverse signal excitation.

[0072] Based on the dynamic disturbance perception dataset, the crossover probability and mutation probability are dynamically adjusted according to the population fitness distribution. The population fitness distribution is quantified by calculating the ratio of the standard deviation to the mean of the individual fitness values. When the population diversity index is lower than the preset threshold, the mutation probability is adaptively increased to the range of 0.2 - 0.4. The fitness function takes the signal-to-noise ratio of the disturbance features extracted by the short-time window moving average algorithm and the trend fitting error as evaluation indicators, and combines the roulette wheel selection mechanism to screen high-fitness individuals, balancing the requirements of global search and local optimization.

[0073] The elite retention strategy is adopted to directly transfer the optimal individual of each generation to the next generation, accelerating the parameter convergence process. The combination of the window length and weight coefficient of the elite individual is optimized based on the distribution characteristics of the disturbance feature data, including the fluctuation amplitude and change gradient of the frequency deviation. Through the tournament selection mechanism, the top 10% individuals with the highest fitness in the current population are screened, and their window parameters are used as the initial solution of the moving average algorithm to guide the subsequent iteration direction and improve the feature extraction efficiency.

[0074] The genetic algorithm is combined with GPU parallel computing to perform parallel evaluation on the window length and weight coefficient of the short-time window moving average algorithm. The parallel computing tasks are divided into multiple thread blocks, and each thread block independently calculates the noise suppression effect of different parameter combinations. Through the high-speed read and write characteristics of the GPU video memory, the fitness scores of multiple parameter combinations are synchronously processed. After integrating the evaluation results of each thread block, the optimal window parameter combination is screened. The parallel computing framework is implemented based on the Compute Unified Device Architecture, significantly shortening the single iteration time.

[0075] Based on the historical disturbance data in the dynamic disturbance perception dataset, the amplitude parameter of the test signal is initialized to generate a reverse test signal opposite to the current disturbance direction. The timing parameters of the test signal are dynamically adjusted according to the historical disturbance pattern, including the disturbance duration and energy distribution characteristics. The reverse signal excitation is superimposed on the grid equivalent inertia estimation link to suppress the frequency interference in the inertia parameter identification process. The recursive least squares method introduces a forgetting factor in the identification process to reduce the influence of historical noise on the real-time estimation result and improve the parameter identification accuracy.

[0076] The above steps form a complete link for optimizing the short-time window moving average parameters by the genetic algorithm: population diversity monitoring drives the dynamic adjustment of the crossover and mutation probabilities, the elite retention strategy accelerates the convergence direction guidance, GPU parallel computing improves the evaluation efficiency, and historical disturbance data supports the initialization of the test signal parameters. Through the cooperation of multiple mechanisms, the dynamic optimization of the sliding window parameters and noise suppression are realized, providing high-precision disturbance feature data for subsequent inertia estimation and frequency modulation control.

[0077] Specifically, the wind turbine fast frequency modulation method based on disturbance adaptive compensation described in the present invention further includes: Adjust the activation weights of fuzzy rules by using a reinforcement learning mechanism according to the execution feedback data of the compensation power command; Dynamically divide the fuzzy interval of the grid equivalent inertia parameter by a clustering algorithm, and optimize the center point and width parameters of the membership function; Establish a parameter correlation matrix between the adaptive sliding mode control model and the fuzzy controller, and synchronously update the boundary layer thickness and input range; When a high-frequency disturbance mode is detected, preferentially optimize the high-frequency segment membership function parameters of the grid equivalent inertia parameter and enhance the weight allocation of the corresponding rules.

[0078] Adjust the activation weights of fuzzy rules by using a reinforcement learning mechanism according to the execution feedback data of the compensation power command. The execution feedback data includes the frequency fluctuation suppression rate, frequency modulation energy consumption, and power overshoot. After normalization processing, a state-action reward matrix is constructed. The Q-learning algorithm iteratively updates the rule weight coefficients, and preferentially strengthens the rule activation path in the high-frequency disturbance scenario. The historical disturbance mode matching results are used to weight the reward values of similar scenarios, improve the dynamic adaptability of the fuzzy controller, and optimize the response efficiency of the compensation command.

[0079] Dynamically divide the fuzzy interval of the grid equivalent inertia parameter by a clustering algorithm, and optimize the center point and width parameters of the membership function. The clustering algorithm constructs a feature vector based on the mean value, variance, and change gradient of the inertia parameter statistically calculated by a sliding window, and uses an improved K-means algorithm to calculate the optimal clustering center. When it is detected that the distribution deviation of the inertia parameter exceeds a preset threshold, trigger the recalibration of the membership function parameters, and synchronously update the input range and rule matching threshold of the fuzzy controller to adapt to the dynamic characteristics of the grid inertia.

[0080] Establish a parameter correlation matrix between the adaptive sliding mode control model and the fuzzy controller to achieve synchronous update of the boundary layer thickness and input range. The correlation matrix defines the mapping relationship between the sliding mode switching gain parameter and the fuzzy input universe, and dynamically corrects the matrix weight coefficient through the real-time frequency deviation gradient. When the boundary layer parameters of the sliding mode model are updated, the correlation matrix drives the adjustment of the input range of the fuzzy controller, eliminates the response delay between the two-level models, and avoids control command oscillation.

[0081] When a high-frequency disturbance mode is detected, preferentially optimize the high-frequency segment membership function parameters of the grid equivalent inertia parameter and enhance the weight allocation of the corresponding rules. Analyze the spectral characteristics of the inertia parameter by fast Fourier transform, and identify high-frequency disturbance events where the frequency deviation change rate exceeds a preset threshold. For high-frequency segment disturbances, narrow the width of the membership function to 60% of the reference value to improve the rule matching resolution. The activation weights corresponding to the high-frequency segment in the fuzzy rule base are increased through a weighted summation mechanism, and high-frequency compensation commands are preferentially triggered to accelerate the power response speed.

[0082] The above steps form a closed-loop optimization link: the execution effect of the compensation instruction is fed back to the reinforcement learning module to optimize the rule weights; the real-time inertia parameters are clustered to divide the fuzzy interval to support precise control; the sliding mode and the fuzzy model parameters are associated and updated to improve the collaborative efficiency; the high-frequency disturbance characteristics drive the priority adjustment to achieve scenario adaptation. Through multi-level technology collaboration, the frequency regulation accuracy under complex disturbances is enhanced, and the loss of frequency regulation resources and equipment impact are reduced.

[0083] Specifically, for the wind turbine fast frequency regulation method based on disturbance adaptive compensation described in the present invention, the step S6 includes: In the containerized deployment framework of the model predictive control, the power distribution ratio between the rotor kinetic energy reserve and the converter capacity is optimized through the dynamic programming microservice; The communication protocols of the multi-source frequency regulation devices of the power grid are integrated through the API gateway to generate a cross-device frequency regulation instruction synchronization protocol; Based on the dynamic disturbance perception data set, the disturbance energy entropy is calculated, and an event-driven priority weight microservice is designed to dynamically adjust the output ratio; When a power conflict is detected, a redundancy elimination algorithm is triggered to reallocate the frequency regulation tasks of the virtual synchronous machine and the energy storage device.

[0084] In the containerized deployment framework of the model predictive control, the power distribution ratio between the rotor kinetic energy reserve and the converter capacity is optimized through the dynamic programming microservice. The dynamic programming microservice parses the compensation power instruction, and inputs the rotor kinetic energy reserve capacity of the wind turbine, the output upper limit of the converter, and the power grid frequency recovery requirement into the multi-objective optimization model. Based on the frequency stability constraint within the rolling time domain, a linear programming algorithm is used to solve the optimal combination of the rotor kinetic energy release ratio and the converter power output, and a power distribution instruction adapted to the dynamic disturbance scenario is generated. During the optimization process, the constraint boundary is dynamically adjusted in combination with the real-time inertia parameters, and the distribution instruction is distributed to the frequency regulation device nodes through a lightweight message queue.

[0085] The heterogeneous communication protocols of the multi-source frequency regulation devices of the power grid are integrated through the API gateway to generate a cross-device frequency regulation instruction synchronization protocol. The API gateway defines a standardized instruction interaction interface based on the RESTful architecture, and encapsulates the proprietary communication protocols of the virtual synchronous machine, the energy storage device, and the traditional unit into a unified JSON format. The gateway monitors the communication delay and bandwidth occupancy rate of each device in real time, dynamically adjusts the data packet transmission priority, and ensures the timing consistency of the cross-device frequency regulation instructions. The instruction synchronization protocol includes device status query, power instruction issuance, and execution feedback confirmation processes, and improves the instruction transmission reliability through the timeout retransmission mechanism.

[0086] Calculate the disturbance energy entropy based on the dynamic disturbance perception dataset, and design an event-driven priority weight microservice to dynamically adjust the output ratio. The disturbance energy entropy quantifies the concentration degree of disturbance energy by statistically analyzing the spatio-temporal distribution characteristics of frequency deviation, change rate, and power deficit through a sliding time window. When the entropy value exceeds the preset threshold, the priority weight microservice dynamically allocates the output weights of the virtual synchronous machine and the energy storage device according to the real-time available capacity, response rate, and frequency modulation cost coefficient of multiple-source frequency modulation devices. The weight allocation result is shared through the distributed cache of the microservice cluster to support each frequency modulation node to synchronously adjust the output target in proportion.

[0087] When a power conflict is detected, trigger the redundancy elimination algorithm to reallocate the frequency modulation tasks of the virtual synchronous machine and the energy storage device. The conflict detection module monitors the output superposition effect of multiple-source devices in real time to identify power overshoot or contradictory output directions. The redundancy elimination algorithm constructs an optimization model based on the real-time operating state parameters of the conflicting devices, and preferentially adjusts the output ratio of the devices whose response delay exceeds the threshold. The genetic algorithm initializes the population with the historical feasible solution set, quickly solves the conflict resolution scheme within the hard constraint space, and redistributes the corrected frequency modulation instructions through containerized microservices to achieve the coordinated output consistency of multiple-source devices.

[0088] The above steps form a complete closed-loop for multi-source coordinated frequency modulation: dynamic programming generates the initial allocation strategy, the API gateway ensures the cross-device synchronous transmission of instructions, the disturbance energy entropy drives the dynamic adjustment of priorities, and the redundancy elimination algorithm solves the output conflict. Through the elastic expansion ability of containerized deployment and microservice architecture, it adapts to the frequency modulation requirements of power grids of different scales, and improves the response efficiency of control instructions and the system frequency stability under multi-disturbance scenarios.

[0089] Specifically, for the wind turbine fast frequency modulation method based on disturbance adaptive compensation described in the present invention, step S7 includes: Input the historical disturbance data and real-time inertia parameters in the dynamic disturbance perception dataset into the rolling horizon optimization model to predict the frequency trend in the next 3-5 sampling periods; Use the multi-island genetic algorithm to divide the initial population of the genetic algorithm, and regularly migrate elite individuals to generate a differential trigger threshold; When the change rate of the frequency deviation exceeds the start threshold, switch to the fast power support mode to release the rotor kinetic energy reserve; In the droop control stage, generate a power slope curve based on the compensation power instruction, and dynamically update the slope rate limit by combining the fan speed recovery rate calculated from the change gradient of the fan speed after the release of the rotor kinetic energy reserve of the wind turbine.

[0090] Input the historical disturbance data and real-time inertia parameters into the rolling horizon optimization model to predict the frequency trend in the next 3-5 sampling periods. The rolling horizon optimization model intercepts the historical frequency deviation, change rate, and inertia parameter time series data based on a sliding time window to construct a prediction input vector. Generate the frequency trend curve for the future sampling period through the autoregressive integrated moving average algorithm. The prediction results include the extreme values of frequency fluctuations, the recovery time window, and the integral characteristic of the deviation. The prediction data is used to optimize the trigger threshold parameters, support the generation of the dynamic switching strategy, and improve the foresight of mode switching.

[0091] Use the multi-island genetic algorithm to divide the initial population of the genetic algorithm and regularly migrate elite individuals to generate differentiated trigger thresholds. The multi-island genetic algorithm divides the search space into multiple independently evolving subgroups, and each subgroup independently optimizes the trigger threshold combination. Every preset number of iterations, migrate the top 10% of the elite individuals with the highest fitness to the adjacent subgroups to generate differentiated threshold candidate solutions through crossover operations. The fitness function takes the frequency stability index as the core evaluation parameter and combines the switching action times constraint to screen the optimal threshold combination to balance the response speed and system stability.

[0092] When the change rate of the frequency deviation exceeds the preset starting threshold, switch to the fast power support mode to release the rotor kinetic energy reserve. The starting threshold is dynamically set based on the historical disturbance data and real-time inertia parameters, and the threshold optimization algorithm is used to reduce the risk of false triggering. In the fast power support mode, preferentially release 80%-90% of the pre-stored rotor kinetic energy and quickly inject compensation power through the converter. The release ratio is dynamically adjusted according to the frequency recovery demand predicted by the rolling horizon, and the inertial compensation algorithm is combined to smooth the power output curve and reduce the secondary impact on the power grid.

[0093] Generate the power slope curve based on the compensation power command in the droop control stage and dynamically update the slope rate limit in combination with the fan speed recovery rate. The initial value of the power slope is set according to the frequency recovery target, and the power reduction amplitude is dynamically adjusted by real-time monitoring the gradient change of the fan speed recovery rate. When the speed recovery rate is lower than the safety threshold, reduce the power slope to reduce mechanical stress; when the rate stabilizes, gradually increase the slope to accelerate the exit from frequency modulation. The dead zone control logic is introduced in the slope adjustment process to shield the interference of small speed fluctuations on parameter updates and ensure the continuity of control commands.

[0094] The above steps generate the trigger threshold through the collaboration of rolling prediction and multi-island optimization, realizing the seamless switching between fast power support and droop control modes. The historical disturbance data-driven prediction model improves the forward-looking nature of the strategy, the multi-island genetic algorithm enhances the global nature of threshold search, and the dynamic slope adjustment balances the FM exit speed and equipment safety. The data flow and control instructions in each link are tightly coupled, forming a closed-loop control link adaptable to the time-varying inertia and multi-disturbance scenarios, significantly improving the FM response accuracy of the wind turbine and the system frequency stability.

[0095] Specifically, the method for fast frequency modulation of a wind turbine based on disturbance adaptive compensation described in the present invention further includes: Taking the fan speed recovery rate and frequency deviation as hard constraint conditions, constructing a gradient search space; solving the optimal power descent gradient parameter in the constraint space through the genetic algorithm; When a risk of rotational speed mutation is detected, triggering a dynamic relaxation mechanism to expand the constraint boundary of the sliding mode control model; Initializing the search space with the optimal gradient parameter in the historical FM exit data to shorten the generation time of the safety recovery strategy.

[0096] Taking the fan speed recovery rate and frequency deviation as hard constraint conditions, constructing a gradient search space. The hard constraint conditions are set based on the mechanical strength threshold of the fan and the dead zone range of the grid frequency, defining the boundary of the multi-dimensional feasible solution region. The upper limit of the constraint on the speed recovery rate is determined by the fan rotor stress analysis model, and the frequency deviation constraint is dynamically adjusted according to the grid frequency modulation regulations. The coupling relationship between speed and frequency is transformed into a mathematical constraint through the hyperplane equation, restricting the search range of the genetic algorithm and ensuring the physical feasibility of the power descent strategy.

[0097] Solving the optimal power descent gradient parameter in the constraint space through the genetic algorithm. The initial population of the genetic algorithm is generated based on the feasible solution set in the historical FM exit data, and the fitness function takes the frequency secondary drop suppression effect and the rotational speed recovery smoothness during the power descent process as evaluation indicators. A constraint violation penalty mechanism is used to screen the candidate solutions that meet the hard constraint conditions, and iterative optimization is performed through crossover and mutation operations to output a power ramp rate parameter combination that takes into account both safety and economy, supporting the smooth transition during the FM exit stage.

[0098] When a risk of rotational speed mutation is detected, triggering a dynamic relaxation mechanism to expand the constraint boundary of the sliding mode control model. The risk of rotational speed mutation is identified by real-time monitoring of the absolute value of the rotational speed change gradient. When the gradient value exceeds 80% of the preset safety threshold, the thickness adjustment of the constraint boundary layer is triggered. The dynamic relaxation mechanism temporarily expands the boundary of the feasible solution region, allowing the genetic algorithm to re-solve in the expanded search space, and gradually shrinks the relaxed constraint range in combination with the rotational speed recovery trend until the mutation risk is lifted and the original constraint conditions are restored.

[0099] Initialize the search space by combining the optimal gradient parameters in the historical frequency modulation exit data, and shorten the generation time of the safety recovery strategy. The historical disturbance data is stored by disturbance scenario classification, including the frequency deviation range, inertia parameters, and speed recovery characteristics. When the current disturbance mode is detected to match the historical record, the optimal solution in the same type of scenario is preferentially loaded as the initial individual of the genetic algorithm. The fitness score of the historical solution is corrected through a similarity weighting mechanism to accelerate the population convergence. The historical disturbance data is dynamically maintained through a sliding time window update mechanism, and records with insufficient timeliness are removed to ensure the real-time effectiveness of the initialization strategy.

[0100] The above steps form a safety recovery closed loop in the frequency modulation exit stage: the hard constraints define to ensure the safety of the strategy, the genetic algorithm searches for the optimal solution within the feasible space, the dynamic relaxation deals with sudden anomalies, and the injection of historical disturbance data improves the optimization efficiency. Through the coordination of multiple mechanisms, the generation time of the power reduction strategy is shortened, the risk of mechanical damage caused by sudden speed changes is reduced, and at the same time, the secondary frequency drop is suppressed, realizing a seamless switch of the wind turbine from the frequency modulation mode to the maximum power point tracking mode.

[0101] Specifically, for the fast frequency modulation method of a wind turbine based on disturbance adaptive compensation described in the present invention, step S8 includes: Collect the real-time data stream of the continuous disturbance event through a sliding time window and perform principal component analysis for dimensionality reduction. Use the recursive least squares method to update the boundary layer parameters of the adaptive sliding mode control model only based on the incremental data. When data anomalies are detected, trigger the local model rollback mechanism to restore to the effective parameters of the previous time period. Dynamically adjust the learning rate according to the change rate of the model prediction error, and increase the parameter update speed when the error gradient increases.

[0102] Collect the real-time data stream of the continuous disturbance event through a sliding time window and perform principal component analysis for dimensionality reduction. The length of the sliding time window is dynamically adjusted according to the typical disturbance duration, and the window contains the time series data of frequency deviation, change rate, and power deficit. Principal component analysis calculates the eigenvalue distribution of the covariance matrix, screens the principal components with a contribution rate exceeding a preset threshold, extracts the low-dimensional feature vector representing the disturbance characteristics, eliminates redundant noise interference, generates an input data set suitable for the sliding mode control model, and improves the subsequent parameter update efficiency.

[0103] The recursive least squares method is used to update the boundary layer parameters of the adaptive sliding mode control model based only on incremental data. The recursive update process retains the covariance matrix and residual information of the historical parameters, and each time a new data window arrives, only the incremental data is used to calculate the parameter correction. The boundary layer thickness and switching gain parameters are updated through iterative formulas to avoid the computational overhead of retraining the full amount of data. The updated parameters directly affect the switching frequency and tracking accuracy of the sliding mode control, balancing the control response speed and system stability.

[0104] When data anomalies are detected, the local model rollback mechanism is triggered to restore to the valid parameters of the previous period. The anomaly detection is achieved by monitoring the statistical characteristics of the data stream, including mean drift, variance mutation, and outlier identification beyond the confidence interval. The rollback mechanism calls the historical parameter database, loads the sliding mode control parameters that have been verified to be valid in the previous period, and overwrites the update results of the current abnormal window. The model after rollback maintains stable output until the subsequent continuous window data passes the integrity check and the incremental update process is restored to ensure the continuity of the control instructions.

[0105] The learning rate is dynamically adjusted according to the rate of change of the model prediction error, and the parameter update speed is increased when the error gradient increases. The error change rate is calculated by the prediction error difference in the sliding window, combined with the exponential weighted average to eliminate the interference of instantaneous fluctuations. When the absolute value of the error gradient exceeds the preset threshold, the learning rate is increased according to the logarithmic function law to accelerate parameter convergence; when the error gradient approaches zero, the learning rate is gradually reduced to the baseline value to refine the parameter search accuracy. The dynamic adjustment mechanism balances the parameter update speed and stability to avoid overfitting or underfitting problems.

[0106] The above steps form a complete closed loop for online self-adaptation of the sliding mode control model: the sliding window collects data to support feature extraction, and recursive updates reduce the computational load; abnormal rollback ensures the robustness of the model, and the learning rate is dynamically adjusted to optimize the convergence process. Through the coordination of multiple mechanisms, the control parameters can be adapted to the dynamic disturbance characteristics in real time, the response speed and anti-interference ability of the frequency regulation control of wind turbines in complex scenarios can be improved, and the frequency stability of the power system can be enhanced.

[0107] The explanations of the various technical features in the technical solution of the present invention are as follows: Dynamic disturbance perception data set: Through the frequency measurement unit, phasor measurement unit and wind turbine status monitoring module deployed at the power grid node, data such as frequency deviation, frequency deviation change rate and wind turbine speed-power mapping parameters are collected in real time. After the collected raw data is aligned with the timestamp, the sliding window mechanism is used to remove outliers to form a structured data set containing time series features and statistical features. This data set provides the basic input for subsequent disturbance feature extraction and parameter identification, and its dynamic update characteristics adapt to the time-varying characteristics of power grid disturbances.

[0108] Genetic Algorithm Optimized Short-Time Window Moving Average Algorithm: The genetic algorithm constructs an initial population with the sliding window length and weight coefficient as optimization variables. The fitness function evaluates the individual performance based on the signal-to-noise ratio of the disturbance feature and the trend fitting error. High-fitness individuals are selected through the roulette wheel selection mechanism, and the crossover probability and mutation probability are dynamically adjusted (the mutation probability is increased to the range of 0.2 - 0.4 when the population diversity is insufficient). Combined with GPU parallel computing to accelerate the evaluation of the noise suppression effect of window parameter combinations, the optimal parameter combination is output to suppress the interference of grid background noise on the extraction of disturbance features.

[0109] Recursive Least Squares Inertia Parameter Estimation: A reverse test signal is generated based on the disturbance feature data after noise suppression. The signal amplitude is initialized according to the historical disturbance energy distribution, and the timing parameters are synchronously adjusted with the current disturbance features. The recursive least squares method introduces a forgetting factor (0.95 - 0.98) to optimize the weight ratio of historical disturbance data, and the estimated values of the grid equivalent inertia and damping parameters are corrected in real time through incremental updates to improve the tracking accuracy of the parameters in the inertia time-varying scenario.

[0110] Online Incremental Learning to Update Sliding Mode Control Parameters: The sliding time window intercepts the data stream of continuous disturbance events, and after dimensionality reduction by principal component analysis, it is input into the sliding mode control model. The recursive least squares method updates the boundary layer thickness and switching gain parameters only based on incremental data, reducing the computational complexity. When a data mean drift or variance mutation is detected, a local model rollback mechanism is triggered to load the parameters that have been verified effective in the previous time period to maintain control stability.

[0111] Fuzzy Adaptive Proportional-Integral Controller: The membership function is dynamically divided by the K-means clustering algorithm, and the number of clustering centers is set to 3 - 5 according to the distribution characteristics of the inertia parameters. The reinforcement learning mechanism uses the frequency modulation energy consumption and frequency overshoot as reward indicators to iteratively update the activation weights of the fuzzy rules. In the high-frequency disturbance scenario, the width of the membership function in the high-frequency band is reduced to 60% of the reference value, enhancing the weight allocation of the corresponding rules and improving the dynamic response speed of the compensation command.

[0112] Multi-source Cooperative Control of Microservice Architecture: The dynamic programming microservice analyzes the compensation power command, and optimizes the rotor kinetic energy release ratio and converter power output with the frequency stability as the constraint condition. The API gateway encapsulates the heterogeneous communication protocols of the virtual synchronous machine and energy storage device into standardized JSON commands, and ensures cross-device command synchronization through the timeout retransmission mechanism. The disturbance energy entropy quantifies the concentration degree of disturbance energy, triggers the event-driven priority weight adjustment, and the redundancy elimination algorithm reallocates the frequency modulation tasks of conflicting devices based on the genetic algorithm within 10 - 20 seconds.

[0113] Genetic Algorithm Optimization of Mode Switching Threshold: The rolling horizon optimization model inputs historical disturbance data and real-time inertia parameters to predict the frequency trend in the next 3 - 5 sampling periods. The multi-island genetic algorithm divides the search space into 4 - 6 subpopulations and periodically migrates elite individuals to generate differentiated trigger thresholds. When the rate of change of frequency deviation exceeds 120% of the dynamically set threshold, it switches to the fast power support mode to release 80% - 90% of the pre-stored rotor kinetic energy; during the droop control stage, the power slope curve is dynamically adjusted according to the gradient of the speed recovery rate, and the initial slope is set to 5% / second of the rated power.

[0114] Adaptive Power Ramp Rate Exit Mechanism: The genetic algorithm takes the speed recovery rate ≤ 2% / second and frequency deviation ≤ 0.05Hz as hard constraints, and solves for the optimal power descent gradient parameter in the search space initialized with the historical feasible solution set. When the speed gradient mutation exceeds 80% of the safety threshold, the dynamic relaxation mechanism temporarily expands the constraint boundary by 20% - 30%, and combines historical frequency modulation exit data to accelerate the optimization process, realizing a seamless switch from the frequency modulation mode to the maximum power point tracking mode.

[0115] Construction of Dynamic Disturbance Perception Dataset: Through the frequency measurement unit, phasor measurement unit, and fan status monitoring module deployed at the power grid nodes, the frequency deviation, rate of change of frequency deviation, and fan speed-power mapping parameters are collected in real time. After the collected data is aligned by timestamp and the outliers in the sliding window are removed, a structured dataset containing time series and statistical features is formed. This dataset adapts to the power grid disturbance characteristics through a dynamic update mechanism, providing basic data support for subsequent noise suppression and parameter identification.

[0116] Genetic Algorithm Optimization of Short-Time Window Moving Average Algorithm: Taking the sliding window length and weight coefficient as optimization variables, an initial population is constructed and the fitness of each individual is evaluated. The fitness function takes the signal-to-noise ratio of disturbance characteristics and the trend fitting error as evaluation indicators, and uses the roulette wheel selection and elite retention strategies to screen out individuals with high fitness. When the population diversity index is lower than the threshold, the mutation probability is dynamically increased to the range of 0.2 - 0.4, and combined with GPU parallel computing to accelerate the evaluation of parameter combinations, and the optimal window parameters are output to suppress the interference of power grid background noise on the extraction of disturbance characteristics.

[0117] Recursive Least Squares Inertia Parameter Estimation: Based on the disturbance characteristics after noise suppression, a reverse test signal is generated, and the signal amplitude is initialized according to the 80th percentile value of the historical disturbance energy distribution. The equivalent inertia and damping parameters of the power grid are updated online through the recursive least squares method, and a forgetting factor (0.95 - 0.98) is introduced to reduce the weight of historical noise, and the parameter estimation value is corrected in real time to improve the identification accuracy in the time-varying inertia scenario.

[0118] Online incremental learning to update sliding mode control parameters: A sliding time window is used to intercept the continuous disturbance event data stream, which is input into the sliding mode control model after dimensionality reduction by principal component analysis. The recursive least squares method updates the boundary layer thickness and switching gain parameters only based on incremental data. When the detected data mean drift exceeds 2 times the standard deviation, a local model rollback mechanism is triggered to restore to the effective parameters of the previous time period, maintaining the continuity and stability of the control command.

[0119] Fuzzy adaptive proportional-integral controller: The fuzzy interval of the grid equivalent inertia parameter is dynamically divided by the K-means clustering algorithm, and the number of cluster centers is set to 3 - 5, optimizing the center point and width of the membership function. The reinforcement learning mechanism takes the frequency modulation energy consumption and frequency overshoot as the reward indicators to iteratively update the activation weights of the fuzzy rules. In the high-frequency disturbance scenario, the width of the membership function in the high-frequency band is reduced to 60% of the reference value, enhancing the priority of the corresponding rules and improving the dynamic response speed of the compensation command.

[0120] Multi-source collaborative control of microservice architecture: The microservices are dynamically planned to analyze and compensate the power command. With the frequency stability as the constraint condition, the linear programming algorithm is used to optimize the rotor kinetic energy release ratio and the converter power output. The API gateway encapsulates the Modbus and IEC 61850 protocols of the virtual synchronous machine and energy storage device into standardized JSON commands, and ensures cross-device command synchronization through the timeout retransmission mechanism. The disturbance energy entropy calculates the concentration degree of the disturbance energy, triggering the priority weight adjustment driven by events. The redundancy elimination algorithm redistributes the frequency modulation tasks of the conflicting devices based on the genetic algorithm within 10 - 20 seconds.

[0121] Genetic algorithm to optimize the mode switching threshold: The rolling horizon optimization model inputs historical disturbance data and real-time inertia parameters to predict the frequency trend in the next 3 - 5 sampling periods. The multi-island genetic algorithm divides the search space into 4 - 6 subpopulations, and migrates elite individuals every 10 generations to generate different triggering thresholds. When the rate of change of the frequency deviation exceeds 120% of the dynamically tuned threshold, it switches to the fast power support mode to release 80% - 90% of the pre-stored rotor kinetic energy, and combines the inertia compensation algorithm to smooth the power output curve.

[0122] Adaptive power ramp rate exit mechanism: With the rotational speed recovery rate ≤ 2% / s and the frequency deviation ≤ 0.05Hz as the hard constraint conditions, the genetic algorithm solves the optimal power descent gradient parameter in the search space initialized by the historical feasible solution set. The dynamic relaxation mechanism temporarily expands the constraint boundary by 20% - 30% when the rotational speed gradient mutation exceeds 80% of the safety threshold, and adjusts the power ramp rate in combination with the recursive update strategy to achieve a seamless switch from the frequency modulation mode to the maximum power point tracking mode.

[0123] The specific implementation of the present invention is based on the frequency regulation requirements in the scenarios of dynamic multi-disturbance and inertia time-variation, and realizes the fast frequency regulation and multi-source collaborative control of wind turbines through the following technical solutions: In the dynamic disturbance perception stage, the frequency measurement unit, phasor measurement unit and wind turbine state monitoring module deployed at key grid nodes collect the frequency deviation, the rate of change of frequency deviation and the wind turbine speed-power mapping parameters in real time. The sampling frequency is dynamically adjusted to 5-10 times per second according to the grid disturbance intensity. After the collected raw data is aligned by time stamps and the outliers are removed by the sliding window, a multi-dimensional dynamic disturbance perception data set containing time series features and statistical features is constructed. The window length and weight coefficient of the short-time window sliding mean algorithm are dynamically optimized by the genetic algorithm. The initial population size is set to 50-100 groups of parameter combinations. When the population diversity is lower than the threshold, the mutation probability is increased to the range of 0.2-0.4. Combined with GPU parallel computing to accelerate the fitness evaluation, the window parameters with the optimal signal-to-noise ratio are selected to generate the disturbance feature data after noise suppression. This data drives the generation of the reverse test signal. The signal amplitude is initialized based on the 80th percentile value of the historical disturbance energy distribution. The equivalent inertia parameter of the power grid is identified online by the recursive least squares method, and the forgetting factor is set to 0.95-0.98 to balance the weight of historical disturbance data and improve the parameter tracking accuracy in the inertia time-variation scenario.

[0124] In the control strategy optimization stage, the online incremental learning algorithm is used to model continuous disturbance events. The sliding time window length is set to 3-5 disturbance cycles. The principal components with a contribution rate exceeding 85% are retained by the principal component analysis, and the dimensionality-reduced feature vector is input into the sliding mode control model. The recursive least squares method updates the boundary layer thickness and switching gain parameters only based on the incremental data. When it is detected that the data mean drift exceeds 2 times the standard deviation, the local model rollback mechanism is triggered to load the effective parameters of the previous time period. The membership function of the fuzzy adaptive proportional-integral controller is dynamically divided by K-means clustering, and the number of cluster centers is set to 3-5. The reinforcement learning mechanism takes the frequency regulation energy consumption and the frequency overshoot as the reward indicators to iteratively update the activation weights. In the microservice architecture, the dynamic programming microservice solves the optimal allocation ratio of the rotor kinetic energy and the converter power with the frequency stability as the constraint. The API gateway integrates the Modbus and IEC 61850 protocols of the virtual synchronous machine and the energy storage device to generate the cross-device frequency regulation instruction in JSON format. When the disturbance energy entropy exceeds the threshold, the priority weight microservice is triggered to dynamically adjust the output ratio. The redundancy elimination algorithm generates a conflict resolution scheme based on the genetic algorithm within 10-20 seconds.

[0125] During the mode switching and safe exit phase, the rolling horizon optimization model predicts the frequency trend in the next 3 - 5 sampling periods based on historical disturbance data and real - time inertia parameters. The multi - island genetic algorithm divides into 4 - 6 sub - populations, and migrates elite individuals every 10 generations to generate a differential triggering threshold. When the rate of change of frequency deviation exceeds 120% of the dynamically set threshold, it switches to the fast power support mode to release 80% - 90% of the pre - stored rotor kinetic energy. In the droop control phase, the power slope curve is dynamically adjusted based on the speed recovery rate gradient, and the initial value of the slope is set to 5% / s of the rated power. When the frequency modulation exits, the genetic algorithm solves for the optimal power ramp rate with the hard constraints of the speed recovery rate ≤ 2% / s and the frequency deviation ≤ 0.05 Hz. The historical feasible solution set initializes the population size to 30 - 50 groups. The dynamic relaxation mechanism temporarily expands the constraint boundary by 20% - 30% when the speed gradient mutation exceeds 80% of the safety threshold, realizing a seamless switch to the maximum power point tracking operation.

[0126] The above - mentioned implementation method solves the problem of response lag in the traditional method under the scenarios of time - varying inertia and superposition of multiple disturbances through a closed - loop link of dynamic perception, parameter optimization, coordinated control and safe exit. The data interaction and instruction coordination of each technical link strictly match the features in the claims, improving the dynamic adaptability of the frequency modulation instruction and the coordinated efficiency of multi - source devices, and finally realizing the high - precision stable control of the power system frequency.

[0127] The technical solution of the present invention solves the problem of the decline in the power system frequency stability caused by the lag of the wind turbine frequency modulation response and the insufficient coordination of multi - source frequency modulation resources under the conditions of dynamic multi - disturbances and time - varying inertia in the following way: First, through the construction of a dynamic disturbance perception data set and genetic algorithm optimization, the accuracy of disturbance feature extraction and the efficiency of inertia parameter identification are improved. The frequency deviation, rate of change of frequency deviation and fan operation state parameters of the power system are collected in real - time to construct a dynamic disturbance perception data set; based on the genetic algorithm, the window length and weight coefficient of the short - time window sliding mean algorithm are dynamically optimized to suppress the interference of grid background noise and generate disturbance feature data with high signal - to - noise ratio; combined with the recursive least - squares method, the equivalent inertia and damping parameters of the power grid are estimated in real - time to provide accurate inputs for the subsequent control model. The above - mentioned technologies work together to overcome the problems of perception lag and insufficient parameter identification accuracy in the traditional method under dynamic disturbance scenarios, and shorten the frequency modulation response time.

[0128] Secondly, online incremental learning algorithm and fuzzy adaptive control are used for collaborative optimization to enhance the dynamic adaptability of frequency modulation instructions and the coordination ability of multi-source resources. The continuous disturbance event data stream is collected through a sliding time window, and the boundary layer thickness and switching gain parameters of the sliding mode control model are updated using the recursive least squares method to adapt to changes in disturbance characteristics in real time; the compensation power instruction is generated based on the fuzzy adaptive proportional-integral controller, and the fuzzy rule weights are dynamically adjusted in combination with the reinforcement learning mechanism to optimize the instruction response speed under high-frequency disturbances; the allocation of wind turbine rotor kinetic energy reserves and the output of multi-source frequency modulation equipment are coordinated through the microservice architecture, and an event-driven priority weight dynamic adjustment strategy is designed to eliminate power redundancy and conflict. The above technologies realize self-optimization of control parameters and synchronization of cross-device instructions, and improve the collaborative efficiency of multi-source frequency modulation resources.

[0129] Finally, a genetic algorithm is introduced to optimize the dynamic switching threshold and safe exit mechanism to balance the switching speed of the frequency modulation mode and the stability of the system. In the model predictive control, the rolling time domain optimization model is used to predict the future frequency trend, and the multi-island genetic algorithm is combined to generate differentiated trigger thresholds, and the fast power support mode and the droop control mode are switched according to the frequency deviation change rate; the wind turbine speed recovery rate and frequency deviation are used as hard constraints, and the genetic algorithm is used to solve the optimal power reduction gradient parameters, combined with the dynamic relaxation mechanism to deal with the risk of speed mutation; the microservice is smoothly exited from the frequency modulation through the adaptive power ramp rate limit, and switched to the maximum power point tracking operation. The above technologies ensure the safety and economy of the frequency modulation strategy in the scenario of time-varying inertia, suppress the secondary frequency drop, and ultimately achieve an overall improvement in the frequency stability of the power system.

Claims

1. A method for rapid frequency regulation of wind turbines based on adaptive disturbance compensation, characterized in that: include: Step S1, collecting the frequency deviation, frequency deviation change rate and wind turbine operating state parameters of the power system to construct a dynamic disturbance perception data set, wherein the wind turbine operating state parameters include real-time monitoring data of the kinetic energy reserve of the wind turbine rotor; Step S2, based on the dynamic disturbance perception data set, dynamically optimizing the window length and weight coefficient of the short-time window sliding mean algorithm through a genetic algorithm to generate disturbance feature data after noise suppression; Step S3, generating a reverse test signal according to the disturbance characteristic data, and estimating the equivalent inertia and damping parameters of the power grid in real time by combining the recursive least squares method; Step S4, using an online incremental learning algorithm to perform time series modeling on the continuous disturbance events in the dynamic disturbance perception data set, and updating the boundary layer thickness and switching gain parameters of the preset adaptive sliding mode control model; Step S5, based on the equivalent inertia and damping parameters of the power grid, generating a compensation power instruction through a preset fuzzy adaptive proportional and preset integral controller, wherein the compensation power instruction includes triggering conditions of a fast power support mode and a droop control mode; Step S6, according to the compensation power instruction, coordinating the wind turbine rotor kinetic energy reserve allocation and the power response of the power grid multi-source frequency regulation equipment through a preset microservice architecture, wherein the power grid multi-source frequency regulation equipment includes a virtual synchronous machine and an energy storage device; Step S7, introducing the genetic algorithm to optimize the trigger threshold of dynamic switching in the adaptive sliding mode control model predictive control, and switching between the fast power support mode and the droop control mode according to the frequency deviation change rate; Step S8: When the frequency deviation returns to the preset dead zone range, the microservice exits the frequency modulation and switches to the maximum power point tracking operation through the adaptive power ramp rate limiting microservice.

2. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 1 is characterized in that: The step S2 comprises: Based on the dynamic disturbance perception data set, the crossover probability and mutation probability are dynamically adjusted according to the population fitness distribution, and the mutation probability is increased to the range of 0.2-0.4 when the population diversity is lower than the set threshold; An elite retention strategy is adopted to directly pass the best individuals of each generation to the next generation, and the distribution characteristics of the disturbance feature data are combined to accelerate parameter convergence; Combining the genetic algorithm with GPU parallel computing, the window length and weight coefficient of the short-time window sliding mean algorithm are evaluated in parallel; The test signal amplitude parameter is initialized based on the historical disturbance data in the dynamic disturbance perception data set, and the frequency interference of inertia estimation is reduced by reverse signal excitation.

3. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 1, characterized in that: Also includes: According to the execution feedback data of the compensation power instruction, the activation weight of the fuzzy rule is adjusted by using a reinforcement learning mechanism; Dynamically divide the fuzzy interval of the equivalent inertia parameter of the power grid by a clustering algorithm, and optimize the center point and width parameter of the membership function; Establishing the parameter association matrix of the adaptive sliding mode control model and the fuzzy controller, and synchronously updating the boundary layer thickness and the input range; When a high-frequency disturbance mode is detected, the high-frequency band membership function parameters of the power grid equivalent inertia parameters are preferentially optimized and the weight distribution of the corresponding rules is enhanced.

4. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 1, characterized in that: The step S6 comprises: In the containerized deployment framework of the model predictive control, the power allocation ratio of the rotor kinetic energy reserve and the converter capacity is optimized through dynamic programming microservices; Integrate the communication protocol of the multi-source frequency modulation equipment of the power grid through the API gateway to generate a cross-device frequency modulation instruction synchronization protocol; Calculate disturbance energy entropy based on the dynamic disturbance perception data set, and design an event-driven priority weight microservice to dynamically adjust the output ratio; When a power conflict is detected, a redundancy elimination algorithm is triggered to reallocate the frequency regulation tasks of the virtual synchronous machine and the energy storage device.

5. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 1, characterized in that: The step S7 comprises: Inputting historical disturbance data and real-time inertia parameters in the dynamic disturbance perception data set into a rolling time domain optimization model to predict the frequency trend of the next 3-5 sampling periods; The initial population of the genetic algorithm is divided by using a multi-island genetic algorithm, and elite individuals are regularly migrated to generate a differentiated trigger threshold; When the frequency deviation change rate exceeds the start threshold, switching to a fast power support mode to release the rotor kinetic energy reserve; In the droop control stage, a power slope curve is generated based on the compensation power command, and the wind turbine speed recovery rate is calculated based on the speed change gradient after the wind turbine rotor kinetic energy reserve is released. The slope rate limit is dynamically updated based on the wind turbine speed recovery rate.

6. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 5 is characterized in that: Also includes: Taking the fan speed recovery rate and frequency deviation as hard constraints, a gradient search space is constructed; Solving the optimal power reduction gradient parameter in the constraint space by using the genetic algorithm; When a risk of a sudden change in speed is detected, a dynamic relaxation mechanism is triggered to expand the constraint boundary of the sliding mode control model; The search space is initialized by combining the optimal gradient parameters in the historical frequency modulation exit data to shorten the generation time of the safety recovery strategy.

7. The method for rapid frequency regulation of wind turbines based on disturbance adaptive compensation according to claim 1, characterized in that: The step S8 comprises: The real-time data stream of the continuous disturbance event is collected through a sliding time window, and a principal component analysis and dimensionality reduction process is performed; Using a recursive least squares method to update boundary layer parameters of the adaptive sliding mode control model based only on incremental data; When data anomalies are detected, the local model rollback mechanism is triggered to restore to the valid parameters of the previous period; The learning rate is dynamically adjusted according to the rate of change of the model prediction error, and the parameter update speed is increased when the error gradient increases.

Images